Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Refactor tuning script #190

Merged
merged 27 commits into from
Feb 14, 2024
Merged

Refactor tuning script #190

merged 27 commits into from
Feb 14, 2024

Conversation

maleadt
Copy link
Member

@maleadt maleadt commented Jan 30, 2024 •
edited
Loading

The goal is to make the script more resilient, i.e., automatically restarting workers if they fail. It should also be much faster, as I switched from doing relatively brute-force measurements, to waiting for the device to be unthrottled and doing only few measurements. Finally, I've also split off the WMMA/GEMM-specific parts so that it should be easier to add support for e.g. FPU/GEMM or TC tuning.

@maleadt
Copy link
Member Author

maleadt commented Jan 30, 2024 •
edited
Loading

Benchmark results for commit 158ce73 (comparing to 480924c):

test master PR Δmin
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 128, 16) Warps (4, 1) OP (16, 16, 16) 8.74 μs ± 1.7% (8.34 … 9.06 μs) 9.09 μs ± 1.93% (8.82 … 9.3 μs) +5.7% ❌
FPU GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) OP (8, 16, 2), base shape (4, 8, 1) 40.9 μs ± 0.482% (40.3 … 41.5 μs) 43.3 μs ± 0.491% (42.7 … 43.9 μs) +5.9% ❌
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 64, 64) Warps (1, 4) OP (16, 16, 16) 3.88 μs ± 3.16% (3.81 … 4.29 μs) 3.86 μs ± 2.74% (3.58 … 4.29 μs) -6.2% ✅
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (128, 64, 16) Warps (1, 4) OP (16, 16, 16) 9.67 μs ± 1.55% (9.3 … 10.0 μs) 10.1 μs ± 1.43% (9.78 … 10.3 μs) +5.1% ❌
Tropical GEMM Float32*Float32=Float32 (256×256) · (256×256) (TT) OP (8, 16, 2), base shape (4, 8, 1) 81.5 μs ± 0.248% (81.1 … 82.3 μs) 73.3 μs ± 0.278% (72.7 … 74.1 μs) -10.3% ✅
FPU GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) OP (8, 16, 2), base shape (4, 8, 1) 41.0 μs ± 0.536% (40.3 … 41.5 μs) 43.0 μs ± 0.521% (42.4 … 44.1 μs) +5.3% ❌
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 64, 32) Warps (4, 2) OP (16, 16, 16) 3.63 μs ± 3.03% (3.58 … 4.05 μs) 3.6 μs ± 2.29% (3.34 … 4.05 μs) -6.7% ✅
Tropical GEMM Float32*Float32=Float32 (256×256) · (256×256) (NT) OP (8, 16, 2), base shape (4, 8, 1) 81.7 μs ± 0.251% (81.3 … 82.3 μs) 73.3 μs ± 0.275% (72.7 … 73.9 μs) -10.6% ✅
Tropical GEMM Float32*Float32=Float32 (256×128) · (128×256) (NT) OP (8, 16, 2), base shape (4, 8, 1) 41.9 μs ± 0.358% (41.2 … 42.4 μs) 37.9 μs ± 0.413% (37.4 … 38.4 μs) -9.2% ✅
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 64, 32) Warps (2, 2) OP (16, 16, 16) 4.11 μs ± 2.77% (3.81 … 4.53 μs) 4.14 μs ± 3.35% (4.05 … 4.53 μs) +6.2% ❌
FPU GEMM Float32*Float32=Float32 (256×256) · (256×256) (TT) OP (8, 16, 2), base shape (4, 8, 1) 81.7 μs ± 0.254% (81.1 … 82.3 μs) 73.8 μs ± 0.297% (73.2 … 74.4 μs) -9.7% ✅
Tropical GEMM Float32*Float32=Float32 (256×128) · (128×256) (TT) OP (8, 16, 2), base shape (4, 8, 1) 41.8 μs ± 0.348% (41.2 … 42.2 μs) 37.9 μs ± 0.408% (37.2 … 38.1 μs) -9.8% ✅
FPU GEMM Float32*Float32=Float32 (256×256) · (256×256) (NN) OP (8, 16, 2), base shape (4, 8, 1) 81.5 μs ± 0.25% (81.1 … 82.3 μs) 73.6 μs ± 0.279% (73.2 … 74.1 μs) -9.7% ✅
FPU GEMM Int32*Int32=Int32 (256×256) · (256×256) (NT) OP (8, 16, 2), base shape (4, 8, 1) 81.7 μs ± 0.251% (81.3 … 82.3 μs) 73.4 μs ± 0.277% (73.0 … 74.1 μs) -10.3% ✅
FPU GEMM Int32*Int32=Int32 (256×256) · (256×256) (TN) OP (8, 16, 2), base shape (4, 8, 1) 82.0 μs ± 0.248% (81.5 … 82.5 μs) 73.6 μs ± 0.297% (73.2 … 74.4 μs) -10.2% ✅
Tropical GEMM Float32*Float32=Float32 (256×256) · (256×256) (NN) OP (8, 16, 2), base shape (4, 8, 1) 81.5 μs ± 0.252% (80.8 … 82.0 μs) 73.3 μs ± 0.274% (73.0 … 73.9 μs) -9.7% ✅
FPU GEMM Float32*Float32=Float32 (256×256) · (256×256) (TN) OP (8, 16, 2), base shape (4, 8, 1) 81.8 μs ± 0.252% (81.3 … 82.5 μs) 73.8 μs ± 0.304% (73.2 … 74.6 μs) -10.0% ✅
FPU GEMM Int32*Int32=Int32 (256×256) · (256×256) (NN) OP (8, 16, 2), base shape (4, 8, 1) 81.8 μs ± 0.249% (81.3 … 82.3 μs) 73.6 μs ± 0.281% (73.2 … 74.4 μs) -10.0% ✅
FPU GEMM Int16*Int16=Int16 (256×256) · (256×256) (TN) OP (8, 16, 2), base shape (4, 8, 1) 39.8 μs ± 0.577% (39.1 … 40.5 μs) 42.1 μs ± 0.447% (41.7 … 42.9 μs) +6.7% ❌
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 64, 32) Warps (2, 4) OP (16, 16, 16) 3.66 μs ± 3.33% (3.58 … 4.05 μs) 3.62 μs ± 2.95% (3.34 … 4.05 μs) -6.7% ✅
FPU GEMM Int16*Int16=Int16 (256×256) · (256×256) (TT) OP (8, 16, 2), base shape (4, 8, 1) 40.3 μs ± 0.58% (39.6 … 41.0 μs) 42.5 μs ± 0.485% (42.0 … 43.2 μs) +6.0% ❌
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 64, 64) Warps (1, 4) OP (16, 16, 16) 3.6 μs ± 2.25% (3.58 … 4.05 μs) 3.59 μs ± 1.92% (3.34 … 4.05 μs) -6.7% ✅
Tropical GEMM Float32*Float32=Float32 (256×256) · (256×256) (TN) OP (8, 16, 2), base shape (4, 8, 1) 81.5 μs ± 0.24% (81.1 … 82.0 μs) 73.3 μs ± 0.265% (72.7 … 73.9 μs) -10.3% ✅
FPU GEMM Int32*Int32=Int32 (256×256) · (256×256) (TT) OP (8, 16, 2), base shape (4, 8, 1) 81.7 μs ± 0.251% (81.1 … 82.3 μs) 73.4 μs ± 0.286% (73.0 … 74.1 μs) -10.0% ✅
FPU GEMM Float32*Float32=Float32 (256×256) · (256×256) (NT) OP (8, 16, 2), base shape (4, 8, 1) 81.7 μs ± 0.254% (81.1 … 82.3 μs) 73.5 μs ± 0.286% (73.0 … 74.1 μs) -10.0% ✅
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 64, 32) Warps (4, 1) OP (16, 16, 16) 3.63 μs ± 2.93% (3.58 … 4.05 μs) 3.62 μs ± 2.82% (3.34 … 4.05 μs) -6.7% ✅
FPU GEMM Int16*Int16=Int16 (256×256) · (256×256) (NN) OP (8, 16, 2), base shape (4, 8, 1) 39.7 μs ± 0.558% (39.1 … 40.3 μs) 41.9 μs ± 0.462% (41.2 … 42.9 μs) +5.5% ❌
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 64, 64) Warps (1, 4) OP (16, 16, 16) 3.57 μs ± 1.98% (3.1 … 3.81 μs) 3.61 μs ± 2.48% (3.34 … 4.05 μs) +7.7% ❌
Tropical GEMM Float32*Float32=Float32 (256×128) · (128×256) (TN) OP (8, 16, 2), base shape (4, 8, 1) 41.7 μs ± 0.316% (41.2 … 42.2 μs) 37.8 μs ± 0.436% (37.2 … 38.1 μs) -9.8% ✅
Tropical GEMM Float32*Float32=Float32 (256×128) · (128×256) (NN) OP (8, 16, 2), base shape (4, 8, 1) 41.8 μs ± 0.309% (41.2 … 42.2 μs) 37.8 μs ± 0.417% (37.4 … 38.1 μs) -9.2% ✅

Comparison with baseline

test GemmKernels Baseline %
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 128, 16) Warps (4, 2) OP (16, 16, 16) 8.17 μs ± 1.76% (7.87 … 8.58 μs) 3.22 μs ± 4.27% (2.86 … 3.34 μs) 36.4
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (128, 128, 32) Warps (4, 4) OP (16, 16, 16) 10.5 μs ± 1.09% (10.5 … 11.0 μs) 3.47 μs ± 3.87% (3.1 … 3.81 μs) 29.5
WMMA GEMM Float16*Float16+Float16=Float16 (2048×2048) · (2048×2048) (TN) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 3.01 ms ± 0.415% (2.98 … 3.07 ms) 179.0 μs ± 0.341% (178.0 … 181.0 μs) 5.96
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 256, 64) Warps (4, 4) OP (16, 16, 16) 10.3 μs ± 1.17% (10.0 … 10.7 μs) 3.37 μs ± 3.1% (3.1 … 3.81 μs) 31.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 128, 16) Warps (4, 1) OP (16, 16, 16) 9.09 μs ± 1.93% (8.82 … 9.3 μs) 3.22 μs ± 4.41% (2.86 … 3.34 μs) 32.4
WMMA GEMM Float16*Float16+Float32=Float32 (128×256) · (256×128) (TT) Block (128, 128, 64) Warps (4, 2) OP (8, 32, 16) 30.0 μs ± 0.729% (29.3 … 30.5 μs) 2.62 μs ± 5.84% (2.38 … 2.86 μs) 8.13
WMMA GEMM Float16*Float16+Float16=Float16 (2048×2048) · (2048×2048) (NN) Block (128, 128, 64) Warps (4, 2) OP (8, 32, 16) 3.01 ms ± 0.506% (2.98 … 3.07 ms) 175.0 μs ± 0.333% (174.0 … 177.0 μs) 5.84
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 64, 64) Warps (1, 4) OP (16, 16, 16) 3.81 μs ± 2.83% (3.58 … 4.05 μs) 3.47 μs ± 3.83% (3.1 … 3.81 μs) 86.7
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (128, 128, 32) Warps (4, 4) OP (16, 16, 16) 10.4 μs ± 1.07% (10.0 … 10.7 μs) 3.39 μs ± 3.73% (2.86 … 3.81 μs) 28.6
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TN) OP (32, 4, 1), base shape (16, 2, 1) 58.2 μs ± 0.347% (57.7 … 58.7 μs) 3.79 μs ± 3.31% (3.58 … 4.05 μs) 6.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×128) · (128×256) (NT) Block (128, 128, 64) Warps (4, 2) OP (8, 32, 16) 17.8 μs ± 1.0% (17.4 … 18.4 μs) 2.57 μs ± 5.23% (2.15 … 2.86 μs) 12.3
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (128, 64, 16) Warps (4, 1) OP (16, 16, 16) 7.48 μs ± 1.87% (7.39 … 7.87 μs) 3.38 μs ± 3.37% (3.1 … 3.81 μs) 41.9
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 128, 64) Warps (1, 4) OP (16, 16, 16) 36.1 μs ± 0.684% (35.5 … 37.0 μs) 3.29 μs ± 3.85% (2.86 … 3.81 μs) 8.05
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 256, 16) Warps (4, 2) OP (16, 16, 16) 13.0 μs ± 1.19% (12.6 … 13.4 μs) 3.47 μs ± 4.05% (3.1 … 4.05 μs) 24.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 64, 32) Warps (4, 2) OP (16, 16, 16) 3.58 μs ± 1.87% (3.1 … 4.05 μs) 3.39 μs ± 4.02% (2.86 … 3.81 μs) 92.3
WMMA GEMM Float16*Float16+Float32=Float32 (128×256) · (256×128) (NN) Block (128, 128, 64) Warps (4, 2) OP (32, 8, 16) 10.6 μs ± 1.31% (10.5 … 11.0 μs) 2.67 μs ± 6.02% (2.38 … 2.86 μs) 22.7
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (128, 64, 16) Warps (1, 4) OP (16, 16, 16) 8.02 μs ± 2.09% (7.87 … 8.34 μs) 3.56 μs ± 3.34% (3.1 … 4.05 μs) 39.4
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 128, 64) Warps (2, 4) OP (16, 16, 16) 5.65 μs ± 2.42% (5.48 … 5.96 μs) 3.21 μs ± 4.45% (2.86 … 3.34 μs) 52.2
WMMA GEMM Float16*Float16+Float32=Float32 (2048×2048) · (2048×2048) (TT) Block (128, 128, 64) Warps (4, 2) OP (32, 8, 16) 343.0 μs ± 0.308% (340.0 … 347.0 μs) 289.0 μs ± 0.146% (288.0 … 291.0 μs) 84.9
WMMA GEMM Float16*Float16+Float16=Float16 (256×256) · (256×256) (TN) Block (128, 128, 64) Warps (4, 2) OP (32, 8, 16) 7.7 μs ± 1.76% (7.39 … 8.11 μs) 3.21 μs ± 4.51% (2.86 … 3.34 μs) 38.7
WMMA GEMM Float16*Float16+Float32=Float32 (128×128) · (128×128) (NN) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 16.4 μs ± 1.03% (16.0 … 16.7 μs) 1.97 μs ± 8.47% (1.67 … 2.38 μs) 10.4
FPU GEMM Float64*Float64=Float64 (128×128) · (128×128) (TT) OP (8, 16, 2), base shape (4, 8, 1) 139.0 μs ± 0.136% (138.0 … 140.0 μs) 28.8 μs ± 0.552% (28.6 … 29.1 μs) 20.7
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (128, 128, 64) Warps (2, 4) OP (16, 16, 16) 30.1 μs ± 0.6% (29.6 … 30.5 μs) 3.37 μs ± 3.27% (2.86 … 3.81 μs) 9.68
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 128, 32) Warps (2, 4) OP (16, 16, 16) 5.41 μs ± 2.7% (5.25 … 5.72 μs) 3.38 μs ± 3.88% (2.86 … 3.81 μs) 54.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (128, 64, 64) Warps (4, 2) OP (16, 16, 16) 5.98 μs ± 2.7% (5.72 … 6.2 μs) 3.48 μs ± 3.92% (3.1 … 3.81 μs) 54.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 256, 32) Warps (1, 4) OP (16, 16, 16) 19.2 μs ± 0.817% (18.8 … 19.6 μs) 3.28 μs ± 4.02% (2.86 … 3.81 μs) 15.2
FPU GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) OP (8, 16, 2), base shape (4, 8, 1) 43.3 μs ± 0.491% (42.7 … 43.9 μs) 3.28 μs ± 3.59% (2.86 … 3.58 μs) 6.7
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (128, 64, 64) Warps (2, 4) OP (16, 16, 16) 6.09 μs ± 2.46% (5.72 … 6.2 μs) 3.46 μs ± 4.03% (3.1 … 3.81 μs) 54.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 256, 32) Warps (4, 2) OP (16, 16, 16) 9.76 μs ± 1.33% (9.54 … 10.0 μs) 3.39 μs ± 3.78% (2.86 … 3.81 μs) 30.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (128, 128, 16) Warps (2, 4) OP (16, 16, 16) 11.6 μs ± 1.33% (11.4 … 11.9 μs) 3.35 μs ± 2.71% (3.1 … 3.81 μs) 27.1
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (256, 64, 16) Warps (1, 4) OP (16, 16, 16) 22.5 μs ± 0.69% (22.2 … 22.9 μs) 3.49 μs ± 3.69% (3.1 … 3.81 μs) 14.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (128, 128, 64) Warps (4, 4) OP (16, 16, 16) 9.96 μs ± 1.1% (9.54 … 10.3 μs) 3.38 μs ± 3.19% (3.1 … 3.81 μs) 32.5
WMMA GEMM Float16*Float16+Float32=Float32 (128×256) · (256×128) (NN) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 30.2 μs ± 0.739% (29.6 … 31.0 μs) 2.68 μs ± 5.98% (2.38 … 2.86 μs) 8.06
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (256, 64, 32) Warps (1, 4) OP (16, 16, 16) 75.5 μs ± 0.439% (74.4 … 76.5 μs) 3.31 μs ± 2.68% (2.86 … 3.58 μs) 3.85
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 256, 32) Warps (4, 1) OP (16, 16, 16) 63.5 μs ± 0.39% (62.7 … 64.4 μs) 3.22 μs ± 4.42% (2.86 … 3.58 μs) 4.56
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 256, 32) Warps (4, 4) OP (16, 16, 16) 10.2 μs ± 0.916% (9.78 … 10.5 μs) 3.29 μs ± 3.86% (2.86 … 3.58 μs) 29.3
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (128, 64, 32) Warps (4, 4) OP (16, 16, 16) 6.87 μs ± 1.55% (6.44 … 7.39 μs) 3.47 μs ± 4.09% (3.1 … 4.05 μs) 48.1
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 256, 16) Warps (2, 2) OP (16, 16, 16) 15.6 μs ± 1.01% (15.3 … 16.0 μs) 3.38 μs ± 3.58% (3.1 … 3.81 μs) 20.3
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 128, 16) Warps (2, 4) OP (16, 16, 16) 8.02 μs ± 2.07% (7.87 … 8.34 μs) 3.49 μs ± 3.67% (3.1 … 3.81 μs) 39.4
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TN) OP (4, 8, 2), base shape (4, 8, 1) 43.8 μs ± 1.16% (42.4 … 45.8 μs) 3.8 μs ± 2.96% (3.58 … 4.05 μs) 8.43
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (128, 128, 32) Warps (1, 4) OP (16, 16, 16) 13.8 μs ± 0.889% (13.4 … 14.3 μs) 3.47 μs ± 3.98% (3.1 … 4.05 μs) 23.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (128, 64, 32) Warps (4, 1) OP (16, 16, 16) 5.95 μs ± 2.68% (5.72 … 6.2 μs) 3.48 μs ± 3.9% (3.1 … 3.81 μs) 54.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 64, 16) Warps (2, 4) OP (16, 16, 16) 5.22 μs ± 3.31% (5.01 … 5.48 μs) 3.47 μs ± 3.92% (3.1 … 3.81 μs) 61.9
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 64, 32) Warps (4, 4) OP (16, 16, 16) 5.23 μs ± 3.33% (5.01 … 5.48 μs) 3.46 μs ± 4.17% (3.1 … 4.05 μs) 61.9
WMMA GEMM Float16*Float16+Float16=Float16 (512×256) · (256×512) (TT) Block (128, 128, 64) Warps (4, 2) OP (8, 32, 16) 29.7 μs ± 0.637% (29.3 … 30.3 μs) 4.47 μs ± 3.76% (4.29 … 4.77 μs) 14.6
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 128, 64) Warps (4, 1) OP (16, 16, 16) 7.96 μs ± 1.59% (7.63 … 8.34 μs) 3.37 μs ± 3.27% (3.1 … 3.81 μs) 40.6
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 256, 32) Warps (2, 2) OP (16, 16, 16) 19.4 μs ± 0.731% (19.1 … 19.8 μs) 3.3 μs ± 3.06% (2.86 … 3.81 μs) 15.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 256, 16) Warps (4, 1) OP (16, 16, 16) 15.8 μs ± 0.897% (15.5 … 16.2 μs) 3.35 μs ± 2.89% (2.86 … 3.81 μs) 18.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (256, 64, 64) Warps (1, 4) OP (16, 16, 16) 1.76 ms ± 0.202% (1.75 … 1.77 ms) 3.46 μs ± 3.86% (3.1 … 3.81 μs) 0.177
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 256, 16) Warps (2, 2) OP (16, 16, 16) 14.4 μs ± 1.13% (14.1 … 14.8 μs) 3.21 μs ± 4.44% (2.86 … 3.34 μs) 20.3
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 64, 16) Warps (2, 2) OP (16, 16, 16) 4.66 μs ± 3.22% (4.53 … 5.01 μs) 3.31 μs ± 3.69% (2.86 … 3.81 μs) 63.2
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NT) OP (32, 4, 1), base shape (16, 2, 1) 58.4 μs ± 0.335% (57.7 … 58.9 μs) 3.72 μs ± 3.74% (3.58 … 4.05 μs) 6.2
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TT) OP (8, 8, 1), base shape (4, 8, 1) 40.3 μs ± 0.439% (39.8 … 40.8 μs) 3.56 μs ± 2.53% (3.1 … 3.81 μs) 7.78
WMMA GEMM Float16*Float16+Float32=Float32 (128×256) · (256×128) (TT) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 30.6 μs ± 0.627% (30.0 … 31.2 μs) 2.62 μs ± 6.02% (2.38 … 2.86 μs) 7.94
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (256, 64, 32) Warps (2, 4) OP (16, 16, 16) 12.7 μs ± 1.36% (12.4 … 12.9 μs) 3.31 μs ± 2.87% (2.86 … 3.58 μs) 23.1
FPU GEMM Float64*Float64=Float64 (256×256) · (256×256) (NT) OP (8, 16, 2), base shape (4, 8, 1) 274.0 μs ± 0.0981% (273.0 … 275.0 μs) 109.0 μs ± 0.159% (109.0 … 110.0 μs) 40.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 128, 16) Warps (2, 2) OP (16, 16, 16) 7.76 μs ± 2.04% (7.63 … 8.11 μs) 3.35 μs ± 3.5% (2.86 … 3.81 μs) 37.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 64, 16) Warps (2, 2) OP (16, 16, 16) 5.27 μs ± 3.34% (5.01 … 5.48 μs) 3.37 μs ± 3.38% (2.86 … 3.81 μs) 57.1
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (128, 128, 32) Warps (4, 4) OP (16, 16, 16) 9.71 μs ± 1.25% (9.3 … 9.78 μs) 3.37 μs ± 3.21% (2.86 … 3.81 μs) 30.8
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 128, 32) Warps (1, 4) OP (16, 16, 16) 5.47 μs ± 3.15% (5.25 … 5.72 μs) 3.39 μs ± 3.71% (3.1 … 3.81 μs) 59.1
FPU GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) OP (8, 16, 2), base shape (4, 8, 1) 43.0 μs ± 0.521% (42.4 … 44.1 μs) 3.48 μs ± 3.97% (3.1 … 3.81 μs) 7.3
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 256, 64) Warps (2, 4) OP (16, 16, 16) 41.7 μs ± 0.503% (41.0 … 42.4 μs) 3.22 μs ± 4.3% (2.86 … 3.34 μs) 6.98
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 64, 64) Warps (4, 4) OP (16, 16, 16) 3.72 μs ± 3.72% (3.58 … 4.05 μs) 3.29 μs ± 3.96% (2.86 … 3.81 μs) 80.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 64, 32) Warps (4, 2) OP (16, 16, 16) 3.93 μs ± 3.78% (3.81 … 4.29 μs) 3.47 μs ± 3.97% (3.1 … 4.05 μs) 81.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (128, 64, 32) Warps (2, 2) OP (16, 16, 16) 5.93 μs ± 2.39% (5.72 … 6.2 μs) 3.38 μs ± 3.88% (2.86 … 3.81 μs) 50.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (128, 128, 64) Warps (2, 2) OP (16, 16, 16) 16.1 μs ± 0.921% (15.7 … 16.5 μs) 3.38 μs ± 3.48% (3.1 … 3.81 μs) 19.7
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (128, 64, 32) Warps (4, 2) OP (16, 16, 16) 6.11 μs ± 2.17% (5.72 … 6.44 μs) 3.38 μs ± 3.85% (2.86 … 3.81 μs) 50.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 256, 32) Warps (2, 2) OP (16, 16, 16) 21.5 μs ± 0.682% (21.2 … 21.9 μs) 3.37 μs ± 3.7% (2.86 … 3.81 μs) 13.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 64, 64) Warps (4, 2) OP (16, 16, 16) 3.83 μs ± 3.03% (3.58 … 4.29 μs) 3.38 μs ± 3.9% (2.86 … 3.81 μs) 80.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (128, 128, 32) Warps (4, 2) OP (16, 16, 16) 10.4 μs ± 1.3% (10.0 … 10.7 μs) 3.48 μs ± 4.07% (2.86 … 4.05 μs) 28.6
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NN) OP (4, 16, 1), base shape (4, 8, 1) 48.3 μs ± 0.673% (47.2 … 49.6 μs) 3.62 μs ± 2.74% (3.58 … 4.05 μs) 7.58
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (128, 128, 64) Warps (2, 4) OP (16, 16, 16) 30.1 μs ± 0.701% (29.6 … 30.8 μs) 3.47 μs ± 3.93% (3.1 … 4.05 μs) 10.5
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TN) OP (4, 32, 1), base shape (1, 32, 1) 29.0 μs ± 0.575% (28.8 … 29.3 μs) 3.8 μs ± 3.02% (3.58 … 4.05 μs) 12.4
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TN) OP (8, 16, 2), base shape (4, 8, 1) 42.0 μs ± 0.452% (41.5 … 42.7 μs) 3.8 μs ± 3.08% (3.58 … 4.05 μs) 8.62
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (256, 64, 32) Warps (4, 2) OP (16, 16, 16) 11.9 μs ± 1.41% (11.7 … 12.2 μs) 3.31 μs ± 2.83% (2.86 … 3.81 μs) 24.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 64, 16) Warps (4, 1) OP (16, 16, 16) 5.47 μs ± 3.11% (5.25 … 5.72 μs) 3.34 μs ± 3.62% (2.86 … 3.81 μs) 54.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 64, 32) Warps (4, 2) OP (16, 16, 16) 3.82 μs ± 2.76% (3.58 … 4.29 μs) 3.36 μs ± 3.84% (2.86 … 3.81 μs) 80.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (128, 64, 16) Warps (1, 4) OP (16, 16, 16) 8.19 μs ± 1.61% (7.87 … 8.58 μs) 3.35 μs ± 3.6% (2.86 … 3.81 μs) 36.4
FPU GEMM Float32*Float32=Float32 (2048×2048) · (2048×2048) (NN) OP (8, 16, 2), base shape (4, 8, 1) 30.0 ms ± 0.0224% (30.0 … 30.0 ms) 869.0 μs ± 0.591% (852.0 … 875.0 μs) 2.84
WMMA GEMM Float16*Float16+Float32=Float32 (256×128) · (128×256) (TN) Block (128, 128, 64) Warps (4, 2) OP (8, 32, 16) 17.9 μs ± 0.793% (17.6 … 18.4 μs) 2.48 μs ± 6.51% (2.15 … 2.62 μs) 12.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 64, 32) Warps (2, 2) OP (16, 16, 16) 3.92 μs ± 3.58% (3.81 … 4.29 μs) 3.47 μs ± 3.89% (3.1 … 3.81 μs) 81.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (128, 128, 16) Warps (1, 4) OP (16, 16, 16) 12.3 μs ± 1.16% (11.9 … 12.6 μs) 3.31 μs ± 3.7% (2.86 … 3.58 μs) 24.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (256, 64, 16) Warps (4, 1) OP (16, 16, 16) 17.1 μs ± 0.538% (16.7 … 17.4 μs) 3.5 μs ± 3.59% (3.1 … 3.81 μs) 18.6
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TN) OP (4, 8, 1), base shape (4, 8, 1) 54.8 μs ± 0.713% (53.6 … 56.5 μs) 3.8 μs ± 3.17% (3.58 … 4.05 μs) 6.67
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 64, 16) Warps (2, 4) OP (16, 16, 16) 5.45 μs ± 2.94% (5.25 … 5.72 μs) 3.35 μs ± 3.51% (2.86 … 3.81 μs) 54.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 256, 64) Warps (1, 4) OP (16, 16, 16) 23.8 μs ± 0.596% (23.4 … 24.1 μs) 3.39 μs ± 3.92% (3.1 … 3.81 μs) 13.3
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 256, 32) Warps (2, 4) OP (16, 16, 16) 9.77 μs ± 1.29% (9.54 … 10.0 μs) 3.29 μs ± 3.94% (2.86 … 3.81 μs) 30.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 64, 32) Warps (4, 4) OP (16, 16, 16) 5.23 μs ± 3.37% (5.01 … 5.48 μs) 3.39 μs ± 3.66% (3.1 … 3.81 μs) 61.9
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (128, 64, 16) Warps (2, 4) OP (16, 16, 16) 8.08 μs ± 2.05% (7.87 … 8.34 μs) 3.38 μs ± 3.4% (3.1 … 3.81 μs) 39.4
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (256, 64, 64) Warps (4, 4) OP (16, 16, 16) 11.4 μs ± 1.43% (11.2 … 11.7 μs) 3.48 μs ± 3.85% (3.1 … 3.81 μs) 27.7
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 128, 64) Warps (2, 2) OP (16, 16, 16) 31.2 μs ± 0.618% (30.5 … 31.9 μs) 3.22 μs ± 4.37% (2.86 … 3.34 μs) 9.38
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 64, 32) Warps (4, 1) OP (16, 16, 16) 3.92 μs ± 3.62% (3.81 … 4.29 μs) 3.38 μs ± 3.71% (2.86 … 3.81 μs) 75.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 256, 16) Warps (2, 4) OP (16, 16, 16) 12.0 μs ± 1.38% (11.7 … 12.2 μs) 3.34 μs ± 2.77% (2.86 … 3.81 μs) 24.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (128, 64, 32) Warps (4, 2) OP (16, 16, 16) 6.0 μs ± 2.83% (5.72 … 6.2 μs) 3.3 μs ± 2.95% (2.86 … 3.58 μs) 50.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (256, 64, 32) Warps (4, 1) OP (16, 16, 16) 24.6 μs ± 0.493% (24.3 … 25.0 μs) 3.48 μs ± 3.93% (3.1 … 3.81 μs) 12.7
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (128, 64, 16) Warps (2, 2) OP (16, 16, 16) 7.01 μs ± 1.83% (6.68 … 7.39 μs) 3.33 μs ± 3.97% (2.86 … 3.81 μs) 42.9
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (128, 64, 64) Warps (4, 4) OP (16, 16, 16) 6.23 μs ± 2.37% (5.96 … 6.44 μs) 3.46 μs ± 4.1% (3.1 … 4.05 μs) 52.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (128, 128, 32) Warps (2, 4) OP (16, 16, 16) 9.61 μs ± 1.68% (9.3 … 9.78 μs) 3.22 μs ± 4.34% (2.86 … 3.34 μs) 30.8
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 256, 16) Warps (4, 1) OP (16, 16, 16) 17.1 μs ± 0.759% (16.7 … 17.2 μs) 3.35 μs ± 3.8% (2.86 … 3.81 μs) 17.1
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 128, 32) Warps (2, 2) OP (16, 16, 16) 5.53 μs ± 3.11% (5.25 … 5.72 μs) 3.29 μs ± 3.95% (2.86 … 3.81 μs) 54.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (128, 128, 32) Warps (2, 4) OP (16, 16, 16) 10.4 μs ± 1.11% (10.0 … 10.7 μs) 3.46 μs ± 3.87% (3.1 … 3.81 μs) 31.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 256, 32) Warps (4, 1) OP (16, 16, 16) 68.8 μs ± 0.409% (67.7 … 69.9 μs) 3.38 μs ± 3.68% (2.86 … 3.81 μs) 4.23
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 128, 16) Warps (4, 2) OP (16, 16, 16) 8.19 μs ± 1.65% (7.87 … 8.58 μs) 3.49 μs ± 3.82% (3.1 … 3.81 μs) 39.4
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (128, 64, 64) Warps (2, 2) OP (16, 16, 16) 36.9 μs ± 0.691% (36.2 … 37.9 μs) 3.38 μs ± 3.91% (2.86 … 3.81 μs) 7.89
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (256, 64, 16) Warps (2, 2) OP (16, 16, 16) 13.5 μs ± 1.09% (13.1 … 13.6 μs) 3.34 μs ± 2.72% (2.86 … 3.81 μs) 21.8
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (128, 64, 32) Warps (4, 1) OP (16, 16, 16) 5.44 μs ± 2.99% (5.25 … 5.72 μs) 3.22 μs ± 4.5% (2.86 … 3.34 μs) 54.5
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TT) OP (32, 4, 1), base shape (16, 2, 1) 58.4 μs ± 0.328% (57.9 … 58.9 μs) 3.55 μs ± 2.49% (3.1 … 3.81 μs) 5.35
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (128, 128, 16) Warps (4, 2) OP (16, 16, 16) 11.4 μs ± 1.45% (11.2 … 11.7 μs) 3.34 μs ± 2.82% (2.86 … 3.81 μs) 25.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 256, 64) Warps (2, 2) OP (16, 16, 16) 29.3 μs ± 0.823% (28.6 … 30.3 μs) 3.29 μs ± 3.71% (2.86 … 3.81 μs) 10.0
WMMA GEMM Float16*Float16+Float16=Float16 (256×512) · (512×256) (NN) Block (128, 128, 64) Warps (4, 2) OP (32, 8, 16) 11.2 μs ± 1.36% (11.0 … 11.7 μs) 4.82 μs ± 3.65% (4.53 … 5.25 μs) 41.3
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 128, 16) Warps (1, 4) OP (16, 16, 16) 8.94 μs ± 1.57% (8.58 … 9.3 μs) 3.22 μs ± 4.31% (2.86 … 3.34 μs) 33.3
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (128, 128, 32) Warps (1, 4) OP (16, 16, 16) 11.7 μs ± 1.44% (11.4 … 11.9 μs) 3.37 μs ± 3.78% (2.86 … 3.81 μs) 25.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 64, 32) Warps (4, 4) OP (16, 16, 16) 5.52 μs ± 3.15% (5.25 … 5.72 μs) 3.37 μs ± 4.0% (2.86 … 3.81 μs) 54.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (128, 128, 64) Warps (2, 4) OP (16, 16, 16) 29.4 μs ± 0.571% (29.1 … 29.8 μs) 3.29 μs ± 4.14% (2.86 … 3.81 μs) 9.84
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (128, 128, 64) Warps (2, 4) OP (16, 16, 16) 30.8 μs ± 0.541% (30.3 … 31.5 μs) 3.31 μs ± 2.94% (2.86 … 3.81 μs) 9.45
WMMA GEMM Float16*Float16+Float16=Float16 (256×512) · (512×256) (TT) Block (128, 128, 64) Warps (4, 2) OP (8, 32, 16) 50.3 μs ± 0.418% (49.8 … 51.0 μs) 4.85 μs ± 3.12% (4.53 … 5.25 μs) 9.09
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 128, 32) Warps (2, 2) OP (16, 16, 16) 5.89 μs ± 2.04% (5.48 … 6.2 μs) 3.47 μs ± 3.9% (3.1 … 4.05 μs) 56.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (128, 128, 32) Warps (4, 4) OP (16, 16, 16) 9.9 μs ± 1.49% (9.54 … 10.0 μs) 3.39 μs ± 3.67% (2.86 … 3.81 μs) 30.0
WMMA GEMM Float16*Float16+Float16=Float16 (256×256) · (256×256) (TT) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 27.4 μs ± 0.589% (26.9 … 27.9 μs) 3.31 μs ± 2.87% (2.86 … 3.81 μs) 10.6
FPU GEMM Float16*Float16=Float32 (128×128) · (128×128) (TT) OP (8, 16, 2), base shape (4, 8, 1) 16.5 μs ± 1.08% (16.0 … 16.9 μs) 2.17 μs ± 7.87% (1.91 … 2.38 μs) 11.9
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (128, 128, 16) Warps (1, 4) OP (16, 16, 16) 14.6 μs ± 0.76% (14.3 … 15.0 μs) 3.21 μs ± 4.36% (2.86 … 3.34 μs) 20.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (256, 64, 32) Warps (2, 2) OP (16, 16, 16) 20.2 μs ± 0.737% (19.8 … 20.5 μs) 3.29 μs ± 3.94% (2.86 … 3.81 μs) 14.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (256, 64, 64) Warps (2, 2) OP (16, 16, 16) 34.5 μs ± 0.61% (33.9 … 35.3 μs) 3.49 μs ± 3.95% (3.1 … 4.05 μs) 9.15
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 128, 64) Warps (2, 4) OP (16, 16, 16) 5.65 μs ± 2.33% (5.48 … 5.96 μs) 3.38 μs ± 3.98% (2.86 … 3.81 μs) 52.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (256, 64, 64) Warps (4, 1) OP (16, 16, 16) 19.2 μs ± 0.807% (18.8 … 19.6 μs) 3.29 μs ± 3.97% (2.86 … 3.81 μs) 15.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 256, 64) Warps (2, 2) OP (16, 16, 16) 20.9 μs ± 0.712% (20.5 … 21.5 μs) 3.38 μs ± 3.26% (3.1 … 3.81 μs) 15.1
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (256, 64, 64) Warps (4, 1) OP (16, 16, 16) 24.7 μs ± 0.597% (24.3 … 25.0 μs) 3.3 μs ± 2.96% (2.86 … 3.58 μs) 11.8
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 64, 32) Warps (4, 4) OP (16, 16, 16) 5.42 μs ± 2.86% (5.25 … 5.72 μs) 3.38 μs ± 3.34% (3.1 … 3.81 μs) 59.1
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NN) OP (4, 8, 2), base shape (4, 8, 1) 40.7 μs ± 0.471% (40.3 … 41.2 μs) 3.61 μs ± 2.53% (3.58 … 4.05 μs) 8.88
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 128, 32) Warps (4, 4) OP (16, 16, 16) 6.26 μs ± 2.57% (5.96 … 6.44 μs) 3.28 μs ± 3.72% (2.86 … 3.58 μs) 48.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 64, 64) Warps (2, 4) OP (16, 16, 16) 3.57 μs ± 2.24% (3.1 … 3.81 μs) 3.39 μs ± 4.15% (2.86 … 3.81 μs) 92.3
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NT) OP (4, 32, 1), base shape (2, 16, 1) 61.1 μs ± 0.353% (60.6 … 61.8 μs) 3.72 μs ± 3.79% (3.58 … 4.05 μs) 5.91
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 64, 64) Warps (4, 1) OP (16, 16, 16) 3.57 μs ± 2.4% (3.1 … 4.05 μs) 3.38 μs ± 3.96% (2.86 … 3.81 μs) 92.3
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (256, 64, 16) Warps (4, 2) OP (16, 16, 16) 11.8 μs ± 1.07% (11.4 … 12.2 μs) 3.48 μs ± 4.0% (3.1 … 4.05 μs) 27.1
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 256, 16) Warps (4, 1) OP (16, 16, 16) 16.4 μs ± 0.847% (16.0 … 16.7 μs) 3.48 μs ± 3.98% (3.1 … 4.05 μs) 19.4
WMMA GEMM Float16*Float16+Float32=Float32 (256×128) · (128×256) (TT) Block (128, 128, 64) Warps (4, 2) OP (32, 8, 16) 7.65 μs ± 1.69% (7.39 … 7.87 μs) 2.51 μs ± 6.07% (2.15 … 2.62 μs) 29.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 256, 16) Warps (4, 4) OP (16, 16, 16) 12.1 μs ± 1.42% (11.7 … 12.4 μs) 3.21 μs ± 4.55% (2.86 … 3.34 μs) 24.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 29.9 μs ± 0.75% (29.3 … 30.5 μs) 3.38 μs ± 3.44% (3.1 … 3.81 μs) 10.6
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 256, 16) Warps (2, 2) OP (16, 16, 16) 13.2 μs ± 1.22% (12.9 … 13.4 μs) 3.55 μs ± 3.45% (3.1 … 4.05 μs) 24.1
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 256, 64) Warps (4, 2) OP (16, 16, 16) 13.1 μs ± 1.17% (12.9 … 13.4 μs) 3.31 μs ± 2.92% (2.86 … 3.81 μs) 22.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (128, 64, 16) Warps (4, 1) OP (16, 16, 16) 7.65 μs ± 1.71% (7.39 … 7.87 μs) 3.49 μs ± 3.77% (3.1 … 3.81 μs) 41.9
FPU GEMM Float64*Float64=Float64 (2048×2048) · (2048×2048) (TN) OP (8, 16, 2), base shape (4, 8, 1) 59.8 ms ± 0.173% (59.7 … 60.1 ms) 41.0 ms ± 0.015% (41.0 … 41.0 ms) 68.7
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 128, 16) Warps (4, 1) OP (16, 16, 16) 9.84 μs ± 1.58% (9.54 … 10.0 μs) 3.38 μs ± 3.41% (3.1 … 3.81 μs) 32.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (128, 64, 64) Warps (4, 1) OP (16, 16, 16) 35.9 μs ± 0.731% (35.3 … 36.7 μs) 3.37 μs ± 3.79% (2.86 … 3.81 μs) 8.11
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 256, 16) Warps (4, 2) OP (16, 16, 16) 15.3 μs ± 1.04% (15.0 … 15.7 μs) 3.38 μs ± 3.42% (3.1 … 3.81 μs) 20.6
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (256, 64, 64) Warps (1, 4) OP (16, 16, 16) 1.76 ms ± 0.163% (1.75 … 1.77 ms) 3.48 μs ± 4.13% (3.1 … 4.05 μs) 0.177
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (128, 64, 16) Warps (4, 1) OP (16, 16, 16) 8.54 μs ± 1.99% (8.34 … 8.82 μs) 3.55 μs ± 3.41% (3.1 … 4.05 μs) 37.1
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 128, 32) Warps (4, 1) OP (16, 16, 16) 5.85 μs ± 2.37% (5.48 … 5.96 μs) 3.37 μs ± 3.84% (2.86 … 3.81 μs) 52.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×128) · (128×256) (NN) Block (128, 128, 64) Warps (4, 2) OP (8, 32, 16) 18.2 μs ± 0.908% (17.9 … 18.6 μs) 2.52 μs ± 5.62% (2.15 … 2.86 μs) 12.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 128, 64) Warps (1, 4) OP (16, 16, 16) 38.3 μs ± 0.689% (37.4 … 39.1 μs) 3.38 μs ± 3.31% (3.1 … 3.81 μs) 8.28
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (128, 64, 16) Warps (2, 2) OP (16, 16, 16) 8.88 μs ± 1.84% (8.58 … 9.06 μs) 3.21 μs ± 4.57% (2.86 … 3.58 μs) 33.3
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (256, 64, 64) Warps (4, 4) OP (16, 16, 16) 11.4 μs ± 1.46% (11.2 … 11.7 μs) 3.3 μs ± 2.97% (2.86 … 3.58 μs) 25.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (256, 64, 16) Warps (4, 4) OP (16, 16, 16) 12.4 μs ± 1.25% (11.9 … 12.6 μs) 3.32 μs ± 3.55% (2.86 … 3.81 μs) 24.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (128, 128, 16) Warps (4, 2) OP (16, 16, 16) 12.3 μs ± 1.06% (11.9 … 12.6 μs) 3.35 μs ± 3.71% (2.86 … 3.81 μs) 24.0
WMMA GEMM Float16*Float16+Float16=Float16 (256×256) · (256×256) (TN) Block (128, 128, 64) Warps (4, 2) OP (8, 32, 16) 26.5 μs ± 0.711% (26.0 … 27.2 μs) 3.2 μs ± 4.49% (2.86 … 3.34 μs) 11.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (128, 128, 16) Warps (4, 4) OP (16, 16, 16) 11.4 μs ± 1.48% (11.0 … 11.9 μs) 3.47 μs ± 3.9% (3.1 … 3.81 μs) 28.3
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (128, 64, 16) Warps (2, 4) OP (16, 16, 16) 7.84 μs ± 2.04% (7.63 … 8.11 μs) 3.35 μs ± 3.79% (2.86 … 3.81 μs) 37.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 64, 64) Warps (4, 2) OP (16, 16, 16) 3.72 μs ± 3.81% (3.58 … 4.05 μs) 3.47 μs ± 3.88% (3.1 … 3.81 μs) 86.7
WMMA GEMM Float16*Float16+Float16=Float16 (2048×2048) · (2048×2048) (NN) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 3.01 ms ± 0.478% (2.98 … 3.07 ms) 175.0 μs ± 0.308% (174.0 … 178.0 μs) 5.85
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TN) OP (4, 32, 1), base shape (2, 16, 1) 61.0 μs ± 0.329% (60.6 … 61.5 μs) 3.8 μs ± 3.05% (3.58 … 4.05 μs) 5.91
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (128, 128, 32) Warps (4, 4) OP (16, 16, 16) 10.6 μs ± 1.24% (10.5 … 11.0 μs) 3.47 μs ± 3.97% (3.1 … 4.05 μs) 29.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (128, 128, 16) Warps (2, 4) OP (16, 16, 16) 11.4 μs ± 1.42% (11.2 … 11.7 μs) 3.48 μs ± 3.91% (3.1 … 3.81 μs) 27.7
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 128, 32) Warps (4, 4) OP (16, 16, 16) 6.6 μs ± 1.88% (6.2 … 6.91 μs) 3.38 μs ± 3.81% (2.86 … 3.81 μs) 46.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (128, 128, 64) Warps (1, 4) OP (16, 16, 16) 18.7 μs ± 1.05% (18.4 … 19.3 μs) 3.29 μs ± 4.12% (2.86 … 3.81 μs) 15.6
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 64, 64) Warps (4, 1) OP (16, 16, 16) 3.49 μs ± 3.85% (3.1 … 3.81 μs) 3.38 μs ± 3.17% (3.1 … 3.81 μs) 100.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 64, 16) Warps (2, 2) OP (16, 16, 16) 5.06 μs ± 3.22% (4.77 … 5.25 μs) 3.35 μs ± 3.45% (2.86 … 3.81 μs) 60.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 31.0 μs ± 0.478% (30.5 … 31.7 μs) 3.39 μs ± 3.84% (2.86 … 3.81 μs) 9.38
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 256, 16) Warps (1, 4) OP (16, 16, 16) 13.8 μs ± 0.544% (13.4 … 14.3 μs) 3.49 μs ± 3.8% (3.1 … 3.81 μs) 23.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 128, 32) Warps (4, 2) OP (16, 16, 16) 5.97 μs ± 2.62% (5.72 … 6.2 μs) 3.47 μs ± 3.94% (3.1 … 3.81 μs) 54.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 128, 16) Warps (4, 1) OP (16, 16, 16) 8.22 μs ± 1.7% (7.87 … 8.58 μs) 3.49 μs ± 3.59% (3.1 … 3.81 μs) 39.4
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 64, 16) Warps (1, 4) OP (16, 16, 16) 5.85 μs ± 2.4% (5.48 … 5.96 μs) 3.55 μs ± 3.55% (3.1 … 4.05 μs) 56.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (128, 64, 32) Warps (2, 2) OP (16, 16, 16) 5.94 μs ± 2.48% (5.72 … 6.2 μs) 3.47 μs ± 3.76% (3.1 … 3.81 μs) 54.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 64, 32) Warps (4, 4) OP (16, 16, 16) 5.3 μs ± 3.21% (5.01 … 5.48 μs) 3.29 μs ± 3.87% (2.86 … 3.81 μs) 57.1
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (256, 64, 16) Warps (4, 2) OP (16, 16, 16) 14.9 μs ± 0.984% (14.8 … 15.3 μs) 3.56 μs ± 3.58% (3.1 … 4.05 μs) 21.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 256, 64) Warps (4, 4) OP (16, 16, 16) 10.3 μs ± 1.21% (10.0 … 10.7 μs) 3.22 μs ± 4.4% (2.86 … 3.34 μs) 28.6
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (128, 128, 64) Warps (2, 2) OP (16, 16, 16) 17.1 μs ± 0.82% (16.7 … 17.4 μs) 3.37 μs ± 3.71% (2.86 … 3.81 μs) 17.1
WMMA GEMM Float16*Float16+Float16=Float16 (512×512) · (512×512) (TT) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 56.2 μs ± 0.439% (55.6 … 57.2 μs) 6.5 μs ± 2.26% (6.2 … 6.68 μs) 11.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 64, 32) Warps (2, 2) OP (16, 16, 16) 3.82 μs ± 2.42% (3.58 … 4.05 μs) 3.47 μs ± 3.95% (3.1 … 4.05 μs) 86.7
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (128, 64, 64) Warps (1, 4) OP (16, 16, 16) 8.85 μs ± 1.98% (8.58 … 9.06 μs) 3.47 μs ± 3.9% (3.1 … 4.05 μs) 36.1
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (128, 128, 16) Warps (2, 2) OP (16, 16, 16) 13.0 μs ± 0.923% (12.6 … 13.4 μs) 3.49 μs ± 3.73% (3.1 … 3.81 μs) 24.5
WMMA GEMM Float16*Float16+Float16=Float16 (512×512) · (512×512) (NT) Block (128, 128, 64) Warps (4, 2) OP (32, 8, 16) 11.3 μs ± 1.49% (11.0 … 11.7 μs) 6.4 μs ± 1.87% (5.96 … 6.68 μs) 54.3
WMMA GEMM Float16*Float16+Float32=Float32 (256×128) · (128×256) (NN) Block (128, 128, 64) Warps (4, 2) OP (32, 8, 16) 7.72 μs ± 1.76% (7.63 … 8.11 μs) 2.53 μs ± 5.51% (2.15 … 2.86 μs) 28.1
WMMA GEMM Float16*Float16+Float32=Float32 (128×128) · (128×128) (NN) Block (128, 128, 64) Warps (4, 2) OP (32, 8, 16) 7.57 μs ± 2.02% (7.39 … 7.87 μs) 1.97 μs ± 8.62% (1.67 … 2.15 μs) 22.6
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (128, 64, 64) Warps (2, 2) OP (16, 16, 16) 37.9 μs ± 0.765% (37.0 … 38.9 μs) 3.48 μs ± 4.09% (3.1 … 4.05 μs) 8.39
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (256, 64, 16) Warps (4, 1) OP (16, 16, 16) 15.8 μs ± 1.01% (15.5 … 16.0 μs) 3.22 μs ± 4.37% (2.86 … 3.34 μs) 18.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (128, 128, 16) Warps (2, 4) OP (16, 16, 16) 12.3 μs ± 1.09% (12.2 … 12.6 μs) 3.55 μs ± 3.25% (3.1 … 4.05 μs) 25.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 256, 32) Warps (4, 4) OP (16, 16, 16) 10.9 μs ± 1.33% (10.7 … 11.2 μs) 3.38 μs ± 3.79% (2.86 … 3.81 μs) 26.7
WMMA GEMM Float16*Float16+Float16=Float16 (256×256) · (256×256) (TN) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 27.2 μs ± 0.587% (26.7 … 27.7 μs) 3.2 μs ± 4.35% (2.86 … 3.34 μs) 10.7
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TN) OP (16, 16, 1), base shape (8, 4, 1) 41.3 μs ± 0.474% (40.8 … 42.0 μs) 3.79 μs ± 3.25% (3.58 … 4.05 μs) 8.77
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 256, 16) Warps (4, 4) OP (16, 16, 16) 12.1 μs ± 1.55% (11.7 … 12.6 μs) 3.34 μs ± 2.48% (3.1 … 3.58 μs) 26.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 128, 64) Warps (4, 4) OP (16, 16, 16) 5.82 μs ± 2.63% (5.48 … 5.96 μs) 3.38 μs ± 3.27% (3.1 … 3.81 μs) 56.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (256, 64, 16) Warps (1, 4) OP (16, 16, 16) 17.4 μs ± 0.48% (16.9 … 17.6 μs) 3.47 μs ± 3.97% (3.1 … 4.05 μs) 18.3
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 256, 16) Warps (4, 1) OP (16, 16, 16) 17.3 μs ± 0.734% (16.9 … 17.6 μs) 3.55 μs ± 3.48% (3.1 … 4.05 μs) 18.3
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (128, 64, 16) Warps (2, 4) OP (16, 16, 16) 7.49 μs ± 1.94% (7.15 … 7.87 μs) 3.21 μs ± 4.54% (2.86 … 3.34 μs) 40.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (128, 128, 16) Warps (2, 2) OP (16, 16, 16) 12.5 μs ± 1.16% (12.2 … 12.9 μs) 3.21 μs ± 4.56% (2.86 … 3.34 μs) 23.5
WMMA GEMM Float16*Float16+Float16=Float16 (2048×2048) · (2048×2048) (NT) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 2.99 ms ± 0.46% (2.96 … 3.05 ms) 177.0 μs ± 1.56% (173.0 … 183.0 μs) 5.85
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (256, 64, 32) Warps (4, 4) OP (16, 16, 16) 10.2 μs ± 1.23% (9.78 … 10.5 μs) 3.22 μs ± 4.22% (2.86 … 3.34 μs) 29.3
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (256, 64, 64) Warps (2, 4) OP (16, 16, 16) 15.4 μs ± 1.06% (15.0 … 16.0 μs) 3.38 μs ± 3.96% (2.86 … 3.81 μs) 19.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (128, 128, 64) Warps (4, 4) OP (16, 16, 16) 10.8 μs ± 0.929% (10.5 … 11.2 μs) 3.31 μs ± 2.88% (2.86 … 3.58 μs) 27.3
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 256, 32) Warps (1, 4) OP (16, 16, 16) 17.4 μs ± 0.492% (16.9 … 17.9 μs) 3.22 μs ± 4.41% (2.86 … 3.34 μs) 16.9
WMMA GEMM Float16*Float16+Float32=Float32 (2048×2048) · (2048×2048) (NN) Block (128, 128, 64) Warps (4, 2) OP (8, 32, 16) 3.14 ms ± 0.446% (3.11 … 3.18 ms) 290.0 μs ± 0.173% (288.0 … 292.0 μs) 9.27
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 128, 16) Warps (2, 4) OP (16, 16, 16) 6.71 μs ± 1.76% (6.44 … 6.91 μs) 3.32 μs ± 3.71% (2.86 … 3.81 μs) 44.4
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (128, 64, 16) Warps (2, 4) OP (16, 16, 16) 7.47 μs ± 1.8% (7.39 … 7.87 μs) 3.49 μs ± 3.62% (3.1 … 3.81 μs) 41.9
WMMA GEMM Float16*Float16+Float16=Float16 (256×256) · (256×256) (TT) Block (128, 128, 64) Warps (4, 2) OP (8, 32, 16) 26.9 μs ± 0.678% (26.5 … 27.4 μs) 3.3 μs ± 3.05% (2.86 … 3.81 μs) 10.8
WMMA GEMM Float16*Float16+Float16=Float16 (512×256) · (256×512) (NN) Block (128, 128, 64) Warps (4, 2) OP (8, 32, 16) 29.5 μs ± 0.707% (29.1 … 30.0 μs) 4.33 μs ± 3.24% (4.05 … 4.77 μs) 13.9
FPU GEMM Float32*Float32=Float32 (2048×2048) · (2048×2048) (TT) OP (8, 16, 2), base shape (4, 8, 1) 30.1 ms ± 0.0235% (30.1 … 30.1 ms) 946.0 μs ± 0.693% (913.0 … 952.0 μs) 3.04
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 64, 16) Warps (4, 1) OP (16, 16, 16) 5.53 μs ± 3.1% (5.25 … 5.72 μs) 3.35 μs ± 2.71% (2.86 … 3.81 μs) 54.5
FPU GEMM Float16*Float16=Float32 (2048×2048) · (2048×2048) (NN) OP (8, 16, 2), base shape (4, 8, 1) 15.7 ms ± 0.0711% (15.6 … 15.7 ms) 290.0 μs ± 0.0883% (289.0 … 291.0 μs) 1.85
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (128, 64, 32) Warps (1, 4) OP (16, 16, 16) 5.48 μs ± 3.17% (5.25 … 5.72 μs) 3.39 μs ± 3.92% (2.86 … 3.81 μs) 54.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 256, 16) Warps (4, 2) OP (16, 16, 16) 12.9 μs ± 1.25% (12.6 … 13.1 μs) 3.35 μs ± 3.85% (2.86 … 3.81 μs) 22.6
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (128, 64, 32) Warps (4, 1) OP (16, 16, 16) 5.42 μs ± 2.87% (5.25 … 5.72 μs) 3.38 μs ± 3.2% (3.1 … 3.81 μs) 59.1
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 64, 16) Warps (4, 2) OP (16, 16, 16) 4.7 μs ± 3.47% (4.53 … 5.01 μs) 3.47 μs ± 4.14% (3.1 … 4.05 μs) 68.4
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (128, 64, 64) Warps (1, 4) OP (16, 16, 16) 8.74 μs ± 1.8% (8.34 … 9.06 μs) 3.31 μs ± 2.93% (2.86 … 3.81 μs) 34.3
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (128, 64, 32) Warps (4, 2) OP (16, 16, 16) 5.48 μs ± 3.09% (5.25 … 5.72 μs) 3.38 μs ± 3.32% (3.1 … 3.81 μs) 59.1
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 256, 64) Warps (4, 1) OP (16, 16, 16) 1.37 ms ± 0.175% (1.36 … 1.38 ms) 3.47 μs ± 3.86% (3.1 … 3.81 μs) 0.228
FPU GEMM Float16*Float16=Float32 (128×128) · (128×128) (TN) OP (8, 16, 2), base shape (4, 8, 1) 16.4 μs ± 1.01% (16.0 … 16.7 μs) 2.07 μs ± 6.92% (1.67 … 2.38 μs) 10.4
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (128, 128, 32) Warps (2, 2) OP (16, 16, 16) 9.8 μs ± 1.48% (9.54 … 10.0 μs) 3.29 μs ± 4.03% (2.86 … 3.81 μs) 30.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 128, 64) Warps (4, 2) OP (16, 16, 16) 5.97 μs ± 2.71% (5.72 … 6.2 μs) 3.38 μs ± 3.95% (2.86 … 3.81 μs) 50.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (128, 128, 32) Warps (4, 4) OP (16, 16, 16) 10.4 μs ± 1.25% (10.0 … 10.7 μs) 3.31 μs ± 2.98% (2.86 … 3.81 μs) 28.6
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 64, 64) Warps (4, 1) OP (16, 16, 16) 3.74 μs ± 3.76% (3.58 … 4.05 μs) 3.31 μs ± 2.88% (2.86 … 3.58 μs) 80.0
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NN) OP (32, 4, 1), base shape (32, 1, 1) 108.0 μs ± 0.521% (107.0 … 110.0 μs) 3.6 μs ± 2.13% (3.58 … 4.05 μs) 3.34
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 256, 32) Warps (4, 2) OP (16, 16, 16) 9.85 μs ± 1.56% (9.54 … 10.0 μs) 3.38 μs ± 3.2% (3.1 … 3.58 μs) 32.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (256, 64, 16) Warps (2, 4) OP (16, 16, 16) 14.4 μs ± 0.792% (14.1 … 14.8 μs) 3.38 μs ± 3.28% (3.1 … 3.81 μs) 22.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 64, 32) Warps (2, 2) OP (16, 16, 16) 3.61 μs ± 2.47% (3.58 … 4.05 μs) 3.29 μs ± 3.94% (2.86 … 3.81 μs) 80.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (256, 64, 16) Warps (1, 4) OP (16, 16, 16) 20.1 μs ± 0.788% (19.8 … 20.3 μs) 3.38 μs ± 3.27% (3.1 … 3.81 μs) 15.7
WMMA GEMM Float16*Float16+Float32=Float32 (128×128) · (128×128) (NT) Block (128, 128, 64) Warps (4, 2) OP (8, 32, 16) 17.4 μs ± 0.913% (16.9 … 18.1 μs) 1.99 μs ± 8.29% (1.67 … 2.15 μs) 9.86
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (128, 64, 64) Warps (2, 2) OP (16, 16, 16) 36.5 μs ± 0.692% (35.8 … 37.2 μs) 3.3 μs ± 3.56% (2.86 … 3.58 μs) 8.0
FPU GEMM Float64*Float64=Float64 (128×128) · (128×128) (NT) OP (8, 16, 2), base shape (4, 8, 1) 138.0 μs ± 0.178% (138.0 … 139.0 μs) 28.8 μs ± 0.572% (28.6 … 29.1 μs) 20.8
WMMA GEMM Float16*Float16+Float16=Float16 (512×512) · (512×512) (NT) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 56.4 μs ± 0.533% (55.6 … 57.5 μs) 6.41 μs ± 1.78% (5.96 … 6.68 μs) 10.7
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (128, 128, 32) Warps (4, 1) OP (16, 16, 16) 13.1 μs ± 1.32% (12.6 … 13.4 μs) 3.47 μs ± 3.9% (3.1 … 3.81 μs) 24.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 128, 64) Warps (2, 4) OP (16, 16, 16) 6.08 μs ± 2.42% (5.72 … 6.2 μs) 3.47 μs ± 3.82% (3.1 … 3.81 μs) 54.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 30.8 μs ± 0.541% (30.3 … 31.2 μs) 3.3 μs ± 2.88% (2.86 … 3.58 μs) 9.45
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (128, 64, 32) Warps (4, 4) OP (16, 16, 16) 6.27 μs ± 2.47% (5.96 … 6.44 μs) 3.3 μs ± 3.02% (2.86 … 3.58 μs) 48.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 64, 64) Warps (4, 2) OP (16, 16, 16) 3.47 μs ± 3.95% (3.1 … 3.81 μs) 3.37 μs ± 3.27% (3.1 … 3.81 μs) 100.0
WMMA GEMM Float16*Float16+Float16=Float16 (256×512) · (512×256) (TT) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 51.4 μs ± 0.391% (50.8 … 52.0 μs) 4.84 μs ± 3.07% (4.53 … 5.25 μs) 8.92
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 128, 32) Warps (4, 4) OP (16, 16, 16) 6.96 μs ± 1.46% (6.68 … 7.39 μs) 3.46 μs ± 3.97% (3.1 … 3.81 μs) 46.4
WMMA GEMM Float16*Float16+Float16=Float16 (512×256) · (256×512) (TN) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 30.8 μs ± 0.671% (30.0 … 31.2 μs) 4.46 μs ± 3.69% (4.29 … 4.77 μs) 14.3
WMMA GEMM Float16*Float16+Float16=Float16 (512×256) · (256×512) (TT) Block (128, 128, 64) Warps (4, 2) OP (32, 8, 16) 7.91 μs ± 2.22% (7.63 … 8.34 μs) 4.47 μs ± 3.73% (4.29 … 4.77 μs) 56.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 64, 64) Warps (4, 4) OP (16, 16, 16) 3.86 μs ± 2.58% (3.58 … 4.29 μs) 3.22 μs ± 4.26% (2.86 … 3.34 μs) 80.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 256, 64) Warps (4, 2) OP (16, 16, 16) 15.5 μs ± 1.21% (15.0 … 16.0 μs) 3.37 μs ± 3.63% (2.86 … 3.81 μs) 19.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 128, 32) Warps (4, 1) OP (16, 16, 16) 5.43 μs ± 2.99% (5.25 … 5.72 μs) 3.38 μs ± 3.36% (3.1 … 3.81 μs) 59.1
WMMA GEMM Float16*Float16+Float32=Float32 (2048×2048) · (2048×2048) (TN) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 3.23 ms ± 0.178% (3.21 … 3.25 ms) 290.0 μs ± 0.2% (289.0 … 295.0 μs) 8.98
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (128, 64, 16) Warps (2, 2) OP (16, 16, 16) 7.53 μs ± 2.07% (7.39 … 7.87 μs) 3.49 μs ± 3.84% (3.1 … 3.81 μs) 41.9
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (256, 64, 16) Warps (2, 4) OP (16, 16, 16) 12.3 μs ± 1.16% (11.9 … 12.6 μs) 3.33 μs ± 3.96% (2.86 … 3.81 μs) 24.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 256, 16) Warps (1, 4) OP (16, 16, 16) 13.8 μs ± 0.898% (13.4 … 13.8 μs) 3.38 μs ± 3.47% (3.1 … 3.81 μs) 23.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 256, 16) Warps (2, 2) OP (16, 16, 16) 12.4 μs ± 1.41% (12.2 … 12.6 μs) 3.47 μs ± 3.78% (3.1 … 3.81 μs) 25.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 64, 16) Warps (4, 4) OP (16, 16, 16) 5.19 μs ± 3.16% (5.01 … 5.48 μs) 3.21 μs ± 4.38% (2.86 … 3.34 μs) 57.1
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (128, 128, 16) Warps (2, 2) OP (16, 16, 16) 11.8 μs ± 1.17% (11.4 … 11.9 μs) 3.47 μs ± 4.07% (3.1 … 4.05 μs) 27.1
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TT) OP (4, 32, 1), base shape (1, 32, 1) 28.8 μs ± 0.58% (28.6 … 29.1 μs) 3.56 μs ± 2.71% (3.1 … 4.05 μs) 10.8
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (128, 64, 16) Warps (4, 1) OP (16, 16, 16) 6.98 μs ± 1.73% (6.68 … 7.39 μs) 3.21 μs ± 4.48% (2.86 … 3.34 μs) 42.9
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 256, 32) Warps (4, 4) OP (16, 16, 16) 11.0 μs ± 1.09% (10.7 … 11.4 μs) 3.46 μs ± 3.91% (3.1 … 3.81 μs) 28.9
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 256, 16) Warps (4, 2) OP (16, 16, 16) 14.0 μs ± 0.938% (13.6 … 14.3 μs) 3.55 μs ± 3.44% (3.1 … 4.05 μs) 22.8
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NT) OP (4, 16, 1), base shape (4, 8, 1) 46.6 μs ± 0.662% (45.8 … 47.9 μs) 3.71 μs ± 3.75% (3.58 … 4.05 μs) 7.81
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (128, 64, 32) Warps (2, 4) OP (16, 16, 16) 6.33 μs ± 2.26% (5.96 … 6.44 μs) 3.3 μs ± 3.24% (2.86 … 3.58 μs) 48.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (256, 64, 16) Warps (4, 1) OP (16, 16, 16) 18.0 μs ± 0.893% (17.9 … 18.4 μs) 3.35 μs ± 3.58% (2.86 … 3.81 μs) 16.0
FPU GEMM Float16*Float16=Float32 (128×128) · (128×128) (NN) OP (8, 16, 2), base shape (4, 8, 1) 17.2 μs ± 0.815% (16.9 … 17.9 μs) 1.98 μs ± 8.31% (1.67 … 2.15 μs) 9.86
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 256, 32) Warps (4, 1) OP (16, 16, 16) 64.7 μs ± 0.4% (63.9 … 65.3 μs) 3.31 μs ± 2.94% (2.86 … 3.58 μs) 4.48
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (128, 64, 32) Warps (2, 4) OP (16, 16, 16) 6.06 μs ± 2.61% (5.72 … 6.2 μs) 3.38 μs ± 3.87% (2.86 … 3.81 μs) 50.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (128, 64, 16) Warps (4, 1) OP (16, 16, 16) 8.0 μs ± 1.9% (7.87 … 8.34 μs) 3.48 μs ± 3.85% (3.1 … 3.81 μs) 39.4
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 64, 32) Warps (2, 4) OP (16, 16, 16) 3.96 μs ± 3.83% (3.81 … 4.29 μs) 3.47 μs ± 3.84% (3.1 … 3.81 μs) 81.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (128, 64, 64) Warps (4, 4) OP (16, 16, 16) 5.77 μs ± 2.92% (5.48 … 5.96 μs) 3.38 μs ± 3.42% (3.1 … 3.81 μs) 56.5
FPU GEMM Float32*Float32=Float32 (256×256) · (256×256) (TN) OP (8, 16, 2), base shape (4, 8, 1) 73.8 μs ± 0.304% (73.2 … 74.6 μs) 6.26 μs ± 2.49% (5.96 … 6.44 μs) 8.14
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (128, 128, 32) Warps (4, 2) OP (16, 16, 16) 9.58 μs ± 1.68% (9.3 … 9.78 μs) 3.27 μs ± 4.05% (2.86 … 3.81 μs) 30.8
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 128, 16) Warps (2, 4) OP (16, 16, 16) 7.22 μs ± 1.68% (6.91 … 7.63 μs) 3.38 μs ± 3.23% (3.1 … 3.81 μs) 44.8
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 128, 16) Warps (1, 4) OP (16, 16, 16) 7.87 μs ± 1.96% (7.63 … 8.11 μs) 3.5 μs ± 3.65% (3.1 … 3.81 μs) 40.6
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (128, 128, 64) Warps (2, 2) OP (16, 16, 16) 16.9 μs ± 0.777% (16.5 … 17.2 μs) 3.47 μs ± 3.94% (3.1 … 3.81 μs) 18.8
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 128, 32) Warps (2, 4) OP (16, 16, 16) 5.84 μs ± 2.58% (5.48 … 5.96 μs) 3.38 μs ± 3.71% (2.86 … 3.81 μs) 52.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 64, 64) Warps (4, 1) OP (16, 16, 16) 3.75 μs ± 3.64% (3.58 … 4.05 μs) 3.46 μs ± 3.87% (3.1 … 3.81 μs) 86.7
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 128, 32) Warps (1, 4) OP (16, 16, 16) 5.93 μs ± 2.45% (5.72 … 6.2 μs) 3.31 μs ± 2.85% (2.86 … 3.58 μs) 50.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 128, 32) Warps (4, 4) OP (16, 16, 16) 5.81 μs ± 2.84% (5.48 … 5.96 μs) 3.39 μs ± 3.94% (2.86 … 3.81 μs) 52.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (256, 64, 16) Warps (4, 2) OP (16, 16, 16) 14.6 μs ± 0.915% (14.5 … 15.0 μs) 3.39 μs ± 3.42% (3.1 … 3.81 μs) 21.3
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 64, 16) Warps (4, 4) OP (16, 16, 16) 5.33 μs ± 2.86% (5.01 … 5.48 μs) 3.49 μs ± 3.75% (3.1 … 3.81 μs) 61.9
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (128, 64, 64) Warps (4, 4) OP (16, 16, 16) 6.33 μs ± 2.3% (5.96 … 6.44 μs) 3.31 μs ± 2.83% (2.86 … 3.58 μs) 48.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 64, 32) Warps (4, 4) OP (16, 16, 16) 5.07 μs ± 3.09% (4.77 … 5.25 μs) 3.3 μs ± 3.25% (2.86 … 3.58 μs) 60.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (256, 64, 16) Warps (2, 4) OP (16, 16, 16) 15.1 μs ± 0.954% (14.8 … 15.5 μs) 3.35 μs ± 3.7% (2.86 … 3.81 μs) 19.4
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (256, 64, 32) Warps (4, 1) OP (16, 16, 16) 24.2 μs ± 0.547% (23.6 … 24.6 μs) 3.47 μs ± 3.96% (3.1 … 3.81 μs) 13.1
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (256, 64, 64) Warps (1, 4) OP (16, 16, 16) 1.71 ms ± 0.142% (1.71 … 1.72 ms) 3.3 μs ± 3.22% (2.86 … 3.58 μs) 0.168
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 64, 16) Warps (4, 4) OP (16, 16, 16) 5.18 μs ± 3.03% (5.01 … 5.48 μs) 3.34 μs ± 2.69% (2.86 … 3.81 μs) 57.1
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 256, 64) Warps (2, 2) OP (16, 16, 16) 30.0 μs ± 0.803% (29.3 … 30.8 μs) 3.39 μs ± 3.91% (2.86 … 3.81 μs) 9.76
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 64, 16) Warps (2, 2) OP (16, 16, 16) 5.65 μs ± 2.3% (5.25 … 5.96 μs) 3.55 μs ± 3.5% (3.1 … 4.05 μs) 59.1
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 128, 64) Warps (1, 4) OP (16, 16, 16) 38.6 μs ± 0.641% (38.1 … 39.6 μs) 3.46 μs ± 3.92% (3.1 … 3.81 μs) 8.12
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 256, 64) Warps (4, 2) OP (16, 16, 16) 14.8 μs ± 1.2% (14.3 … 15.3 μs) 3.39 μs ± 3.82% (2.86 … 3.81 μs) 20.0
WMMA GEMM Float16*Float16+Float16=Float16 (256×256) · (256×256) (NN) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 27.6 μs ± 0.562% (27.2 … 28.1 μs) 3.36 μs ± 3.04% (3.1 … 3.81 μs) 11.4
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 256, 64) Warps (2, 4) OP (16, 16, 16) 49.4 μs ± 0.541% (48.6 … 50.3 μs) 3.38 μs ± 3.68% (2.86 … 3.81 μs) 5.88
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NN) OP (4, 32, 1), base shape (1, 32, 1) 13.8 μs ± 0.71% (13.4 … 14.1 μs) 3.61 μs ± 2.51% (3.58 … 4.05 μs) 26.8
WMMA GEMM Float16*Float16+Float16=Float16 (256×256) · (256×256) (NN) Block (128, 128, 64) Warps (4, 2) OP (8, 32, 16) 26.4 μs ± 0.685% (26.0 … 26.9 μs) 3.36 μs ± 3.12% (2.86 … 3.81 μs) 11.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 128, 64) Warps (4, 4) OP (16, 16, 16) 6.18 μs ± 2.05% (5.72 … 6.44 μs) 3.47 μs ± 4.04% (3.1 … 3.81 μs) 54.2
WMMA GEMM Float16*Float16+Float16=Float16 (2048×2048) · (2048×2048) (TT) Block (128, 128, 64) Warps (4, 2) OP (32, 8, 16) 200.0 μs ± 0.641% (198.0 … 204.0 μs) 178.0 μs ± 0.447% (176.0 … 181.0 μs) 88.9
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (256, 64, 64) Warps (1, 4) OP (16, 16, 16) 1.76 ms ± 0.214% (1.75 … 1.77 ms) 3.38 μs ± 3.34% (2.86 … 3.81 μs) 0.164
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 64, 64) Warps (4, 2) OP (16, 16, 16) 3.52 μs ± 3.36% (3.1 … 3.81 μs) 3.21 μs ± 4.3% (2.86 … 3.34 μs) 92.3
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (256, 64, 16) Warps (4, 4) OP (16, 16, 16) 12.0 μs ± 1.29% (11.7 … 12.2 μs) 3.49 μs ± 3.76% (3.1 … 3.81 μs) 26.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (128, 128, 16) Warps (4, 4) OP (16, 16, 16) 11.2 μs ± 1.4% (11.0 … 11.4 μs) 3.23 μs ± 4.3% (2.86 … 3.34 μs) 26.1
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 64, 16) Warps (2, 4) OP (16, 16, 16) 5.5 μs ± 3.13% (5.25 … 5.72 μs) 3.56 μs ± 3.52% (3.1 … 3.81 μs) 59.1
WMMA GEMM Float16*Float16+Float16=Float16 (512×256) · (256×512) (NN) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 30.7 μs ± 0.668% (30.0 … 31.2 μs) 4.43 μs ± 3.54% (4.29 … 4.77 μs) 14.3
FPU GEMM Float64*Float64=Float64 (128×128) · (128×128) (NN) OP (8, 16, 2), base shape (4, 8, 1) 139.0 μs ± 0.136% (138.0 … 139.0 μs) 28.8 μs ± 0.564% (28.6 … 29.1 μs) 20.7
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 128, 32) Warps (2, 2) OP (16, 16, 16) 5.75 μs ± 2.97% (5.48 … 5.96 μs) 3.39 μs ± 3.98% (2.86 … 3.81 μs) 52.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (128, 64, 64) Warps (4, 2) OP (16, 16, 16) 6.08 μs ± 2.52% (5.72 … 6.2 μs) 3.31 μs ± 3.02% (2.86 … 3.58 μs) 50.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 64, 16) Warps (4, 2) OP (16, 16, 16) 5.16 μs ± 2.96% (5.01 … 5.48 μs) 3.22 μs ± 4.31% (2.86 … 3.34 μs) 57.1
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (128, 64, 16) Warps (4, 2) OP (16, 16, 16) 6.92 μs ± 0.839% (6.68 … 7.39 μs) 3.3 μs ± 3.6% (2.86 … 3.58 μs) 42.9
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (128, 64, 64) Warps (2, 2) OP (16, 16, 16) 37.3 μs ± 0.801% (36.5 … 38.4 μs) 3.41 μs ± 3.86% (2.86 … 3.81 μs) 7.84
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 64, 16) Warps (1, 4) OP (16, 16, 16) 5.63 μs ± 2.21% (5.25 … 5.96 μs) 3.31 μs ± 3.72% (2.86 … 3.81 μs) 54.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 64, 32) Warps (4, 4) OP (16, 16, 16) 4.88 μs ± 2.78% (4.53 … 5.25 μs) 3.22 μs ± 4.43% (2.86 … 3.34 μs) 63.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (128, 64, 16) Warps (1, 4) OP (16, 16, 16) 9.36 μs ± 1.79% (9.06 … 9.54 μs) 3.21 μs ± 4.36% (2.86 … 3.34 μs) 31.6
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (128, 64, 64) Warps (4, 1) OP (16, 16, 16) 32.9 μs ± 0.769% (32.2 … 33.9 μs) 3.38 μs ± 3.34% (3.1 … 3.81 μs) 9.63
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (256, 64, 32) Warps (4, 1) OP (16, 16, 16) 21.8 μs ± 0.752% (21.5 … 22.2 μs) 3.38 μs ± 4.05% (2.86 … 3.81 μs) 13.3
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 128, 16) Warps (4, 2) OP (16, 16, 16) 7.31 μs ± 1.99% (7.15 … 7.63 μs) 3.47 μs ± 4.01% (3.1 … 3.81 μs) 43.3
WMMA GEMM Float16*Float16+Float16=Float16 (512×512) · (512×512) (NN) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 56.5 μs ± 0.5% (55.8 … 57.5 μs) 6.36 μs ± 2.16% (5.96 … 6.68 μs) 10.7
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (256, 64, 32) Warps (2, 4) OP (16, 16, 16) 10.4 μs ± 1.28% (10.0 … 10.7 μs) 3.28 μs ± 3.96% (2.86 … 3.81 μs) 28.6
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (128, 128, 64) Warps (4, 2) OP (8, 32, 16) 30.3 μs ± 0.587% (29.8 … 31.0 μs) 3.43 μs ± 3.9% (3.1 … 3.81 μs) 10.4
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 128, 32) Warps (4, 2) OP (16, 16, 16) 5.46 μs ± 3.06% (5.25 … 5.72 μs) 3.38 μs ± 3.37% (3.1 … 3.81 μs) 59.1
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (256, 64, 64) Warps (4, 4) OP (16, 16, 16) 11.3 μs ± 1.34% (11.2 … 11.7 μs) 3.38 μs ± 3.91% (2.86 … 3.81 μs) 25.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (256, 64, 16) Warps (1, 4) OP (16, 16, 16) 16.6 μs ± 0.87% (16.2 … 16.9 μs) 3.21 μs ± 4.36% (2.86 … 3.34 μs) 17.6
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (256, 64, 16) Warps (4, 2) OP (16, 16, 16) 14.4 μs ± 1.06% (14.3 … 14.8 μs) 3.21 μs ± 4.51% (2.86 … 3.34 μs) 20.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 256, 32) Warps (2, 4) OP (16, 16, 16) 10.6 μs ± 1.3% (10.5 … 11.0 μs) 3.47 μs ± 3.88% (3.1 … 3.81 μs) 29.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (128, 128, 16) Warps (2, 2) OP (16, 16, 16) 13.5 μs ± 1.04% (13.1 … 13.6 μs) 3.38 μs ± 3.47% (3.1 … 3.81 μs) 23.6
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (128, 64, 16) Warps (2, 2) OP (16, 16, 16) 7.45 μs ± 1.53% (7.15 … 7.87 μs) 3.56 μs ± 3.36% (3.1 … 4.05 μs) 43.3
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 128, 16) Warps (2, 2) OP (16, 16, 16) 7.49 μs ± 1.87% (7.39 … 7.87 μs) 3.38 μs ± 3.43% (3.1 … 3.81 μs) 41.9
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (128, 128, 16) Warps (4, 2) OP (16, 16, 16) 11.6 μs ± 1.49% (11.4 … 11.9 μs) 3.21 μs ± 4.54% (2.86 … 3.34 μs) 25.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 256, 32) Warps (2, 4) OP (16, 16, 16) 9.81 μs ± 1.53% (9.54 … 10.0 μs) 3.4 μs ± 4.01% (2.86 … 3.81 μs) 30.0
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NN) OP (4, 32, 1), base shape (2, 16, 1) 60.9 μs ± 0.395% (60.3 … 61.5 μs) 3.6 μs ± 2.13% (3.58 … 4.05 μs) 5.93
WMMA GEMM Float16*Float16+Float16=Float16 (512×512) · (512×512) (TN) Block (128, 128, 64) Warps (4, 2) OP (8, 32, 16) 54.9 μs ± 0.497% (54.1 … 55.6 μs) 6.57 μs ± 2.12% (6.2 … 6.91 μs) 11.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 128, 32) Warps (4, 2) OP (16, 16, 16) 5.43 μs ± 2.92% (5.25 … 5.72 μs) 3.39 μs ± 3.86% (2.86 … 3.81 μs) 54.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 256, 16) Warps (1, 4) OP (16, 16, 16) 13.0 μs ± 1.01% (12.6 … 13.1 μs) 3.55 μs ± 3.48% (3.1 … 4.05 μs) 24.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 128, 64) Warps (2, 2) OP (16, 16, 16) 33.1 μs ± 0.871% (32.2 … 33.9 μs) 3.39 μs ± 3.59% (2.86 … 3.81 μs) 8.89
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (256, 64, 64) Warps (4, 1) OP (16, 16, 16) 29.1 μs ± 0.579% (28.8 … 29.6 μs) 3.48 μs ± 3.99% (3.1 … 4.05 μs) 10.7
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 64, 16) Warps (4, 4) OP (16, 16, 16) 5.06 μs ± 3.2% (4.77 … 5.25 μs) 3.31 μs ± 3.83% (2.86 … 3.58 μs) 60.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 64, 16) Warps (4, 4) OP (16, 16, 16) 5.24 μs ± 3.33% (5.01 … 5.48 μs) 3.55 μs ± 3.57% (3.1 … 4.05 μs) 61.9
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (256, 64, 16) Warps (2, 2) OP (16, 16, 16) 17.2 μs ± 0.741% (16.9 … 17.6 μs) 3.56 μs ± 3.51% (3.1 … 4.05 μs) 18.3
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (128, 64, 32) Warps (4, 2) OP (16, 16, 16) 5.88 μs ± 2.08% (5.48 … 6.2 μs) 3.48 μs ± 3.9% (3.1 … 3.81 μs) 56.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 128, 32) Warps (2, 2) OP (16, 16, 16) 5.9 μs ± 2.01% (5.48 … 6.2 μs) 3.3 μs ± 3.16% (2.86 … 3.58 μs) 52.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (256, 64, 64) Warps (4, 1) OP (16, 16, 16) 24.6 μs ± 0.532% (24.3 … 25.0 μs) 3.38 μs ± 3.47% (2.86 … 3.81 μs) 11.8
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 64, 64) Warps (4, 4) OP (16, 16, 16) 3.99 μs ± 3.68% (3.81 … 4.29 μs) 3.31 μs ± 2.83% (2.86 … 3.58 μs) 75.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 128, 32) Warps (4, 1) OP (16, 16, 16) 5.88 μs ± 2.07% (5.48 … 6.2 μs) 3.47 μs ± 3.94% (3.1 … 3.81 μs) 56.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (256, 64, 64) Warps (2, 4) OP (16, 16, 16) 13.4 μs ± 0.828% (12.9 … 13.6 μs) 3.31 μs ± 3.5% (2.86 … 3.58 μs) 22.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 256, 16) Warps (2, 2) OP (16, 16, 16) 12.3 μs ± 1.04% (11.9 … 12.6 μs) 3.31 μs ± 3.67% (2.86 … 3.81 μs) 24.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 256, 64) Warps (2, 2) OP (16, 16, 16) 22.2 μs ± 0.869% (21.7 … 22.6 μs) 3.31 μs ± 2.88% (2.86 … 3.81 μs) 13.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×128) · (128×256) (NT) Block (128, 128, 64) Warps (4, 2) OP (32, 8, 16) 7.71 μs ± 1.69% (7.39 … 8.11 μs) 2.57 μs ± 5.26% (2.15 … 2.86 μs) 29.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 64, 16) Warps (4, 4) OP (16, 16, 16) 5.48 μs ± 3.1% (5.25 … 5.72 μs) 3.37 μs ± 3.37% (3.1 … 3.81 μs) 59.1
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (256, 64, 16) Warps (2, 2) OP (16, 16, 16) 13.4 μs ± 1.08% (13.1 … 13.6 μs) 3.48 μs ± 4.15% (3.1 … 4.05 μs) 23.6
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (128, 128, 64) Warps (2, 4) OP (16, 16, 16) 28.8 μs ± 0.65% (28.4 … 29.3 μs) 3.39 μs ± 3.95% (2.86 … 4.05 μs) 10.1
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 64, 16) Warps (4, 2) OP (16, 16, 16) 5.23 μs ± 3.31% (5.01 … 5.48 μs) 3.49 μs ± 3.91% (3.1 … 3.81 μs) 61.9
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NN) OP (16, 16, 1), base shape (4, 8, 1) 40.5 μs ± 0.379% (40.1 … 40.8 μs) 3.61 μs ± 2.62% (3.58 … 4.05 μs) 8.93
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (256, 64, 32) Warps (4, 2) OP (16, 16, 16) 12.0 μs ± 1.38% (11.7 … 12.2 μs) 3.48 μs ± 4.12% (3.1 … 3.81 μs) 26.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (128, 64, 32) Warps (4, 1) OP (16, 16, 16) 5.46 μs ± 3.03% (5.25 … 5.72 μs) 3.4 μs ± 3.81% (2.86 … 3.81 μs) 54.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (128, 64, 16) Warps (1, 4) OP (16, 16, 16) 8.04 μs ± 2.12% (7.87 … 8.34 μs) 3.48 μs ± 3.89% (3.1 … 3.81 μs) 39.4
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (128, 128, 32) Warps (4, 4) OP (16, 16, 16) 9.64 μs ± 1.62% (9.3 … 9.78 μs) 3.21 μs ± 4.51% (2.86 … 3.34 μs) 30.8
WMMA GEMM Float16*Float16+Float16=Float16 (2048×2048) · (2048×2048) (NT) Block (128, 128, 64) Warps (4, 2) OP (32, 8, 16) 200.0 μs ± 0.658% (198.0 … 204.0 μs) 177.0 μs ± 1.55% (173.0 … 183.0 μs) 87.6
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 256, 32) Warps (2, 4) OP (16, 16, 16) 9.66 μs ± 1.56% (9.3 … 9.78 μs) 3.22 μs ± 4.35% (2.86 … 3.34 μs) 30.8
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 128, 64) Warps (1, 4) OP (16, 16, 16) 36.1 μs ± 0.645% (35.5 … 37.0 μs) 3.31 μs ± 2.78% (2.86 … 3.58 μs) 8.05
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (256, 64, 16) Warps (1, 4) OP (16, 16, 16) 16.3 μs ± 0.988% (16.0 … 16.5 μs) 3.31 μs ± 3.87% (2.86 … 3.81 μs) 17.9
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 256, 32) Warps (2, 4) OP (16, 16, 16) 10.7 μs ± 1.24% (10.5 … 11.0 μs) 3.46 μs ± 3.8% (3.1 … 3.81 μs) 29.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (128, 128, 32) Warps (2, 2) OP (16, 16, 16) 9.79 μs ± 1.46% (9.54 … 10.0 μs) 3.38 μs ± 3.41% (3.1 … 3.81 μs) 32.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 128, 16) Warps (4, 1) OP (16, 16, 16) 7.58 μs ± 2.07% (7.39 … 7.87 μs) 3.48 μs ± 3.82% (3.1 … 3.81 μs) 41.9
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 64, 32) Warps (2, 4) OP (16, 16, 16) 3.89 μs ± 3.17% (3.81 … 4.29 μs) 3.48 μs ± 4.07% (3.1 … 4.05 μs) 81.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (128, 128, 16) Warps (4, 4) OP (16, 16, 16) 12.0 μs ± 1.13% (11.7 … 12.4 μs) 3.49 μs ± 3.87% (3.1 … 3.81 μs) 26.5
WMMA GEMM Float16*Float16+Float16=Float16 (256×512) · (512×256) (NN) Block (128, 128, 64) Warps (4, 2) OP (8, 32, 16) 49.5 μs ± 0.458% (48.9 … 50.3 μs) 4.82 μs ± 3.83% (4.53 … 5.25 μs) 9.27
WMMA GEMM Float16*Float16+Float32=Float32 (2048×2048) · (2048×2048) (NT) Block (128, 128, 64) Warps (4, 2) OP (8, 32, 16) 3.12 ms ± 0.297% (3.1 … 3.15 ms) 289.0 μs ± 0.175% (288.0 … 292.0 μs) 9.29
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (128, 64, 32) Warps (4, 2) OP (16, 16, 16) 5.46 μs ± 3.09% (5.25 … 5.72 μs) 3.39 μs ± 3.93% (2.86 … 3.81 μs) 54.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 256, 32) Warps (4, 2) OP (16, 16, 16) 10.6 μs ± 1.15% (10.5 … 11.0 μs) 3.47 μs ± 3.96% (3.1 … 3.81 μs) 29.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 64, 64) Warps (4, 2) OP (16, 16, 16) 3.81 μs ± 2.85% (3.58 … 4.29 μs) 3.31 μs ± 2.74% (2.86 … 3.58 μs) 80.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (256, 64, 32) Warps (2, 2) OP (16, 16, 16) 25.7 μs ± 0.664% (25.5 … 26.0 μs) 3.39 μs ± 3.96% (2.86 … 3.81 μs) 11.2
WMMA GEMM Float16*Float16+Float16=Float16 (2048×2048) · (2048×2048) (TN) Block (128, 128, 64) Warps (4, 2) OP (32, 8, 16) 205.0 μs ± 0.515% (202.0 … 208.0 μs) 179.0 μs ± 0.337% (178.0 … 181.0 μs) 88.2
WMMA GEMM Float16*Float16+Float16=Float16 (512×512) · (512×512) (NN) Block (128, 128, 64) Warps (4, 2) OP (8, 32, 16) 54.7 μs ± 0.484% (53.9 … 55.6 μs) 6.41 μs ± 1.81% (5.96 … 6.68 μs) 11.1
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 128, 32) Warps (4, 2) OP (16, 16, 16) 5.9 μs ± 2.04% (5.72 … 6.2 μs) 3.3 μs ± 3.16% (2.86 … 3.81 μs) 50.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (256, 64, 64) Warps (4, 1) OP (16, 16, 16) 24.3 μs ± 0.447% (23.8 … 24.8 μs) 3.37 μs ± 3.97% (2.86 … 3.81 μs) 12.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 256, 64) Warps (1, 4) OP (16, 16, 16) 24.7 μs ± 0.606% (24.3 … 25.3 μs) 3.47 μs ± 4.29% (3.1 … 4.05 μs) 12.7
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 256, 64) Warps (2, 2) OP (16, 16, 16) 30.4 μs ± 0.771% (29.8 … 31.2 μs) 3.47 μs ± 3.92% (3.1 … 3.81 μs) 10.4
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (256, 64, 16) Warps (4, 4) OP (16, 16, 16) 12.0 μs ± 1.41% (11.7 … 12.4 μs) 3.56 μs ± 3.52% (3.1 … 4.05 μs) 26.5
WMMA GEMM Float16*Float16+Float16=Float16 (512×512) · (512×512) (NN) Block (128, 128, 64) Warps (4, 2) OP (32, 8, 16) 11.4 μs ± 1.46% (11.2 … 11.9 μs) 6.41 μs ± 1.87% (5.96 … 6.68 μs) 53.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (128, 128, 16) Warps (2, 4) OP (16, 16, 16) 12.5 μs ± 1.24% (12.2 … 12.6 μs) 3.38 μs ± 3.26% (3.1 … 3.81 μs) 25.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (128, 128, 32) Warps (1, 4) OP (16, 16, 16) 10.9 μs ± 1.35% (10.7 … 11.2 μs) 3.37 μs ± 3.2% (3.1 … 3.81 μs) 28.9
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (128, 64, 64) Warps (4, 1) OP (16, 16, 16) 33.4 μs ± 0.768% (32.7 … 34.1 μs) 3.46 μs ± 3.91% (3.1 … 3.81 μs) 9.49
WMMA GEMM Float16*Float16+Float32=Float32 (256×128) · (128×256) (NN) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 16.8 μs ± 0.986% (16.5 … 17.2 μs) 2.54 μs ± 5.21% (2.15 … 2.86 μs) 13.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (256, 64, 32) Warps (2, 4) OP (16, 16, 16) 12.1 μs ± 1.4% (11.9 … 12.4 μs) 3.47 μs ± 3.76% (3.1 … 3.81 μs) 26.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 128, 64) Warps (4, 4) OP (16, 16, 16) 6.18 μs ± 2.22% (5.96 … 6.44 μs) 3.48 μs ± 3.92% (3.1 … 3.81 μs) 52.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 64, 16) Warps (2, 4) OP (16, 16, 16) 5.39 μs ± 2.38% (5.01 … 5.72 μs) 3.34 μs ± 2.63% (2.86 … 3.81 μs) 57.1
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 256, 16) Warps (2, 4) OP (16, 16, 16) 12.0 μs ± 1.21% (11.7 … 12.2 μs) 3.3 μs ± 3.82% (2.86 … 3.81 μs) 24.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (128, 64, 16) Warps (4, 2) OP (16, 16, 16) 7.85 μs ± 1.98% (7.63 … 8.11 μs) 3.38 μs ± 3.4% (2.86 … 3.81 μs) 37.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 128, 32) Warps (2, 4) OP (16, 16, 16) 5.9 μs ± 2.21% (5.48 … 6.2 μs) 3.47 μs ± 3.88% (3.1 … 3.81 μs) 56.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 64, 32) Warps (1, 4) OP (16, 16, 16) 3.68 μs ± 3.74% (3.58 … 4.05 μs) 3.3 μs ± 3.75% (2.86 … 3.81 μs) 80.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (256, 64, 16) Warps (4, 2) OP (16, 16, 16) 13.2 μs ± 1.09% (12.9 … 13.4 μs) 3.34 μs ± 3.06% (2.86 … 3.81 μs) 22.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (128, 128, 16) Warps (4, 2) OP (16, 16, 16) 12.1 μs ± 1.44% (11.9 … 12.4 μs) 3.38 μs ± 3.25% (3.1 … 3.81 μs) 26.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 128, 64) Warps (4, 1) OP (16, 16, 16) 8.4 μs ± 1.88% (8.11 … 8.82 μs) 3.46 μs ± 4.01% (3.1 … 3.81 μs) 38.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (256, 64, 32) Warps (1, 4) OP (16, 16, 16) 72.9 μs ± 0.404% (72.0 … 73.9 μs) 3.38 μs ± 3.31% (3.1 … 3.81 μs) 4.3
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 256, 64) Warps (2, 4) OP (16, 16, 16) 44.6 μs ± 0.569% (43.9 … 45.5 μs) 3.47 μs ± 3.91% (3.1 … 3.81 μs) 7.07
WMMA GEMM Float16*Float16+Float32=Float32 (128×128) · (128×128) (NT) Block (128, 128, 64) Warps (4, 2) OP (32, 8, 16) 7.54 μs ± 2.06% (7.39 … 7.87 μs) 1.98 μs ± 8.14% (1.67 … 2.38 μs) 22.6
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 64, 64) Warps (1, 4) OP (16, 16, 16) 3.62 μs ± 2.91% (3.58 … 4.05 μs) 3.39 μs ± 3.87% (2.86 … 3.81 μs) 80.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 256, 64) Warps (4, 2) OP (16, 16, 16) 15.6 μs ± 1.25% (15.0 … 16.2 μs) 3.47 μs ± 4.1% (3.1 … 3.81 μs) 20.6
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 256, 64) Warps (4, 4) OP (16, 16, 16) 10.5 μs ± 0.994% (10.0 … 10.7 μs) 3.38 μs ± 4.0% (2.86 … 3.81 μs) 28.6
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 128, 16) Warps (2, 4) OP (16, 16, 16) 7.06 μs ± 1.89% (6.68 … 7.39 μs) 3.48 μs ± 3.86% (3.1 … 4.05 μs) 46.4
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (128, 128, 32) Warps (4, 1) OP (16, 16, 16) 12.2 μs ± 1.44% (11.9 … 12.6 μs) 3.37 μs ± 3.22% (3.1 … 3.81 μs) 26.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (128, 64, 64) Warps (1, 4) OP (16, 16, 16) 8.29 μs ± 2.01% (8.11 … 8.58 μs) 3.29 μs ± 4.02% (2.86 … 3.81 μs) 35.3
WMMA GEMM Float16*Float16+Float16=Float16 (256×512) · (512×256) (NT) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 51.2 μs ± 0.382% (50.8 … 51.7 μs) 3.92 μs ± 3.6% (3.81 … 4.29 μs) 7.51
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (128, 128, 32) Warps (2, 4) OP (16, 16, 16) 9.63 μs ± 1.67% (9.3 … 9.78 μs) 3.38 μs ± 3.22% (3.1 … 3.81 μs) 33.3
FPU GEMM Float32*Float32=Float32 (256×256) · (256×256) (NT) OP (8, 16, 2), base shape (4, 8, 1) 73.5 μs ± 0.286% (73.0 … 74.1 μs) 5.88 μs ± 2.42% (5.48 … 6.2 μs) 7.52
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (128, 128, 32) Warps (4, 1) OP (16, 16, 16) 11.7 μs ± 1.37% (11.4 … 12.2 μs) 3.31 μs ± 2.8% (2.86 … 3.58 μs) 25.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (128, 64, 64) Warps (2, 4) OP (16, 16, 16) 5.71 μs ± 2.85% (5.48 … 5.96 μs) 3.39 μs ± 3.88% (2.86 … 3.81 μs) 52.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 128, 32) Warps (2, 4) OP (16, 16, 16) 5.53 μs ± 3.03% (5.25 … 5.72 μs) 3.38 μs ± 3.28% (3.1 … 3.81 μs) 59.1
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (128, 64, 16) Warps (4, 4) OP (16, 16, 16) 7.88 μs ± 1.97% (7.63 … 8.34 μs) 3.5 μs ± 3.65% (3.1 … 3.81 μs) 40.6
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 128, 16) Warps (4, 4) OP (16, 16, 16) 7.18 μs ± 1.25% (6.91 … 7.63 μs) 3.49 μs ± 3.7% (3.1 … 3.81 μs) 44.8
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (128, 128, 64) Warps (1, 4) OP (16, 16, 16) 18.7 μs ± 0.945% (18.4 … 19.1 μs) 3.38 μs ± 3.35% (3.1 … 3.81 μs) 16.9
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (256, 64, 32) Warps (4, 4) OP (16, 16, 16) 10.1 μs ± 1.37% (9.78 … 10.3 μs) 3.4 μs ± 3.97% (2.86 … 3.81 μs) 29.3
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (128, 64, 32) Warps (4, 1) OP (16, 16, 16) 5.89 μs ± 1.98% (5.72 … 6.2 μs) 3.31 μs ± 2.94% (2.86 … 3.81 μs) 50.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 128, 64) Warps (4, 2) OP (16, 16, 16) 5.65 μs ± 2.45% (5.48 … 5.96 μs) 3.39 μs ± 3.9% (2.86 … 3.81 μs) 52.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (128, 64, 64) Warps (4, 1) OP (16, 16, 16) 31.2 μs ± 0.646% (30.5 … 31.9 μs) 3.23 μs ± 4.2% (2.86 … 3.34 μs) 9.38
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (128, 128, 16) Warps (1, 4) OP (16, 16, 16) 13.1 μs ± 0.88% (12.6 … 13.4 μs) 3.49 μs ± 3.72% (3.1 … 3.81 μs) 24.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (128, 128, 16) Warps (4, 1) OP (16, 16, 16) 12.4 μs ± 1.39% (12.2 … 12.6 μs) 3.32 μs ± 3.8% (2.86 … 3.81 μs) 23.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (128, 128, 16) Warps (4, 1) OP (16, 16, 16) 14.2 μs ± 0.995% (14.1 … 14.5 μs) 3.21 μs ± 4.38% (2.86 … 3.34 μs) 20.3
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 128, 16) Warps (2, 2) OP (16, 16, 16) 8.2 μs ± 1.68% (7.87 … 8.58 μs) 3.49 μs ± 3.64% (3.1 … 3.81 μs) 39.4
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 256, 32) Warps (2, 2) OP (16, 16, 16) 19.9 μs ± 0.831% (19.6 … 20.0 μs) 3.47 μs ± 3.9% (3.1 … 3.81 μs) 15.9
FPU GEMM Float64*Float64=Float64 (256×256) · (256×256) (TN) OP (8, 16, 2), base shape (4, 8, 1) 278.0 μs ± 0.0762% (277.0 … 278.0 μs) 109.0 μs ± 0.163% (109.0 … 110.0 μs) 39.4
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (128, 64, 32) Warps (1, 4) OP (16, 16, 16) 5.92 μs ± 2.44% (5.72 … 6.2 μs) 3.47 μs ± 3.92% (3.1 … 3.81 μs) 54.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 256, 32) Warps (2, 2) OP (16, 16, 16) 18.5 μs ± 0.872% (18.4 … 18.8 μs) 3.22 μs ± 4.25% (2.86 … 3.34 μs) 15.6
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (128, 128, 32) Warps (2, 2) OP (16, 16, 16) 9.79 μs ± 1.41% (9.54 … 10.0 μs) 3.38 μs ± 4.01% (2.86 … 3.81 μs) 30.0
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NN) OP (8, 16, 2), base shape (4, 8, 1) 45.2 μs ± 0.39% (44.8 … 46.0 μs) 3.61 μs ± 2.35% (3.58 … 4.05 μs) 7.98
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (256, 64, 64) Warps (2, 2) OP (16, 16, 16) 34.6 μs ± 0.572% (34.1 … 35.3 μs) 3.46 μs ± 3.99% (3.1 … 3.81 μs) 9.09
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 256, 32) Warps (4, 4) OP (16, 16, 16) 11.0 μs ± 1.03% (10.7 … 11.4 μs) 3.47 μs ± 3.95% (3.1 … 3.81 μs) 28.9
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (128, 128, 64) Warps (2, 2) OP (16, 16, 16) 17.1 μs ± 0.669% (16.7 … 17.4 μs) 3.31 μs ± 2.83% (2.86 … 3.81 μs) 17.1
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (256, 64, 32) Warps (2, 2) OP (16, 16, 16) 26.6 μs ± 0.558% (26.2 … 26.9 μs) 3.46 μs ± 4.15% (3.1 … 4.05 μs) 11.8
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (128, 64, 64) Warps (2, 2) OP (16, 16, 16) 36.3 μs ± 0.754% (35.5 … 37.0 μs) 3.3 μs ± 3.04% (2.86 … 3.81 μs) 8.05
FPU GEMM Float32*Float32=Float32 (2048×2048) · (2048×2048) (TN) OP (8, 16, 2), base shape (4, 8, 1) 30.0 ms ± 0.0843% (29.9 … 30.1 ms) 973.0 μs ± 0.615% (942.0 … 985.0 μs) 3.15
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (128, 128, 64) Warps (4, 1) OP (16, 16, 16) 19.9 μs ± 0.934% (19.3 … 20.3 μs) 3.38 μs ± 3.91% (2.86 … 3.81 μs) 14.8
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (256, 64, 32) Warps (4, 2) OP (16, 16, 16) 10.0 μs ± 1.11% (9.78 … 10.3 μs) 3.28 μs ± 3.88% (2.86 … 3.58 μs) 29.3
FPU GEMM Float64*Float64=Float64 (2048×2048) · (2048×2048) (TT) OP (8, 16, 2), base shape (4, 8, 1) 59.9 ms ± 0.168% (59.8 … 60.3 ms) 41.0 ms ± 0.0101% (41.0 … 41.0 ms) 68.6
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (128, 128, 32) Warps (2, 4) OP (16, 16, 16) 10.5 μs ± 0.788% (10.0 … 11.0 μs) 3.31 μs ± 3.04% (2.86 … 3.58 μs) 28.6
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 128, 32) Warps (2, 4) OP (16, 16, 16) 5.92 μs ± 2.38% (5.72 … 6.2 μs) 3.46 μs ± 3.98% (3.1 … 3.81 μs) 54.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (128, 128, 32) Warps (4, 2) OP (16, 16, 16) 10.4 μs ± 1.2% (10.0 … 10.7 μs) 3.47 μs ± 3.93% (3.1 … 3.81 μs) 31.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 64, 64) Warps (2, 4) OP (16, 16, 16) 3.53 μs ± 3.1% (3.1 … 3.81 μs) 3.22 μs ± 4.34% (2.86 … 3.34 μs) 92.3
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 256, 64) Warps (4, 2) OP (16, 16, 16) 14.7 μs ± 1.24% (14.3 … 15.3 μs) 3.3 μs ± 3.81% (2.86 … 3.81 μs) 20.0
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NT) OP (4, 8, 2), base shape (4, 8, 1) 41.7 μs ± 0.411% (41.2 … 42.4 μs) 3.72 μs ± 3.92% (3.58 … 4.05 μs) 8.67
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 256, 16) Warps (4, 2) OP (16, 16, 16) 14.2 μs ± 1.0% (14.1 … 14.5 μs) 3.5 μs ± 3.71% (3.1 … 3.81 μs) 22.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (256, 64, 64) Warps (1, 4) OP (16, 16, 16) 1.76 ms ± 0.153% (1.75 … 1.77 ms) 3.39 μs ± 3.8% (2.86 … 3.81 μs) 0.163
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (128, 128, 64) Warps (4, 4) OP (16, 16, 16) 9.96 μs ± 1.12% (9.54 … 10.3 μs) 3.39 μs ± 3.87% (2.86 … 3.81 μs) 30.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 256, 64) Warps (2, 4) OP (16, 16, 16) 42.9 μs ± 0.583% (42.2 … 43.6 μs) 3.31 μs ± 2.75% (2.86 … 3.81 μs) 6.78
WMMA GEMM Float16*Float16+Float16=Float16 (512×256) · (256×512) (NT) Block (128, 128, 64) Warps (4, 2) OP (8, 32, 16) 29.3 μs ± 0.756% (28.8 … 29.8 μs) 4.34 μs ± 3.02% (4.05 … 4.77 μs) 14.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (256, 64, 32) Warps (4, 2) OP (16, 16, 16) 12.0 μs ± 1.38% (11.7 … 12.2 μs) 3.47 μs ± 3.97% (3.1 … 3.81 μs) 26.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (128, 128, 16) Warps (2, 2) OP (16, 16, 16) 12.3 μs ± 1.11% (11.9 … 12.6 μs) 3.34 μs ± 2.62% (2.86 … 3.81 μs) 24.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (256, 64, 16) Warps (4, 1) OP (16, 16, 16) 13.2 μs ± 1.1% (12.9 … 13.4 μs) 3.34 μs ± 2.74% (2.86 … 3.81 μs) 22.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (256, 64, 32) Warps (1, 4) OP (16, 16, 16) 71.1 μs ± 0.404% (70.3 … 72.0 μs) 3.3 μs ± 3.68% (2.86 … 3.81 μs) 4.07
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 64, 32) Warps (2, 2) OP (16, 16, 16) 3.79 μs ± 3.17% (3.58 … 4.05 μs) 3.22 μs ± 4.25% (2.86 … 3.34 μs) 80.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (128, 128, 64) Warps (2, 2) OP (16, 16, 16) 16.4 μs ± 0.881% (16.0 … 16.7 μs) 3.39 μs ± 3.95% (2.86 … 3.81 μs) 17.9
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 64, 16) Warps (1, 4) OP (16, 16, 16) 6.24 μs ± 2.49% (5.96 … 6.44 μs) 3.48 μs ± 3.83% (3.1 … 3.81 μs) 52.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 256, 64) Warps (4, 4) OP (16, 16, 16) 11.1 μs ± 1.27% (11.0 … 11.4 μs) 3.31 μs ± 2.8% (2.86 … 3.81 μs) 26.1
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (128, 128, 16) Warps (2, 4) OP (16, 16, 16) 13.1 μs ± 1.08% (12.6 … 13.4 μs) 3.22 μs ± 4.43% (2.86 … 3.34 μs) 22.6
WMMA GEMM Float16*Float16+Float16=Float16 (512×256) · (256×512) (NT) Block (128, 128, 64) Warps (4, 2) OP (32, 8, 16) 7.94 μs ± 2.19% (7.63 … 8.34 μs) 4.35 μs ± 2.92% (4.05 … 4.77 μs) 53.1
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (128, 64, 64) Warps (4, 1) OP (16, 16, 16) 32.9 μs ± 0.841% (32.2 … 33.6 μs) 3.39 μs ± 4.04% (2.86 … 3.81 μs) 8.89
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 256, 32) Warps (4, 2) OP (16, 16, 16) 10.5 μs ± 0.584% (10.3 … 11.0 μs) 3.38 μs ± 4.06% (2.86 … 3.81 μs) 27.9
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (128, 128, 32) Warps (4, 1) OP (16, 16, 16) 11.1 μs ± 1.55% (10.7 … 11.4 μs) 3.21 μs ± 4.39% (2.86 … 3.34 μs) 26.7
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 64, 32) Warps (4, 1) OP (16, 16, 16) 3.62 μs ± 2.82% (3.34 … 4.05 μs) 3.39 μs ± 3.77% (2.86 … 3.81 μs) 85.7
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (128, 128, 32) Warps (4, 1) OP (16, 16, 16) 15.1 μs ± 1.14% (14.8 … 15.5 μs) 3.47 μs ± 3.82% (3.1 … 3.81 μs) 21.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 256, 32) Warps (4, 2) OP (16, 16, 16) 9.79 μs ± 1.45% (9.54 … 10.0 μs) 3.22 μs ± 4.41% (2.86 … 3.34 μs) 30.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 64, 64) Warps (4, 1) OP (16, 16, 16) 3.75 μs ± 3.76% (3.58 … 4.05 μs) 3.37 μs ± 4.08% (2.86 … 3.81 μs) 80.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 128, 64) Warps (4, 4) OP (16, 16, 16) 5.75 μs ± 3.0% (5.48 … 5.96 μs) 3.29 μs ± 4.07% (2.86 … 3.81 μs) 52.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (128, 128, 16) Warps (2, 4) OP (16, 16, 16) 11.6 μs ± 1.43% (11.4 … 11.9 μs) 3.32 μs ± 3.99% (2.86 … 3.81 μs) 25.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (128, 128, 64) Warps (4, 1) OP (16, 16, 16) 19.0 μs ± 0.996% (18.6 … 19.6 μs) 3.38 μs ± 3.39% (3.1 … 3.81 μs) 16.7
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 64, 64) Warps (4, 2) OP (16, 16, 16) 3.74 μs ± 3.79% (3.58 … 4.05 μs) 3.47 μs ± 3.84% (3.1 … 3.81 μs) 86.7
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 128, 64) Warps (4, 4) OP (16, 16, 16) 6.04 μs ± 2.69% (5.72 … 6.2 μs) 3.3 μs ± 3.19% (2.86 … 3.58 μs) 50.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 128, 16) Warps (2, 4) OP (16, 16, 16) 6.72 μs ± 1.87% (6.44 … 6.91 μs) 3.22 μs ± 4.41% (2.86 … 3.34 μs) 44.4
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (256, 64, 32) Warps (4, 4) OP (16, 16, 16) 10.0 μs ± 1.28% (9.54 … 10.3 μs) 3.38 μs ± 3.3% (3.1 … 3.81 μs) 32.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 256, 16) Warps (4, 1) OP (16, 16, 16) 16.3 μs ± 1.05% (16.0 … 16.5 μs) 3.49 μs ± 3.76% (3.1 … 3.81 μs) 19.4
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (256, 64, 64) Warps (4, 2) OP (16, 16, 16) 49.4 μs ± 0.609% (48.6 … 50.3 μs) 3.31 μs ± 3.04% (2.86 … 3.81 μs) 5.88
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 64, 16) Warps (2, 2) OP (16, 16, 16) 5.57 μs ± 2.9% (5.25 … 5.72 μs) 3.47 μs ± 3.91% (3.1 … 4.05 μs) 59.1
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (128, 64, 16) Warps (1, 4) OP (16, 16, 16) 7.82 μs ± 2.11% (7.63 … 8.11 μs) 3.31 μs ± 3.79% (2.86 … 3.81 μs) 37.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 64, 16) Warps (2, 2) OP (16, 16, 16) 4.92 μs ± 3.25% (4.77 … 5.25 μs) 3.34 μs ± 2.74% (2.86 … 3.81 μs) 60.0
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NT) OP (32, 4, 1), base shape (32, 1, 1) 106.0 μs ± 0.399% (105.0 … 108.0 μs) 3.71 μs ± 3.66% (3.58 … 4.05 μs) 3.4
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 128, 64) Warps (4, 1) OP (16, 16, 16) 8.09 μs ± 2.06% (7.87 … 8.34 μs) 3.29 μs ± 3.87% (2.86 … 3.81 μs) 36.4
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 128, 64) Warps (1, 4) OP (16, 16, 16) 37.4 μs ± 0.731% (36.5 … 38.4 μs) 3.4 μs ± 3.87% (2.86 … 3.81 μs) 7.84
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (128, 64, 32) Warps (4, 4) OP (16, 16, 16) 6.46 μs ± 2.05% (6.2 … 6.68 μs) 3.23 μs ± 4.19% (2.86 … 3.34 μs) 46.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 128, 16) Warps (1, 4) OP (16, 16, 16) 7.84 μs ± 2.13% (7.63 … 8.11 μs) 3.35 μs ± 3.59% (2.86 … 3.81 μs) 37.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 256, 64) Warps (1, 4) OP (16, 16, 16) 23.7 μs ± 0.672% (23.1 … 23.8 μs) 3.29 μs ± 3.77% (2.86 … 3.81 μs) 12.4
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (128, 64, 16) Warps (4, 4) OP (16, 16, 16) 7.61 μs ± 1.9% (7.39 … 7.87 μs) 3.54 μs ± 3.45% (3.1 … 4.05 μs) 41.9
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 256, 16) Warps (1, 4) OP (16, 16, 16) 13.1 μs ± 0.923% (12.6 … 13.4 μs) 3.34 μs ± 2.66% (2.86 … 3.81 μs) 22.6
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (256, 64, 32) Warps (2, 2) OP (16, 16, 16) 26.3 μs ± 0.624% (26.0 … 26.7 μs) 3.3 μs ± 3.04% (2.86 … 3.58 μs) 11.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 64, 32) Warps (2, 4) OP (16, 16, 16) 3.96 μs ± 3.83% (3.81 … 4.29 μs) 3.31 μs ± 2.96% (2.86 … 3.81 μs) 75.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 64, 32) Warps (1, 4) OP (16, 16, 16) 3.86 μs ± 2.66% (3.81 … 4.29 μs) 3.47 μs ± 3.96% (3.1 … 4.05 μs) 81.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (128, 128, 16) Warps (1, 4) OP (16, 16, 16) 11.6 μs ± 1.47% (11.4 … 11.9 μs) 3.47 μs ± 3.93% (3.1 … 3.81 μs) 27.1
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 256, 64) Warps (1, 4) OP (16, 16, 16) 24.6 μs ± 0.562% (24.3 … 25.0 μs) 3.4 μs ± 3.97% (2.86 … 3.81 μs) 11.8
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 256, 32) Warps (4, 2) OP (16, 16, 16) 10.5 μs ± 1.01% (10.5 … 11.0 μs) 3.31 μs ± 2.78% (2.86 … 3.58 μs) 27.3
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (128, 64, 32) Warps (4, 2) OP (16, 16, 16) 5.82 μs ± 2.76% (5.48 … 5.96 μs) 3.47 μs ± 3.93% (3.1 … 3.81 μs) 56.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (128, 64, 32) Warps (2, 2) OP (16, 16, 16) 5.45 μs ± 3.03% (5.25 … 5.72 μs) 3.38 μs ± 3.45% (3.1 … 3.81 μs) 59.1
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (128, 64, 64) Warps (2, 4) OP (16, 16, 16) 5.66 μs ± 2.47% (5.48 … 5.96 μs) 3.22 μs ± 4.34% (2.86 … 3.34 μs) 52.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 128, 16) Warps (4, 2) OP (16, 16, 16) 6.6 μs ± 1.86% (6.2 … 6.68 μs) 3.32 μs ± 4.04% (2.86 … 3.81 μs) 46.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 128, 16) Warps (2, 4) OP (16, 16, 16) 6.6 μs ± 1.91% (6.2 … 6.68 μs) 3.34 μs ± 2.58% (2.86 … 3.81 μs) 46.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 64, 64) Warps (4, 1) OP (16, 16, 16) 3.5 μs ± 3.67% (3.1 … 3.81 μs) 3.22 μs ± 4.44% (2.86 … 3.34 μs) 92.3
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 256, 64) Warps (4, 1) OP (16, 16, 16) 1.55 ms ± 0.186% (1.54 … 1.56 ms) 3.3 μs ± 3.73% (2.86 … 3.81 μs) 0.186
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (128, 64, 16) Warps (2, 4) OP (16, 16, 16) 8.04 μs ± 2.11% (7.87 … 8.34 μs) 3.47 μs ± 3.83% (3.1 … 3.81 μs) 39.4
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 128, 16) Warps (4, 4) OP (16, 16, 16) 6.75 μs ± 2.0% (6.44 … 6.91 μs) 3.3 μs ± 3.9% (2.86 … 3.81 μs) 44.4
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (128, 64, 16) Warps (2, 2) OP (16, 16, 16) 8.02 μs ± 2.01% (7.87 … 8.34 μs) 3.35 μs ± 2.84% (2.86 … 3.81 μs) 36.4
WMMA GEMM Float16*Float16+Float16=Float16 (512×512) · (512×512) (TN) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 56.5 μs ± 0.607% (55.6 … 58.2 μs) 6.58 μs ± 2.12% (6.2 … 6.91 μs) 11.2
WMMA GEMM Float16*Float16+Float16=Float16 (2048×2048) · (2048×2048) (TN) Block (128, 128, 64) Warps (4, 2) OP (8, 32, 16) 3.02 ms ± 0.646% (2.99 … 3.08 ms) 179.0 μs ± 0.353% (178.0 … 182.0 μs) 5.96
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (256, 64, 16) Warps (4, 4) OP (16, 16, 16) 11.0 μs ± 1.14% (10.7 … 11.4 μs) 3.34 μs ± 2.5% (2.86 … 3.58 μs) 26.7
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NT) OP (16, 16, 1), base shape (8, 4, 1) 41.5 μs ± 0.514% (40.8 … 42.4 μs) 3.73 μs ± 3.93% (3.58 … 4.05 μs) 8.77
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (128, 64, 32) Warps (1, 4) OP (16, 16, 16) 6.02 μs ± 2.74% (5.72 … 6.2 μs) 3.46 μs ± 3.9% (3.1 … 3.81 μs) 54.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 128, 16) Warps (2, 2) OP (16, 16, 16) 7.13 μs ± 1.47% (6.91 … 7.39 μs) 3.21 μs ± 4.5% (2.86 … 3.34 μs) 41.4
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (128, 64, 64) Warps (1, 4) OP (16, 16, 16) 8.34 μs ± 2.05% (8.11 … 8.58 μs) 3.4 μs ± 3.68% (2.86 … 3.81 μs) 35.3
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (128, 128, 32) Warps (4, 2) OP (16, 16, 16) 9.54 μs ± 1.61% (9.3 … 9.78 μs) 3.38 μs ± 3.2% (3.1 … 3.81 μs) 33.3
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (128, 64, 16) Warps (4, 4) OP (16, 16, 16) 7.26 μs ± 1.89% (7.15 … 7.63 μs) 3.35 μs ± 3.48% (2.86 … 3.81 μs) 40.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 64, 32) Warps (2, 4) OP (16, 16, 16) 3.64 μs ± 3.28% (3.58 … 4.05 μs) 3.28 μs ± 4.06% (2.86 … 3.81 μs) 80.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (256, 64, 64) Warps (4, 4) OP (16, 16, 16) 11.3 μs ± 1.36% (11.2 … 11.7 μs) 3.47 μs ± 3.84% (2.86 … 3.81 μs) 25.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 64, 32) Warps (1, 4) OP (16, 16, 16) 3.92 μs ± 3.65% (3.81 … 4.29 μs) 3.47 μs ± 3.91% (3.1 … 3.81 μs) 81.2
WMMA GEMM Float16*Float16+Float32=Float32 (128×256) · (256×128) (NT) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 31.0 μs ± 0.628% (30.5 … 31.7 μs) 2.37 μs ± 6.06% (2.15 … 2.62 μs) 7.03
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (128, 64, 16) Warps (4, 2) OP (16, 16, 16) 7.79 μs ± 2.08% (7.63 … 8.11 μs) 3.22 μs ± 4.4% (2.86 … 3.34 μs) 37.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (256, 64, 32) Warps (4, 1) OP (16, 16, 16) 22.3 μs ± 0.722% (21.9 … 22.6 μs) 3.3 μs ± 3.14% (2.86 … 3.81 μs) 13.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 128, 16) Warps (2, 4) OP (16, 16, 16) 7.53 μs ± 2.08% (7.39 … 7.87 μs) 3.36 μs ± 3.81% (2.86 … 3.81 μs) 38.7
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 64, 32) Warps (4, 1) OP (16, 16, 16) 3.85 μs ± 2.53% (3.58 … 4.29 μs) 3.22 μs ± 4.34% (2.86 … 3.34 μs) 80.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 128, 16) Warps (1, 4) OP (16, 16, 16) 7.66 μs ± 1.68% (7.39 … 8.11 μs) 3.38 μs ± 3.27% (3.1 … 3.81 μs) 41.9
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 256, 64) Warps (2, 2) OP (16, 16, 16) 29.6 μs ± 0.849% (29.1 … 30.3 μs) 3.39 μs ± 3.87% (2.86 … 3.81 μs) 9.84
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 256, 16) Warps (2, 4) OP (16, 16, 16) 12.6 μs ± 1.19% (12.2 … 12.9 μs) 3.35 μs ± 3.47% (2.86 … 3.81 μs) 23.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (128, 128, 64) Warps (1, 4) OP (16, 16, 16) 18.6 μs ± 0.986% (18.1 … 19.1 μs) 3.22 μs ± 4.29% (2.86 … 3.34 μs) 15.8
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 128, 16) Warps (4, 1) OP (16, 16, 16) 8.79 μs ± 1.96% (8.58 … 9.06 μs) 3.55 μs ± 3.45% (3.1 … 4.05 μs) 36.1
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 64, 64) Warps (2, 2) OP (16, 16, 16) 3.53 μs ± 3.23% (3.1 … 3.81 μs) 3.22 μs ± 4.37% (2.86 … 3.34 μs) 92.3
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 128, 64) Warps (2, 4) OP (16, 16, 16) 5.67 μs ± 2.6% (5.48 … 5.96 μs) 3.29 μs ± 3.87% (2.86 … 3.81 μs) 52.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (128, 64, 16) Warps (2, 2) OP (16, 16, 16) 7.78 μs ± 2.05% (7.63 … 8.11 μs) 3.47 μs ± 3.96% (3.1 … 4.05 μs) 40.6
FPU GEMM Float16*Float16=Float32 (2048×2048) · (2048×2048) (TT) OP (8, 16, 2), base shape (4, 8, 1) 15.7 ms ± 0.179% (15.6 … 15.7 ms) 290.0 μs ± 0.231% (288.0 … 291.0 μs) 1.85
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 128, 32) Warps (4, 2) OP (16, 16, 16) 5.44 μs ± 2.99% (5.25 … 5.72 μs) 3.21 μs ± 4.51% (2.86 … 3.34 μs) 54.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 256, 32) Warps (1, 4) OP (16, 16, 16) 17.8 μs ± 0.81% (17.6 … 18.1 μs) 3.47 μs ± 3.97% (3.1 … 3.81 μs) 17.6
WMMA GEMM Float16*Float16+Float16=Float16 (256×256) · (256×256) (NT) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 27.3 μs ± 0.591% (26.9 … 27.9 μs) 3.33 μs ± 2.78% (2.86 … 3.81 μs) 10.6
WMMA GEMM Float16*Float16+Float32=Float32 (128×128) · (128×128) (TN) Block (128, 128, 64) Warps (4, 2) OP (8, 32, 16) 17.7 μs ± 0.977% (17.4 … 18.1 μs) 2.07 μs ± 6.31% (1.91 … 2.38 μs) 11.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (256, 64, 32) Warps (2, 2) OP (16, 16, 16) 22.1 μs ± 0.786% (21.9 … 22.4 μs) 3.21 μs ± 4.38% (2.86 … 3.34 μs) 13.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (128, 128, 64) Warps (4, 1) OP (16, 16, 16) 18.5 μs ± 1.0% (18.1 … 19.1 μs) 3.21 μs ± 4.5% (2.86 … 3.34 μs) 15.8
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 256, 16) Warps (1, 4) OP (16, 16, 16) 12.1 μs ± 1.45% (11.9 … 12.4 μs) 3.31 μs ± 3.97% (2.86 … 3.81 μs) 24.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 256, 64) Warps (4, 2) OP (16, 16, 16) 12.3 μs ± 1.21% (11.9 … 12.6 μs) 3.22 μs ± 4.23% (2.86 … 3.34 μs) 24.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (128, 64, 32) Warps (2, 4) OP (16, 16, 16) 6.03 μs ± 2.76% (5.72 … 6.2 μs) 3.48 μs ± 3.79% (3.1 … 3.81 μs) 54.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (256, 64, 16) Warps (4, 1) OP (16, 16, 16) 18.1 μs ± 0.762% (17.9 … 18.6 μs) 3.38 μs ± 3.37% (3.1 … 3.81 μs) 17.3
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (128, 64, 32) Warps (4, 1) OP (16, 16, 16) 5.37 μs ± 2.33% (5.01 … 5.72 μs) 3.29 μs ± 3.77% (2.86 … 3.81 μs) 57.1
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (128, 128, 32) Warps (4, 1) OP (16, 16, 16) 11.9 μs ± 1.48% (11.7 … 12.2 μs) 3.38 μs ± 4.07% (2.86 … 3.81 μs) 24.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 64, 32) Warps (1, 4) OP (16, 16, 16) 3.66 μs ± 3.45% (3.58 … 4.05 μs) 3.39 μs ± 3.71% (3.1 … 3.81 μs) 86.7
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 64, 32) Warps (2, 4) OP (16, 16, 16) 3.66 μs ± 3.63% (3.58 … 4.05 μs) 3.38 μs ± 3.27% (3.1 … 3.81 μs) 86.7
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (256, 64, 64) Warps (4, 1) OP (16, 16, 16) 24.9 μs ± 0.578% (24.6 … 25.3 μs) 3.4 μs ± 4.08% (2.86 … 3.81 μs) 11.7
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (256, 64, 64) Warps (4, 4) OP (16, 16, 16) 10.3 μs ± 0.98% (10.0 … 10.7 μs) 3.31 μs ± 3.66% (2.86 … 3.81 μs) 28.6
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 128, 64) Warps (2, 4) OP (16, 16, 16) 5.64 μs ± 2.19% (5.25 … 5.96 μs) 3.37 μs ± 3.26% (3.1 … 3.81 μs) 59.1
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 256, 64) Warps (1, 4) OP (16, 16, 16) 18.2 μs ± 0.807% (17.9 … 18.6 μs) 3.22 μs ± 4.31% (2.86 … 3.34 μs) 16.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (256, 64, 64) Warps (1, 4) OP (16, 16, 16) 1.71 ms ± 0.114% (1.71 … 1.72 ms) 3.28 μs ± 3.87% (2.86 … 3.81 μs) 0.168
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (256, 64, 16) Warps (2, 4) OP (16, 16, 16) 13.6 μs ± 0.881% (13.4 … 13.8 μs) 3.5 μs ± 3.69% (3.1 … 3.81 μs) 23.2
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TT) OP (8, 16, 2), base shape (4, 8, 1) 45.3 μs ± 0.37% (44.8 … 46.0 μs) 3.56 μs ± 2.37% (3.1 … 3.81 μs) 6.91
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 64, 16) Warps (2, 4) OP (16, 16, 16) 5.15 μs ± 2.74% (5.01 … 5.48 μs) 3.32 μs ± 3.63% (2.86 … 3.81 μs) 57.1
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (256, 64, 16) Warps (1, 4) OP (16, 16, 16) 20.5 μs ± 0.593% (20.0 … 20.7 μs) 3.35 μs ± 3.48% (2.86 … 3.81 μs) 14.3
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 128, 64) Warps (2, 4) OP (16, 16, 16) 5.99 μs ± 2.75% (5.72 … 6.2 μs) 3.3 μs ± 3.03% (2.86 … 3.81 μs) 50.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (256, 64, 16) Warps (4, 4) OP (16, 16, 16) 13.6 μs ± 0.899% (13.1 … 13.8 μs) 3.38 μs ± 3.36% (3.1 … 3.81 μs) 23.6
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 128, 32) Warps (4, 2) OP (16, 16, 16) 5.85 μs ± 2.45% (5.48 … 5.96 μs) 3.37 μs ± 3.83% (2.86 … 3.81 μs) 52.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 128, 64) Warps (4, 2) OP (16, 16, 16) 5.97 μs ± 2.7% (5.72 … 6.2 μs) 3.3 μs ± 3.18% (2.86 … 3.58 μs) 50.0
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TT) OP (4, 8, 2), base shape (4, 8, 1) 43.2 μs ± 0.724% (42.4 … 44.3 μs) 3.56 μs ± 2.49% (3.1 … 3.81 μs) 7.3
WMMA GEMM Float16*Float16+Float32=Float32 (128×256) · (256×128) (TT) Block (128, 128, 64) Warps (4, 2) OP (32, 8, 16) 10.6 μs ± 1.32% (10.5 … 11.0 μs) 2.63 μs ± 5.49% (2.38 … 2.86 μs) 22.7
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (128, 128, 32) Warps (4, 2) OP (16, 16, 16) 10.3 μs ± 1.29% (10.0 … 10.7 μs) 3.31 μs ± 2.75% (2.86 … 3.58 μs) 28.6
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (128, 64, 64) Warps (4, 4) OP (16, 16, 16) 6.1 μs ± 2.34% (5.72 … 6.44 μs) 3.47 μs ± 3.79% (3.1 … 3.81 μs) 54.2
WMMA GEMM Float16*Float16+Float16=Float16 (256×512) · (512×256) (NT) Block (128, 128, 64) Warps (4, 2) OP (32, 8, 16) 11.1 μs ± 1.39% (11.0 … 11.4 μs) 3.92 μs ± 3.56% (3.81 … 4.29 μs) 34.8
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (256, 64, 32) Warps (4, 1) OP (16, 16, 16) 20.8 μs ± 0.543% (20.3 … 21.2 μs) 3.4 μs ± 3.79% (3.1 … 3.81 μs) 15.3
WMMA GEMM Float16*Float16+Float16=Float16 (2048×2048) · (2048×2048) (NN) Block (128, 128, 64) Warps (4, 2) OP (32, 8, 16) 204.0 μs ± 0.814% (200.0 … 209.0 μs) 175.0 μs ± 0.394% (174.0 … 178.0 μs) 86.9
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (128, 128, 64) Warps (4, 1) OP (16, 16, 16) 19.4 μs ± 0.974% (18.8 … 20.0 μs) 3.3 μs ± 3.21% (2.86 … 3.81 μs) 15.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 256, 16) Warps (2, 4) OP (16, 16, 16) 12.7 μs ± 1.28% (12.4 … 12.9 μs) 3.49 μs ± 3.77% (3.1 … 3.81 μs) 25.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 64, 64) Warps (1, 4) OP (16, 16, 16) 3.86 μs ± 2.74% (3.58 … 4.29 μs) 3.31 μs ± 3.14% (2.86 … 3.58 μs) 80.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (256, 64, 32) Warps (4, 2) OP (16, 16, 16) 10.0 μs ± 1.3% (9.78 … 10.3 μs) 3.39 μs ± 3.82% (2.86 … 3.81 μs) 29.3
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (256, 64, 32) Warps (2, 4) OP (16, 16, 16) 10.8 μs ± 1.07% (10.5 … 11.2 μs) 3.21 μs ± 4.34% (2.86 … 3.34 μs) 27.3
WMMA GEMM Float16*Float16+Float32=Float32 (2048×2048) · (2048×2048) (TT) Block (128, 128, 64) Warps (4, 2) OP (8, 32, 16) 3.16 ms ± 0.593% (3.12 … 3.22 ms) 289.0 μs ± 0.188% (288.0 … 292.0 μs) 9.22
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (128, 128, 64) Warps (4, 4) OP (16, 16, 16) 9.99 μs ± 0.9% (9.54 … 10.3 μs) 3.28 μs ± 4.15% (2.86 … 3.81 μs) 30.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (128, 64, 16) Warps (4, 2) OP (16, 16, 16) 7.0 μs ± 1.81% (6.68 … 7.39 μs) 3.48 μs ± 3.93% (3.1 … 4.05 μs) 46.4
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (128, 64, 16) Warps (1, 4) OP (16, 16, 16) 10.1 μs ± 1.43% (9.78 … 10.3 μs) 3.38 μs ± 3.38% (3.1 … 3.81 μs) 31.7
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (128, 64, 16) Warps (4, 2) OP (16, 16, 16) 7.44 μs ± 1.67% (7.15 … 7.87 μs) 3.35 μs ± 3.53% (2.86 … 3.81 μs) 40.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (128, 128, 16) Warps (4, 2) OP (16, 16, 16) 11.5 μs ± 1.28% (11.2 … 11.7 μs) 3.31 μs ± 3.99% (2.86 … 3.81 μs) 25.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (128, 64, 64) Warps (4, 4) OP (16, 16, 16) 6.14 μs ± 1.87% (5.72 … 6.44 μs) 3.38 μs ± 4.13% (2.86 … 3.81 μs) 50.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (128, 64, 32) Warps (1, 4) OP (16, 16, 16) 5.71 μs ± 2.98% (5.48 … 5.96 μs) 3.22 μs ± 4.39% (2.86 … 3.34 μs) 52.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 64, 32) Warps (2, 4) OP (16, 16, 16) 3.68 μs ± 3.71% (3.58 … 4.05 μs) 3.22 μs ± 4.25% (2.86 … 3.34 μs) 80.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 128, 64) Warps (1, 4) OP (16, 16, 16) 35.7 μs ± 0.686% (35.0 … 36.5 μs) 3.21 μs ± 4.44% (2.86 … 3.34 μs) 8.16
WMMA GEMM Float16*Float16+Float32=Float32 (128×128) · (128×128) (NT) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 16.6 μs ± 0.95% (16.2 … 16.9 μs) 1.99 μs ± 7.93% (1.67 … 2.15 μs) 10.3
WMMA GEMM Float16*Float16+Float32=Float32 (128×256) · (256×128) (NT) Block (128, 128, 64) Warps (4, 2) OP (32, 8, 16) 10.6 μs ± 1.31% (10.5 … 11.0 μs) 2.32 μs ± 5.84% (1.91 … 2.62 μs) 18.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (256, 64, 32) Warps (4, 4) OP (16, 16, 16) 11.1 μs ± 1.43% (11.0 … 11.4 μs) 3.48 μs ± 3.91% (3.1 … 4.05 μs) 28.3
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 128, 32) Warps (1, 4) OP (16, 16, 16) 5.46 μs ± 3.1% (5.25 … 5.72 μs) 3.22 μs ± 4.35% (2.86 … 3.34 μs) 54.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 128, 64) Warps (2, 4) OP (16, 16, 16) 5.96 μs ± 2.67% (5.72 … 6.2 μs) 3.47 μs ± 4.0% (3.1 … 4.05 μs) 54.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 31.5 μs ± 0.634% (31.0 … 32.4 μs) 3.47 μs ± 4.13% (3.1 … 4.05 μs) 10.0
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NT) OP (8, 8, 1), base shape (4, 8, 1) 44.3 μs ± 0.874% (43.4 … 46.0 μs) 3.72 μs ± 3.81% (3.58 … 4.05 μs) 8.24
WMMA GEMM Float16*Float16+Float16=Float16 (256×512) · (512×256) (NT) Block (128, 128, 64) Warps (4, 2) OP (8, 32, 16) 49.0 μs ± 0.38% (48.6 … 49.6 μs) 3.92 μs ± 3.68% (3.81 … 4.29 μs) 7.84
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (128, 128, 16) Warps (1, 4) OP (16, 16, 16) 15.6 μs ± 1.02% (15.3 … 15.7 μs) 3.38 μs ± 3.3% (3.1 … 3.81 μs) 20.3
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (256, 64, 16) Warps (4, 2) OP (16, 16, 16) 14.3 μs ± 0.966% (14.1 … 14.5 μs) 3.48 μs ± 3.9% (3.1 … 3.81 μs) 22.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (128, 64, 32) Warps (4, 4) OP (16, 16, 16) 5.79 μs ± 2.93% (5.48 … 5.96 μs) 3.38 μs ± 3.27% (3.1 … 3.81 μs) 56.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (128, 64, 32) Warps (1, 4) OP (16, 16, 16) 5.56 μs ± 2.95% (5.25 … 5.72 μs) 3.38 μs ± 3.39% (3.1 … 3.81 μs) 59.1
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 128, 16) Warps (4, 2) OP (16, 16, 16) 7.8 μs ± 2.06% (7.63 … 8.11 μs) 3.35 μs ± 3.72% (2.86 … 3.81 μs) 37.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 256, 32) Warps (4, 2) OP (16, 16, 16) 10.6 μs ± 1.25% (10.5 … 11.0 μs) 3.47 μs ± 3.94% (3.1 … 3.81 μs) 29.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (256, 64, 32) Warps (1, 4) OP (16, 16, 16) 75.9 μs ± 0.43% (74.9 … 76.8 μs) 3.47 μs ± 3.99% (3.1 … 3.81 μs) 4.14
WMMA GEMM Float16*Float16+Float16=Float16 (256×512) · (512×256) (TN) Block (128, 128, 64) Warps (4, 2) OP (8, 32, 16) 49.6 μs ± 0.435% (49.1 … 50.3 μs) 4.59 μs ± 3.9% (4.29 … 5.01 μs) 8.74
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (256, 64, 64) Warps (4, 2) OP (16, 16, 16) 49.5 μs ± 0.579% (48.9 … 50.5 μs) 3.39 μs ± 4.06% (2.86 … 3.81 μs) 5.85
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 64, 64) Warps (4, 2) OP (16, 16, 16) 3.46 μs ± 4.0% (3.1 … 3.81 μs) 3.39 μs ± 3.98% (2.86 … 3.81 μs) 92.3
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 128, 32) Warps (4, 4) OP (16, 16, 16) 6.25 μs ± 2.5% (5.96 … 6.44 μs) 3.47 μs ± 3.83% (3.1 … 3.81 μs) 52.0
FPU GEMM Float64*Float64=Float64 (128×128) · (128×128) (TN) OP (8, 16, 2), base shape (4, 8, 1) 140.0 μs ± 0.158% (140.0 … 141.0 μs) 28.8 μs ± 0.583% (28.6 … 29.1 μs) 20.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 256, 16) Warps (2, 4) OP (16, 16, 16) 11.9 μs ± 1.43% (11.7 … 12.2 μs) 3.21 μs ± 4.43% (2.86 … 3.34 μs) 24.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (128, 64, 32) Warps (4, 2) OP (16, 16, 16) 5.61 μs ± 2.5% (5.25 … 5.72 μs) 3.29 μs ± 4.05% (2.86 … 3.81 μs) 54.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 128, 16) Warps (2, 2) OP (16, 16, 16) 8.15 μs ± 1.8% (7.87 … 8.34 μs) 3.34 μs ± 2.84% (2.86 … 3.81 μs) 36.4
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 128, 32) Warps (2, 2) OP (16, 16, 16) 5.43 μs ± 2.89% (5.25 … 5.72 μs) 3.22 μs ± 4.34% (2.86 … 3.34 μs) 54.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (256, 64, 64) Warps (4, 4) OP (16, 16, 16) 10.3 μs ± 1.28% (10.0 … 10.7 μs) 3.39 μs ± 3.87% (2.86 … 4.05 μs) 28.6
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 128, 64) Warps (4, 2) OP (16, 16, 16) 5.66 μs ± 2.5% (5.48 … 5.96 μs) 3.38 μs ± 3.24% (3.1 … 3.81 μs) 56.5
WMMA GEMM Float16*Float16+Float32=Float32 (128×128) · (128×128) (TN) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 16.3 μs ± 0.963% (16.0 … 16.7 μs) 2.07 μs ± 6.49% (1.91 … 2.38 μs) 11.9
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 64, 16) Warps (4, 1) OP (16, 16, 16) 5.83 μs ± 2.75% (5.48 … 5.96 μs) 3.48 μs ± 3.91% (3.1 … 3.81 μs) 56.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 64, 64) Warps (2, 4) OP (16, 16, 16) 3.86 μs ± 2.83% (3.58 … 4.29 μs) 3.31 μs ± 2.87% (2.86 … 3.58 μs) 80.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 256, 32) Warps (2, 2) OP (16, 16, 16) 21.7 μs ± 0.747% (21.5 … 21.9 μs) 3.48 μs ± 3.87% (3.1 … 4.05 μs) 14.4
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TT) OP (16, 16, 1), base shape (8, 4, 1) 41.2 μs ± 0.515% (40.5 … 41.7 μs) 3.57 μs ± 2.32% (3.34 … 3.81 μs) 8.24
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NN) OP (32, 4, 1), base shape (16, 2, 1) 58.4 μs ± 0.348% (57.7 … 59.1 μs) 3.61 μs ± 2.37% (3.58 … 4.05 μs) 6.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (128, 64, 64) Warps (4, 1) OP (16, 16, 16) 31.8 μs ± 0.687% (31.2 … 32.4 μs) 3.31 μs ± 2.92% (2.86 … 3.58 μs) 9.16
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 256, 16) Warps (4, 4) OP (16, 16, 16) 11.1 μs ± 1.37% (10.7 … 11.4 μs) 3.31 μs ± 4.1% (2.86 … 3.81 μs) 26.7
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 64, 32) Warps (4, 1) OP (16, 16, 16) 3.65 μs ± 3.42% (3.58 … 4.05 μs) 3.38 μs ± 3.15% (3.1 … 3.81 μs) 86.7
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 64, 64) Warps (4, 4) OP (16, 16, 16) 3.95 μs ± 3.81% (3.81 … 4.29 μs) 3.48 μs ± 3.99% (3.1 … 4.05 μs) 81.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (128, 128, 64) Warps (1, 4) OP (16, 16, 16) 19.4 μs ± 0.976% (18.8 … 19.8 μs) 3.3 μs ± 3.06% (2.86 … 3.81 μs) 15.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 64, 32) Warps (4, 2) OP (16, 16, 16) 3.6 μs ± 2.29% (3.34 … 4.05 μs) 3.38 μs ± 3.31% (2.86 … 3.81 μs) 85.7
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 64, 16) Warps (2, 4) OP (16, 16, 16) 5.71 μs ± 2.9% (5.48 … 5.96 μs) 3.48 μs ± 3.81% (3.1 … 3.81 μs) 56.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 64, 32) Warps (4, 1) OP (16, 16, 16) 4.1 μs ± 2.62% (3.81 … 4.53 μs) 3.31 μs ± 3.11% (2.86 … 3.58 μs) 75.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (256, 64, 64) Warps (2, 4) OP (16, 16, 16) 15.3 μs ± 1.14% (15.0 … 15.7 μs) 3.3 μs ± 2.94% (2.86 … 3.58 μs) 19.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 128, 32) Warps (1, 4) OP (16, 16, 16) 5.97 μs ± 2.61% (5.72 … 6.2 μs) 3.48 μs ± 4.05% (3.1 … 3.81 μs) 54.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (256, 64, 64) Warps (2, 4) OP (16, 16, 16) 17.2 μs ± 0.891% (16.7 … 17.6 μs) 3.47 μs ± 3.82% (3.1 … 3.81 μs) 18.6
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 256, 64) Warps (1, 4) OP (16, 16, 16) 18.9 μs ± 0.813% (18.6 … 19.3 μs) 3.46 μs ± 4.02% (3.1 … 3.81 μs) 16.7
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (128, 64, 32) Warps (2, 4) OP (16, 16, 16) 5.53 μs ± 3.11% (5.25 … 5.72 μs) 3.39 μs ± 3.78% (2.86 … 3.81 μs) 54.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (128, 128, 64) Warps (2, 2) OP (16, 16, 16) 16.3 μs ± 1.02% (16.0 … 16.7 μs) 3.21 μs ± 4.4% (2.86 … 3.34 μs) 17.9
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 64, 16) Warps (4, 2) OP (16, 16, 16) 5.11 μs ± 2.5% (4.77 … 5.25 μs) 3.35 μs ± 3.83% (2.86 … 3.81 μs) 60.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 64, 64) Warps (4, 4) OP (16, 16, 16) 3.91 μs ± 3.6% (3.81 … 4.29 μs) 3.47 μs ± 3.87% (3.1 … 3.81 μs) 81.2
WMMA GEMM Float16*Float16+Float32=Float32 (2048×2048) · (2048×2048) (TT) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 3.18 ms ± 0.261% (3.16 … 3.21 ms) 289.0 μs ± 0.178% (288.0 … 292.0 μs) 9.12
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (128, 128, 32) Warps (4, 1) OP (16, 16, 16) 14.0 μs ± 0.99% (13.6 … 14.5 μs) 3.39 μs ± 3.82% (2.86 … 3.81 μs) 21.1
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 64, 16) Warps (4, 1) OP (16, 16, 16) 5.06 μs ± 3.14% (4.77 … 5.25 μs) 3.47 μs ± 3.87% (3.1 … 3.81 μs) 65.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (256, 64, 16) Warps (2, 2) OP (16, 16, 16) 16.7 μs ± 0.825% (16.5 … 16.9 μs) 3.31 μs ± 3.86% (2.86 … 3.81 μs) 17.4
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (256, 64, 64) Warps (4, 4) OP (16, 16, 16) 10.3 μs ± 1.01% (9.78 … 10.7 μs) 3.22 μs ± 4.29% (2.86 … 3.34 μs) 29.3
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NN) OP (16, 16, 1), base shape (8, 4, 1) 41.6 μs ± 0.513% (41.0 … 42.4 μs) 3.6 μs ± 2.21% (3.58 … 4.05 μs) 8.72
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TT) OP (16, 16, 1), base shape (4, 8, 1) 40.4 μs ± 0.379% (40.1 … 40.8 μs) 3.55 μs ± 2.79% (3.1 … 4.05 μs) 7.74
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (256, 64, 16) Warps (4, 4) OP (16, 16, 16) 11.1 μs ± 1.45% (10.7 … 11.4 μs) 3.47 μs ± 4.07% (3.1 … 4.05 μs) 28.9
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 64, 16) Warps (1, 4) OP (16, 16, 16) 5.28 μs ± 3.29% (5.01 … 5.48 μs) 3.48 μs ± 4.09% (3.1 … 4.05 μs) 61.9
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (128, 64, 16) Warps (2, 2) OP (16, 16, 16) 7.64 μs ± 1.78% (7.39 … 8.11 μs) 3.38 μs ± 3.59% (3.1 … 3.81 μs) 41.9
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 256, 64) Warps (2, 2) OP (16, 16, 16) 21.9 μs ± 0.858% (21.5 … 22.4 μs) 3.47 μs ± 3.88% (3.1 … 3.81 μs) 14.4
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 64, 32) Warps (2, 2) OP (16, 16, 16) 4.14 μs ± 3.35% (4.05 … 4.53 μs) 3.3 μs ± 3.03% (2.86 … 3.58 μs) 70.6
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 64, 64) Warps (2, 4) OP (16, 16, 16) 3.77 μs ± 3.45% (3.58 … 4.05 μs) 3.47 μs ± 4.16% (3.1 … 4.05 μs) 86.7
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (128, 128, 16) Warps (2, 2) OP (16, 16, 16) 13.2 μs ± 1.21% (12.9 … 13.4 μs) 3.35 μs ± 3.39% (2.86 … 3.81 μs) 22.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (128, 128, 32) Warps (1, 4) OP (16, 16, 16) 10.7 μs ± 1.0% (10.5 … 11.2 μs) 3.21 μs ± 4.49% (2.86 … 3.34 μs) 27.3
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (128, 64, 32) Warps (4, 1) OP (16, 16, 16) 5.88 μs ± 2.02% (5.48 … 6.2 μs) 3.37 μs ± 3.86% (2.86 … 3.81 μs) 52.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 256, 64) Warps (4, 1) OP (16, 16, 16) 1.53 ms ± 0.144% (1.52 … 1.53 ms) 3.39 μs ± 3.92% (2.86 … 3.81 μs) 0.188
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (256, 64, 32) Warps (4, 4) OP (16, 16, 16) 11.1 μs ± 1.29% (10.7 … 11.4 μs) 3.46 μs ± 3.95% (3.1 … 3.81 μs) 28.9
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 128, 64) Warps (4, 2) OP (16, 16, 16) 6.0 μs ± 2.71% (5.72 … 6.2 μs) 3.48 μs ± 4.03% (3.1 … 3.81 μs) 54.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (128, 128, 64) Warps (1, 4) OP (16, 16, 16) 19.5 μs ± 0.982% (19.1 … 20.0 μs) 3.37 μs ± 4.02% (2.86 … 3.81 μs) 15.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 64, 16) Warps (1, 4) OP (16, 16, 16) 5.87 μs ± 2.21% (5.48 … 6.2 μs) 3.34 μs ± 3.36% (2.86 … 3.81 μs) 52.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (128, 128, 32) Warps (4, 1) OP (16, 16, 16) 11.0 μs ± 1.09% (10.7 … 11.4 μs) 3.29 μs ± 3.94% (2.86 … 3.81 μs) 26.7
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (128, 64, 32) Warps (1, 4) OP (16, 16, 16) 6.15 μs ± 1.72% (5.72 … 6.44 μs) 3.31 μs ± 2.96% (2.86 … 3.81 μs) 50.0
WMMA GEMM Float16*Float16+Float16=Float16 (512×256) · (256×512) (TN) Block (128, 128, 64) Warps (4, 2) OP (8, 32, 16) 29.8 μs ± 0.709% (29.1 … 30.5 μs) 4.47 μs ± 3.84% (4.29 … 4.77 μs) 14.8
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (128, 128, 16) Warps (1, 4) OP (16, 16, 16) 12.5 μs ± 0.996% (12.2 … 12.9 μs) 3.55 μs ± 3.54% (3.1 … 4.05 μs) 25.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (128, 64, 16) Warps (4, 1) OP (16, 16, 16) 6.7 μs ± 1.58% (6.44 … 6.91 μs) 3.34 μs ± 2.66% (2.86 … 3.81 μs) 44.4
WMMA GEMM Float16*Float16+Float32=Float32 (256×128) · (128×256) (TT) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 16.8 μs ± 0.947% (16.5 … 17.2 μs) 2.52 μs ± 5.79% (2.15 … 2.62 μs) 13.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 256, 32) Warps (1, 4) OP (16, 16, 16) 20.5 μs ± 0.507% (20.0 … 20.7 μs) 3.46 μs ± 4.17% (3.1 … 4.05 μs) 15.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (128, 128, 16) Warps (4, 4) OP (16, 16, 16) 12.0 μs ± 1.23% (11.7 … 12.4 μs) 3.55 μs ± 3.39% (3.1 … 4.05 μs) 26.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 128, 16) Warps (2, 4) OP (16, 16, 16) 7.6 μs ± 1.97% (7.39 … 7.87 μs) 3.56 μs ± 3.53% (3.1 … 4.05 μs) 41.9
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 256, 32) Warps (4, 4) OP (16, 16, 16) 10.3 μs ± 1.15% (10.0 … 10.7 μs) 3.38 μs ± 3.32% (3.1 … 3.81 μs) 31.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (128, 128, 64) Warps (1, 4) OP (16, 16, 16) 18.5 μs ± 1.0% (18.1 … 19.1 μs) 3.38 μs ± 3.79% (2.86 … 3.81 μs) 15.8
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 64, 64) Warps (2, 2) OP (16, 16, 16) 3.76 μs ± 3.53% (3.58 … 4.05 μs) 3.31 μs ± 2.97% (2.86 … 3.81 μs) 80.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 64, 64) Warps (4, 4) OP (16, 16, 16) 3.93 μs ± 3.71% (3.81 … 4.29 μs) 3.37 μs ± 3.97% (2.86 … 3.81 μs) 75.0
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NT) OP (4, 8, 1), base shape (4, 8, 1) 52.1 μs ± 0.923% (50.8 … 53.9 μs) 3.71 μs ± 3.81% (3.58 … 4.05 μs) 7.04
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (128, 128, 64) Warps (4, 1) OP (16, 16, 16) 19.1 μs ± 1.0% (18.6 … 19.6 μs) 3.27 μs ± 4.13% (2.86 … 3.58 μs) 15.4
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (128, 64, 32) Warps (2, 2) OP (16, 16, 16) 5.92 μs ± 2.22% (5.72 … 6.2 μs) 3.48 μs ± 3.87% (3.1 … 4.05 μs) 54.2
WMMA GEMM Float16*Float16+Float32=Float32 (128×128) · (128×128) (TT) Block (128, 128, 64) Warps (4, 2) OP (8, 32, 16) 17.6 μs ± 0.713% (17.4 … 18.1 μs) 2.16 μs ± 7.73% (1.91 … 2.38 μs) 11.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 256, 32) Warps (1, 4) OP (16, 16, 16) 18.2 μs ± 0.767% (17.9 … 18.6 μs) 3.3 μs ± 3.16% (2.86 … 3.58 μs) 16.0
WMMA GEMM Float16*Float16+Float32=Float32 (128×256) · (256×128) (TN) Block (128, 128, 64) Warps (4, 2) OP (8, 32, 16) 30.2 μs ± 0.631% (29.8 … 31.0 μs) 2.49 μs ± 6.21% (2.15 … 2.62 μs) 7.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 256, 64) Warps (4, 1) OP (16, 16, 16) 1.55 ms ± 0.161% (1.54 … 1.56 ms) 3.39 μs ± 3.86% (2.86 … 3.81 μs) 0.185
FPU GEMM Float32*Float32=Float32 (256×256) · (256×256) (TT) OP (8, 16, 2), base shape (4, 8, 1) 73.8 μs ± 0.297% (73.2 … 74.4 μs) 6.27 μs ± 2.55% (5.96 … 6.44 μs) 8.14
WMMA GEMM Float16*Float16+Float32=Float32 (2048×2048) · (2048×2048) (TN) Block (128, 128, 64) Warps (4, 2) OP (32, 8, 16) 344.0 μs ± 0.416% (340.0 … 349.0 μs) 290.0 μs ± 0.162% (289.0 … 292.0 μs) 85.1
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (128, 64, 64) Warps (2, 2) OP (16, 16, 16) 35.9 μs ± 0.801% (35.3 … 36.7 μs) 3.21 μs ± 4.6% (2.86 … 3.34 μs) 8.11
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (128, 128, 64) Warps (2, 4) OP (16, 16, 16) 30.5 μs ± 0.719% (29.8 … 31.2 μs) 3.38 μs ± 3.98% (2.86 … 3.81 μs) 9.6
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (128, 128, 64) Warps (4, 2) OP (32, 8, 16) 10.7 μs ± 1.17% (10.5 … 11.0 μs) 3.27 μs ± 3.85% (2.86 … 3.81 μs) 27.3
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (256, 64, 16) Warps (4, 2) OP (16, 16, 16) 12.1 μs ± 1.41% (11.9 … 12.4 μs) 3.3 μs ± 3.8% (2.86 … 3.81 μs) 24.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (256, 64, 64) Warps (4, 2) OP (16, 16, 16) 47.3 μs ± 0.664% (46.5 … 48.2 μs) 3.21 μs ± 4.52% (2.86 … 3.34 μs) 6.15
WMMA GEMM Float16*Float16+Float16=Float16 (256×512) · (512×256) (TT) Block (128, 128, 64) Warps (4, 2) OP (32, 8, 16) 11.0 μs ± 1.25% (10.7 … 11.4 μs) 4.84 μs ± 3.1% (4.53 … 5.25 μs) 42.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (128, 64, 32) Warps (4, 4) OP (16, 16, 16) 6.8 μs ± 1.98% (6.44 … 6.91 μs) 3.39 μs ± 4.06% (2.86 … 3.81 μs) 44.4
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 64, 32) Warps (4, 1) OP (16, 16, 16) 3.89 μs ± 3.22% (3.81 … 4.29 μs) 3.47 μs ± 3.89% (3.1 … 4.05 μs) 81.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (256, 64, 16) Warps (2, 4) OP (16, 16, 16) 11.6 μs ± 1.39% (11.4 … 11.9 μs) 3.34 μs ± 2.71% (3.1 … 3.81 μs) 27.1
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 256, 32) Warps (4, 4) OP (16, 16, 16) 10.3 μs ± 0.954% (10.0 … 10.7 μs) 3.4 μs ± 4.3% (2.86 … 4.05 μs) 28.6
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 64, 64) Warps (2, 2) OP (16, 16, 16) 3.47 μs ± 3.91% (3.1 … 3.81 μs) 3.38 μs ± 3.33% (2.86 … 3.81 μs) 92.3
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 256, 32) Warps (4, 1) OP (16, 16, 16) 65.7 μs ± 0.435% (64.6 … 66.5 μs) 3.39 μs ± 3.73% (2.86 … 3.81 μs) 4.43
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 64, 32) Warps (4, 1) OP (16, 16, 16) 3.94 μs ± 3.8% (3.81 … 4.29 μs) 3.47 μs ± 4.09% (3.1 … 3.81 μs) 81.2
WMMA GEMM Float16*Float16+Float32=Float32 (128×128) · (128×128) (TT) Block (128, 128, 64) Warps (4, 2) OP (32, 8, 16) 7.57 μs ± 2.1% (7.39 … 7.87 μs) 2.16 μs ± 7.94% (1.91 … 2.38 μs) 25.8
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (256, 64, 16) Warps (4, 1) OP (16, 16, 16) 17.3 μs ± 0.787% (16.9 … 17.6 μs) 3.55 μs ± 3.82% (3.1 … 4.05 μs) 18.3
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 128, 16) Warps (4, 1) OP (16, 16, 16) 8.14 μs ± 1.92% (7.87 … 8.34 μs) 3.35 μs ± 2.68% (2.86 … 3.81 μs) 36.4
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 256, 16) Warps (4, 2) OP (16, 16, 16) 11.6 μs ± 1.24% (11.4 … 11.9 μs) 3.3 μs ± 3.72% (2.86 … 3.81 μs) 25.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (128, 128, 64) Warps (4, 1) OP (16, 16, 16) 18.4 μs ± 0.986% (18.1 … 19.1 μs) 3.39 μs ± 3.91% (2.86 … 3.81 μs) 15.8
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 128, 32) Warps (4, 1) OP (16, 16, 16) 5.9 μs ± 2.1% (5.72 … 6.2 μs) 3.48 μs ± 4.01% (3.1 … 4.05 μs) 54.2
FPU GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) OP (8, 16, 2), base shape (4, 8, 1) 43.2 μs ± 0.566% (42.4 … 43.9 μs) 3.53 μs ± 3.69% (3.1 … 4.05 μs) 7.3
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (256, 64, 16) Warps (2, 2) OP (16, 16, 16) 21.0 μs ± 0.509% (20.7 … 21.5 μs) 3.35 μs ± 3.45% (2.86 … 3.81 μs) 13.8
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 128, 64) Warps (1, 4) OP (16, 16, 16) 36.5 μs ± 0.68% (35.8 … 37.4 μs) 3.38 μs ± 4.17% (2.86 … 3.81 μs) 8.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 64, 32) Warps (2, 2) OP (16, 16, 16) 3.65 μs ± 3.51% (3.58 … 4.05 μs) 3.37 μs ± 3.19% (3.1 … 3.81 μs) 86.7
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 128, 64) Warps (2, 2) OP (16, 16, 16) 35.8 μs ± 0.8% (35.0 … 36.7 μs) 3.38 μs ± 3.35% (3.1 … 3.81 μs) 8.84
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 64, 16) Warps (4, 2) OP (16, 16, 16) 5.48 μs ± 3.11% (5.25 … 5.72 μs) 3.38 μs ± 3.32% (3.1 … 3.81 μs) 59.1
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 256, 16) Warps (4, 4) OP (16, 16, 16) 13.4 μs ± 1.15% (13.1 … 13.8 μs) 3.38 μs ± 3.4% (3.1 … 3.81 μs) 23.6
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (128, 64, 16) Warps (4, 1) OP (16, 16, 16) 6.98 μs ± 1.68% (6.68 … 7.39 μs) 3.3 μs ± 3.69% (2.86 … 3.58 μs) 42.9
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (128, 128, 16) Warps (2, 2) OP (16, 16, 16) 12.3 μs ± 1.07% (11.9 … 12.6 μs) 3.31 μs ± 3.68% (2.86 … 3.81 μs) 24.0
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TT) OP (4, 32, 1), base shape (2, 16, 1) 60.9 μs ± 0.368% (60.3 … 61.5 μs) 3.56 μs ± 2.59% (3.1 … 3.81 μs) 5.14
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (128, 128, 32) Warps (2, 4) OP (16, 16, 16) 9.56 μs ± 1.66% (9.3 … 9.78 μs) 3.29 μs ± 3.74% (2.86 … 3.58 μs) 30.8
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (256, 64, 64) Warps (4, 2) OP (16, 16, 16) 53.9 μs ± 0.479% (53.2 … 54.6 μs) 3.39 μs ± 3.8% (2.86 … 3.81 μs) 5.38
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 128, 32) Warps (2, 4) OP (16, 16, 16) 5.33 μs ± 2.81% (5.01 … 5.48 μs) 3.3 μs ± 3.91% (2.86 … 3.81 μs) 57.1
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (128, 64, 16) Warps (4, 2) OP (16, 16, 16) 7.35 μs ± 1.82% (7.15 … 7.63 μs) 3.34 μs ± 2.6% (3.1 … 3.81 μs) 43.3
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (128, 64, 64) Warps (2, 4) OP (16, 16, 16) 6.16 μs ± 2.0% (5.96 … 6.44 μs) 3.48 μs ± 4.04% (3.1 … 4.05 μs) 52.0
WMMA GEMM Float16*Float16+Float16=Float16 (512×256) · (256×512) (NT) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 30.6 μs ± 0.641% (30.0 … 31.2 μs) 4.34 μs ± 2.97% (4.05 … 4.77 μs) 13.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 64, 16) Warps (2, 2) OP (16, 16, 16) 5.24 μs ± 3.34% (5.01 … 5.48 μs) 3.49 μs ± 3.72% (3.1 … 3.81 μs) 61.9
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (128, 64, 32) Warps (1, 4) OP (16, 16, 16) 6.11 μs ± 2.27% (5.72 … 6.44 μs) 3.38 μs ± 4.02% (2.86 … 3.81 μs) 50.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (128, 128, 16) Warps (1, 4) OP (16, 16, 16) 14.6 μs ± 0.807% (14.3 … 15.0 μs) 3.34 μs ± 2.59% (2.86 … 3.81 μs) 20.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 128, 32) Warps (4, 1) OP (16, 16, 16) 5.41 μs ± 2.73% (5.25 … 5.72 μs) 3.29 μs ± 3.87% (2.86 … 3.81 μs) 54.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (256, 64, 64) Warps (1, 4) OP (16, 16, 16) 1.72 ms ± 0.106% (1.71 … 1.72 ms) 3.38 μs ± 3.83% (2.86 … 3.81 μs) 0.167
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (128, 128, 64) Warps (4, 2) OP (8, 32, 16) 30.6 μs ± 0.73% (30.0 … 31.2 μs) 3.55 μs ± 3.52% (3.1 … 4.05 μs) 10.3
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (256, 64, 32) Warps (4, 4) OP (16, 16, 16) 11.8 μs ± 1.15% (11.4 … 12.2 μs) 3.3 μs ± 3.09% (2.86 … 3.58 μs) 25.0
WMMA GEMM Float16*Float16+Float16=Float16 (256×512) · (512×256) (TN) Block (128, 128, 64) Warps (4, 2) OP (32, 8, 16) 11.1 μs ± 1.35% (10.7 … 11.4 μs) 4.59 μs ± 3.93% (4.29 … 5.01 μs) 40.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 64, 64) Warps (2, 4) OP (16, 16, 16) 3.86 μs ± 2.74% (3.58 … 4.29 μs) 3.37 μs ± 3.98% (2.86 … 3.81 μs) 80.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (256, 64, 64) Warps (4, 1) OP (16, 16, 16) 20.4 μs ± 0.678% (20.0 … 20.7 μs) 3.22 μs ± 4.37% (2.86 … 3.34 μs) 14.3
FPU GEMM Float32*Float32=Float32 (256×256) · (256×256) (NN) OP (8, 16, 2), base shape (4, 8, 1) 73.6 μs ± 0.279% (73.2 … 74.1 μs) 6.0 μs ± 2.71% (5.72 … 6.2 μs) 7.82
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 256, 16) Warps (4, 1) OP (16, 16, 16) 16.1 μs ± 0.908% (15.7 … 16.2 μs) 3.32 μs ± 4.03% (2.86 … 3.81 μs) 18.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 256, 32) Warps (4, 1) OP (16, 16, 16) 66.8 μs ± 0.512% (65.8 … 67.9 μs) 3.47 μs ± 3.87% (3.1 … 3.81 μs) 4.71
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 256, 32) Warps (1, 4) OP (16, 16, 16) 20.1 μs ± 0.797% (19.8 … 20.3 μs) 3.37 μs ± 4.08% (2.86 … 3.81 μs) 14.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 128, 64) Warps (4, 1) OP (16, 16, 16) 7.92 μs ± 1.67% (7.63 … 8.11 μs) 3.22 μs ± 4.47% (2.86 … 3.34 μs) 37.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 64, 64) Warps (2, 4) OP (16, 16, 16) 3.55 μs ± 2.84% (3.1 … 3.81 μs) 3.29 μs ± 4.05% (2.86 … 3.81 μs) 92.3
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 256, 32) Warps (2, 4) OP (16, 16, 16) 9.86 μs ± 1.55% (9.54 … 10.0 μs) 3.38 μs ± 3.25% (3.1 … 3.81 μs) 32.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 64, 32) Warps (2, 2) OP (16, 16, 16) 3.57 μs ± 2.07% (3.1 … 3.81 μs) 3.39 μs ± 3.88% (2.86 … 3.81 μs) 92.3
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 64, 32) Warps (1, 4) OP (16, 16, 16) 3.7 μs ± 3.88% (3.58 … 4.05 μs) 3.21 μs ± 4.22% (2.86 … 3.34 μs) 80.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 128, 64) Warps (2, 4) OP (16, 16, 16) 5.99 μs ± 2.83% (5.72 … 6.2 μs) 3.37 μs ± 3.75% (2.86 … 3.81 μs) 50.0
FPU GEMM Float16*Float16=Float32 (2048×2048) · (2048×2048) (TN) OP (8, 16, 2), base shape (4, 8, 1) 15.6 ms ± 0.225% (15.5 … 15.7 ms) 291.0 μs ± 0.145% (290.0 … 294.0 μs) 1.87
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 128, 16) Warps (4, 4) OP (16, 16, 16) 6.78 μs ± 2.05% (6.44 … 6.91 μs) 3.34 μs ± 2.65% (3.1 … 3.81 μs) 48.1
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 64, 64) Warps (2, 2) OP (16, 16, 16) 3.78 μs ± 3.35% (3.58 … 4.05 μs) 3.48 μs ± 3.89% (3.1 … 4.05 μs) 86.7
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 256, 64) Warps (4, 1) OP (16, 16, 16) 1.36 ms ± 0.173% (1.35 … 1.37 ms) 3.3 μs ± 3.05% (2.86 … 3.58 μs) 0.211
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 128, 16) Warps (4, 2) OP (16, 16, 16) 7.76 μs ± 2.0% (7.63 … 8.11 μs) 3.55 μs ± 3.53% (3.1 … 4.05 μs) 40.6
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 128, 64) Warps (4, 4) OP (16, 16, 16) 5.7 μs ± 2.83% (5.48 … 5.96 μs) 3.22 μs ± 4.38% (2.86 … 3.34 μs) 52.2
WMMA GEMM Float16*Float16+Float16=Float16 (2048×2048) · (2048×2048) (NT) Block (128, 128, 64) Warps (4, 2) OP (8, 32, 16) 2.99 ms ± 0.423% (2.96 … 3.05 ms) 177.0 μs ± 1.55% (173.0 … 183.0 μs) 5.84
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (256, 64, 64) Warps (2, 2) OP (16, 16, 16) 30.4 μs ± 0.693% (29.8 … 31.0 μs) 3.39 μs ± 3.81% (2.86 … 3.81 μs) 9.6
WMMA GEMM Float16*Float16+Float16=Float16 (256×256) · (256×256) (NT) Block (128, 128, 64) Warps (4, 2) OP (8, 32, 16) 26.2 μs ± 0.732% (25.7 … 26.7 μs) 3.34 μs ± 2.71% (2.86 … 3.81 μs) 11.1
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (256, 64, 16) Warps (4, 1) OP (16, 16, 16) 13.3 μs ± 0.815% (13.1 … 13.6 μs) 3.48 μs ± 3.79% (3.1 … 3.81 μs) 23.6
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (128, 128, 32) Warps (1, 4) OP (16, 16, 16) 11.1 μs ± 1.38% (11.0 … 11.4 μs) 3.31 μs ± 2.82% (2.86 … 3.58 μs) 26.1
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (256, 64, 64) Warps (2, 4) OP (16, 16, 16) 17.4 μs ± 0.761% (16.9 … 17.9 μs) 3.47 μs ± 4.07% (3.1 … 4.05 μs) 18.3
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (128, 128, 64) Warps (4, 2) OP (32, 8, 16) 10.7 μs ± 0.804% (10.5 … 11.2 μs) 3.55 μs ± 3.46% (3.1 … 4.05 μs) 29.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 128, 64) Warps (4, 1) OP (16, 16, 16) 8.54 μs ± 2.02% (8.34 … 8.82 μs) 3.38 μs ± 3.98% (2.86 … 3.81 μs) 34.3
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (128, 128, 16) Warps (4, 1) OP (16, 16, 16) 14.4 μs ± 1.04% (14.3 … 14.8 μs) 3.34 μs ± 2.35% (3.1 … 3.58 μs) 21.7
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 256, 64) Warps (4, 2) OP (16, 16, 16) 13.2 μs ± 1.2% (12.9 … 13.6 μs) 3.47 μs ± 3.94% (3.1 … 3.81 μs) 24.1
WMMA GEMM Float16*Float16+Float16=Float16 (512×256) · (256×512) (TT) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 30.6 μs ± 0.652% (30.0 … 31.0 μs) 4.46 μs ± 3.64% (4.29 … 4.77 μs) 14.3
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (256, 64, 16) Warps (1, 4) OP (16, 16, 16) 22.6 μs ± 0.749% (22.4 … 22.9 μs) 3.55 μs ± 3.31% (3.1 … 4.05 μs) 13.8
FPU GEMM Float64*Float64=Float64 (256×256) · (256×256) (NN) OP (8, 16, 2), base shape (4, 8, 1) 275.0 μs ± 0.0875% (274.0 … 276.0 μs) 109.0 μs ± 0.158% (109.0 … 110.0 μs) 39.8
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 128, 32) Warps (4, 2) OP (16, 16, 16) 5.35 μs ± 2.71% (5.01 … 5.72 μs) 3.29 μs ± 4.06% (2.86 … 3.81 μs) 57.1
WMMA GEMM Float16*Float16+Float32=Float32 (128×128) · (128×128) (TT) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 16.6 μs ± 1.03% (16.2 … 16.9 μs) 2.17 μs ± 7.99% (1.91 … 2.38 μs) 11.8
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 128, 32) Warps (2, 4) OP (16, 16, 16) 5.58 μs ± 2.86% (5.25 … 5.72 μs) 3.22 μs ± 4.28% (2.86 … 3.34 μs) 54.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (256, 64, 16) Warps (2, 4) OP (16, 16, 16) 15.2 μs ± 1.11% (15.0 … 15.5 μs) 3.22 μs ± 4.31% (2.86 … 3.34 μs) 19.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (128, 64, 64) Warps (2, 4) OP (16, 16, 16) 5.68 μs ± 2.68% (5.48 … 5.96 μs) 3.38 μs ± 3.32% (3.1 … 3.81 μs) 56.5
WMMA GEMM Float16*Float16+Float32=Float32 (128×128) · (128×128) (NN) Block (128, 128, 64) Warps (4, 2) OP (8, 32, 16) 17.9 μs ± 0.933% (17.6 … 18.4 μs) 1.97 μs ± 8.62% (1.67 … 2.15 μs) 9.46
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (128, 128, 16) Warps (4, 2) OP (16, 16, 16) 12.1 μs ± 1.44% (11.9 … 12.4 μs) 3.49 μs ± 3.82% (3.1 … 3.81 μs) 26.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 64, 64) Warps (2, 4) OP (16, 16, 16) 3.8 μs ± 2.91% (3.58 … 4.05 μs) 3.47 μs ± 3.99% (3.1 … 3.81 μs) 86.7
WMMA GEMM Float16*Float16+Float32=Float32 (128×256) · (256×128) (NT) Block (128, 128, 64) Warps (4, 2) OP (8, 32, 16) 30.1 μs ± 0.728% (29.6 … 30.8 μs) 2.33 μs ± 6.33% (2.15 … 2.62 μs) 7.26
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 256, 64) Warps (1, 4) OP (16, 16, 16) 17.9 μs ± 0.674% (17.6 … 18.4 μs) 3.38 μs ± 3.14% (3.1 … 3.81 μs) 17.6
WMMA GEMM Float16*Float16+Float32=Float32 (2048×2048) · (2048×2048) (NN) Block (128, 128, 64) Warps (4, 2) OP (32, 8, 16) 345.0 μs ± 0.434% (338.0 … 349.0 μs) 289.0 μs ± 0.135% (289.0 … 291.0 μs) 85.3
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 256, 32) Warps (2, 2) OP (16, 16, 16) 20.7 μs ± 0.597% (20.3 … 21.2 μs) 3.38 μs ± 3.71% (3.1 … 3.81 μs) 15.3
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 128, 32) Warps (1, 4) OP (16, 16, 16) 5.46 μs ± 3.13% (5.25 … 5.72 μs) 3.38 μs ± 3.21% (3.1 … 3.81 μs) 59.1
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 128, 16) Warps (4, 4) OP (16, 16, 16) 7.66 μs ± 1.57% (7.39 … 8.11 μs) 3.34 μs ± 3.65% (2.86 … 3.81 μs) 38.7
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 256, 16) Warps (4, 4) OP (16, 16, 16) 12.6 μs ± 1.35% (12.2 … 12.9 μs) 3.47 μs ± 3.91% (3.1 … 3.81 μs) 25.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 64, 32) Warps (2, 2) OP (16, 16, 16) 3.92 μs ± 3.69% (3.81 … 4.29 μs) 3.38 μs ± 3.78% (2.86 … 3.81 μs) 75.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 128, 64) Warps (2, 2) OP (16, 16, 16) 31.6 μs ± 0.641% (31.2 … 32.4 μs) 3.3 μs ± 2.94% (2.86 … 3.58 μs) 9.16
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (128, 128, 64) Warps (4, 4) OP (16, 16, 16) 10.8 μs ± 0.848% (10.5 … 11.2 μs) 3.38 μs ± 3.98% (2.86 … 3.81 μs) 27.3
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 128, 32) Warps (4, 4) OP (16, 16, 16) 6.33 μs ± 2.39% (5.96 … 6.68 μs) 3.38 μs ± 3.22% (3.1 … 3.81 μs) 52.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (128, 64, 64) Warps (2, 4) OP (16, 16, 16) 6.07 μs ± 2.53% (5.72 … 6.2 μs) 3.31 μs ± 2.9% (2.86 … 3.58 μs) 50.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (256, 64, 64) Warps (4, 2) OP (16, 16, 16) 47.3 μs ± 0.613% (46.5 … 48.2 μs) 3.29 μs ± 3.94% (2.86 … 3.81 μs) 6.15
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 128, 64) Warps (4, 2) OP (16, 16, 16) 5.65 μs ± 2.25% (5.25 … 5.96 μs) 3.22 μs ± 4.36% (2.86 … 3.34 μs) 54.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (128, 64, 64) Warps (1, 4) OP (16, 16, 16) 8.25 μs ± 1.93% (8.11 … 8.58 μs) 3.38 μs ± 3.37% (3.1 … 3.81 μs) 38.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 256, 32) Warps (2, 2) OP (16, 16, 16) 20.8 μs ± 0.594% (20.3 … 21.2 μs) 3.29 μs ± 3.74% (2.86 … 3.58 μs) 14.1
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (256, 64, 32) Warps (2, 2) OP (16, 16, 16) 23.1 μs ± 0.688% (22.6 … 23.4 μs) 3.38 μs ± 3.9% (2.86 … 3.81 μs) 12.6
WMMA GEMM Float16*Float16+Float16=Float16 (256×256) · (256×256) (TT) Block (128, 128, 64) Warps (4, 2) OP (32, 8, 16) 7.7 μs ± 2.2% (7.39 … 8.11 μs) 3.3 μs ± 3.1% (2.86 … 3.81 μs) 38.7
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 256, 64) Warps (2, 4) OP (16, 16, 16) 50.2 μs ± 0.542% (49.6 … 51.3 μs) 3.48 μs ± 3.87% (3.1 … 4.05 μs) 6.25
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 128, 16) Warps (2, 2) OP (16, 16, 16) 8.54 μs ± 2.04% (8.34 … 8.82 μs) 3.56 μs ± 3.35% (3.1 … 4.05 μs) 37.1
FPU GEMM Float64*Float64=Float64 (2048×2048) · (2048×2048) (NT) OP (8, 16, 2), base shape (4, 8, 1) 59.7 ms ± 0.169% (59.6 … 60.0 ms) 41.0 ms ± 0.0135% (41.0 … 41.0 ms) 68.8
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (128, 128, 64) Warps (4, 4) OP (16, 16, 16) 10.8 μs ± 1.14% (10.7 … 11.2 μs) 3.47 μs ± 3.81% (3.1 … 3.81 μs) 28.9
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (128, 64, 32) Warps (4, 4) OP (16, 16, 16) 6.3 μs ± 2.44% (5.96 … 6.44 μs) 3.47 μs ± 3.98% (3.1 … 3.81 μs) 52.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (256, 64, 32) Warps (4, 2) OP (16, 16, 16) 12.1 μs ± 1.16% (11.7 … 12.4 μs) 3.37 μs ± 3.85% (2.86 … 3.81 μs) 24.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (128, 128, 64) Warps (4, 2) OP (32, 8, 16) 10.7 μs ± 1.07% (10.5 … 11.0 μs) 3.44 μs ± 4.28% (2.86 … 3.81 μs) 27.3
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 64, 16) Warps (4, 1) OP (16, 16, 16) 5.6 μs ± 2.52% (5.25 … 5.72 μs) 3.22 μs ± 4.32% (2.86 … 3.34 μs) 54.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 256, 16) Warps (4, 4) OP (16, 16, 16) 12.9 μs ± 1.32% (12.4 … 13.4 μs) 3.5 μs ± 3.72% (3.1 … 3.81 μs) 25.0
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NN) OP (4, 8, 1), base shape (4, 8, 1) 55.8 μs ± 0.578% (54.8 … 57.2 μs) 3.62 μs ± 2.79% (3.58 … 4.05 μs) 6.52
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (256, 64, 32) Warps (1, 4) OP (16, 16, 16) 70.9 μs ± 0.443% (70.1 … 72.0 μs) 3.4 μs ± 3.89% (2.86 … 3.81 μs) 4.08
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 128, 32) Warps (4, 4) OP (16, 16, 16) 6.18 μs ± 2.17% (5.72 … 6.44 μs) 3.21 μs ± 4.57% (2.86 … 3.34 μs) 50.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (128, 64, 32) Warps (2, 4) OP (16, 16, 16) 5.81 μs ± 2.81% (5.48 … 5.96 μs) 3.29 μs ± 3.62% (2.86 … 3.58 μs) 52.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (128, 128, 32) Warps (2, 2) OP (16, 16, 16) 10.5 μs ± 1.1% (10.5 … 11.0 μs) 3.38 μs ± 3.96% (2.86 … 3.81 μs) 27.3
WMMA GEMM Float16*Float16+Float32=Float32 (256×128) · (128×256) (TN) Block (128, 128, 64) Warps (4, 2) OP (32, 8, 16) 7.72 μs ± 1.71% (7.63 … 8.11 μs) 2.46 μs ± 6.65% (2.15 … 2.62 μs) 28.1
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 128, 32) Warps (4, 4) OP (16, 16, 16) 6.83 μs ± 1.87% (6.44 … 6.91 μs) 3.3 μs ± 2.97% (2.86 … 3.58 μs) 44.4
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (256, 64, 16) Warps (2, 4) OP (16, 16, 16) 12.3 μs ± 1.24% (12.2 … 12.6 μs) 3.48 μs ± 4.21% (3.1 … 4.05 μs) 25.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 64, 32) Warps (4, 4) OP (16, 16, 16) 5.6 μs ± 2.63% (5.25 … 5.96 μs) 3.47 μs ± 3.84% (3.1 … 3.81 μs) 59.1
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 256, 32) Warps (4, 2) OP (16, 16, 16) 9.74 μs ± 1.13% (9.54 … 10.0 μs) 3.28 μs ± 4.07% (2.86 … 3.58 μs) 30.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (128, 64, 64) Warps (1, 4) OP (16, 16, 16) 8.79 μs ± 1.91% (8.58 … 9.06 μs) 3.39 μs ± 4.06% (2.86 … 3.81 μs) 33.3
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 256, 16) Warps (2, 2) OP (16, 16, 16) 14.1 μs ± 0.763% (13.8 … 14.5 μs) 3.48 μs ± 3.81% (3.1 … 3.81 μs) 22.4
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 256, 64) Warps (4, 2) OP (16, 16, 16) 12.3 μs ± 1.24% (11.9 … 12.6 μs) 3.38 μs ± 3.38% (3.1 … 3.81 μs) 26.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 128, 64) Warps (4, 1) OP (16, 16, 16) 8.52 μs ± 1.88% (8.34 … 8.82 μs) 3.48 μs ± 3.95% (3.1 … 3.81 μs) 37.1
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (128, 128, 64) Warps (4, 4) OP (16, 16, 16) 10.0 μs ± 1.2% (9.78 … 10.3 μs) 3.23 μs ± 4.29% (2.86 … 3.34 μs) 29.3
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 256, 16) Warps (2, 4) OP (16, 16, 16) 13.7 μs ± 0.945% (13.4 … 13.8 μs) 3.55 μs ± 3.62% (3.1 … 4.05 μs) 23.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×128) · (128×256) (NT) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 16.8 μs ± 0.924% (16.5 … 17.2 μs) 2.58 μs ± 5.29% (2.15 … 2.86 μs) 13.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 256, 64) Warps (4, 4) OP (16, 16, 16) 11.0 μs ± 1.03% (10.7 … 11.4 μs) 3.46 μs ± 3.96% (3.1 … 3.81 μs) 28.9
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (128, 128, 16) Warps (2, 4) OP (16, 16, 16) 12.3 μs ± 1.25% (12.2 … 12.6 μs) 3.35 μs ± 3.54% (2.86 … 3.81 μs) 23.5
FPU GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) OP (8, 16, 2), base shape (4, 8, 1) 43.3 μs ± 0.537% (42.7 … 43.9 μs) 3.43 μs ± 3.75% (3.1 … 3.81 μs) 7.26
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (256, 64, 64) Warps (4, 1) OP (16, 16, 16) 28.8 μs ± 0.567% (28.6 … 29.3 μs) 3.47 μs ± 4.09% (3.1 … 3.81 μs) 10.8
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (128, 64, 64) Warps (4, 2) OP (16, 16, 16) 5.62 μs ± 2.26% (5.25 … 5.72 μs) 3.4 μs ± 3.93% (3.1 … 3.81 μs) 59.1
WMMA GEMM Float16*Float16+Float16=Float16 (256×512) · (512×256) (TN) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 51.0 μs ± 0.354% (50.5 … 51.5 μs) 4.58 μs ± 3.91% (4.29 … 5.01 μs) 8.49
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (128, 128, 16) Warps (4, 2) OP (16, 16, 16) 11.4 μs ± 1.51% (11.2 … 11.7 μs) 3.48 μs ± 4.16% (3.1 … 3.81 μs) 27.7
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (128, 128, 32) Warps (2, 2) OP (16, 16, 16) 9.77 μs ± 1.34% (9.54 … 10.0 μs) 3.22 μs ± 4.38% (2.86 … 3.34 μs) 30.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 256, 32) Warps (2, 4) OP (16, 16, 16) 10.5 μs ± 0.639% (10.0 … 11.0 μs) 3.37 μs ± 4.04% (2.86 … 3.81 μs) 28.6
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 128, 64) Warps (4, 2) OP (16, 16, 16) 5.66 μs ± 2.53% (5.48 … 5.96 μs) 3.29 μs ± 3.76% (2.86 … 3.58 μs) 52.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (128, 128, 64) Warps (4, 1) OP (16, 16, 16) 19.7 μs ± 0.89% (19.3 … 20.3 μs) 3.46 μs ± 4.1% (3.1 … 3.81 μs) 16.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (128, 64, 16) Warps (4, 2) OP (16, 16, 16) 7.4 μs ± 1.68% (7.15 … 7.87 μs) 3.55 μs ± 3.45% (3.1 … 4.05 μs) 43.3
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (256, 64, 16) Warps (2, 2) OP (16, 16, 16) 15.7 μs ± 1.1% (15.5 … 16.0 μs) 3.21 μs ± 4.33% (2.86 … 3.34 μs) 18.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (256, 64, 16) Warps (4, 4) OP (16, 16, 16) 12.3 μs ± 1.31% (11.9 … 12.6 μs) 3.21 μs ± 4.49% (2.86 … 3.34 μs) 24.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 64, 64) Warps (1, 4) OP (16, 16, 16) 3.8 μs ± 3.03% (3.58 … 4.05 μs) 3.48 μs ± 3.92% (3.1 … 3.81 μs) 86.7
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 256, 16) Warps (4, 4) OP (16, 16, 16) 11.7 μs ± 1.4% (11.4 … 12.2 μs) 3.55 μs ± 3.35% (3.1 … 4.05 μs) 27.1
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (256, 64, 64) Warps (2, 4) OP (16, 16, 16) 13.3 μs ± 1.08% (12.9 … 13.6 μs) 3.22 μs ± 4.37% (2.86 … 3.34 μs) 22.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (128, 128, 64) Warps (4, 2) OP (8, 32, 16) 30.6 μs ± 0.588% (30.0 … 31.2 μs) 3.27 μs ± 3.72% (2.86 … 3.58 μs) 9.52
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (256, 64, 32) Warps (4, 2) OP (16, 16, 16) 9.93 μs ± 1.34% (9.54 … 10.0 μs) 3.22 μs ± 4.37% (2.86 … 3.34 μs) 30.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (128, 128, 64) Warps (1, 4) OP (16, 16, 16) 19.3 μs ± 0.97% (18.8 … 19.8 μs) 3.47 μs ± 4.02% (3.1 … 3.81 μs) 16.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (128, 64, 16) Warps (1, 4) OP (16, 16, 16) 8.35 μs ± 2.09% (8.11 … 8.58 μs) 3.49 μs ± 3.78% (3.1 … 3.81 μs) 38.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 256, 32) Warps (2, 2) OP (16, 16, 16) 19.1 μs ± 0.922% (18.8 … 19.6 μs) 3.38 μs ± 3.34% (3.1 … 3.81 μs) 16.5
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TT) OP (32, 4, 1), base shape (32, 1, 1) 106.0 μs ± 0.395% (105.0 … 108.0 μs) 3.56 μs ± 2.61% (3.1 … 3.81 μs) 2.95
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 256, 32) Warps (1, 4) OP (16, 16, 16) 19.7 μs ± 0.706% (19.3 … 20.0 μs) 3.39 μs ± 3.87% (3.1 … 4.05 μs) 16.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (256, 64, 64) Warps (2, 2) OP (16, 16, 16) 23.2 μs ± 0.966% (22.4 … 24.1 μs) 3.29 μs ± 3.83% (2.86 … 3.81 μs) 12.8
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (128, 64, 32) Warps (4, 2) OP (16, 16, 16) 5.47 μs ± 3.11% (5.25 … 5.72 μs) 3.22 μs ± 4.42% (2.86 … 3.34 μs) 54.5
WMMA GEMM Float16*Float16+Float32=Float32 (2048×2048) · (2048×2048) (NT) Block (128, 128, 64) Warps (4, 2) OP (32, 8, 16) 346.0 μs ± 0.342% (342.0 … 351.0 μs) 288.0 μs ± 0.0783% (288.0 … 291.0 μs) 84.1
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (128, 64, 32) Warps (1, 4) OP (16, 16, 16) 5.62 μs ± 2.47% (5.25 … 5.96 μs) 3.3 μs ± 3.77% (2.86 … 3.81 μs) 54.5
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TT) OP (4, 8, 1), base shape (4, 8, 1) 55.8 μs ± 1.03% (53.9 … 57.9 μs) 3.56 μs ± 2.47% (3.1 … 3.81 μs) 5.75
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (128, 128, 64) Warps (4, 2) OP (32, 8, 16) 10.8 μs ± 0.841% (10.5 … 11.2 μs) 3.43 μs ± 3.92% (3.1 … 3.81 μs) 29.5
WMMA GEMM Float16*Float16+Float16=Float16 (512×512) · (512×512) (TT) Block (128, 128, 64) Warps (4, 2) OP (8, 32, 16) 55.2 μs ± 0.418% (54.6 … 56.0 μs) 6.51 μs ± 2.28% (6.2 … 6.68 μs) 11.4
WMMA GEMM Float16*Float16+Float32=Float32 (2048×2048) · (2048×2048) (NT) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 3.3 ms ± 0.688% (3.25 … 3.34 ms) 289.0 μs ± 0.17% (288.0 … 292.0 μs) 8.86
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (128, 64, 16) Warps (4, 4) OP (16, 16, 16) 7.47 μs ± 1.77% (7.15 … 7.87 μs) 3.34 μs ± 2.52% (3.1 … 3.81 μs) 43.3
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (256, 64, 16) Warps (2, 2) OP (16, 16, 16) 20.2 μs ± 0.687% (19.8 … 20.5 μs) 3.38 μs ± 3.21% (3.1 … 3.81 μs) 15.7
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 128, 16) Warps (1, 4) OP (16, 16, 16) 8.58 μs ± 1.92% (8.34 … 8.82 μs) 3.56 μs ± 3.63% (3.1 … 4.05 μs) 37.1
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (128, 128, 16) Warps (4, 1) OP (16, 16, 16) 13.2 μs ± 1.2% (12.9 … 13.4 μs) 3.49 μs ± 3.82% (3.1 … 3.81 μs) 24.1
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 128, 32) Warps (1, 4) OP (16, 16, 16) 5.44 μs ± 2.88% (5.25 … 5.72 μs) 3.29 μs ± 3.79% (2.86 … 3.81 μs) 54.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 256, 64) Warps (2, 4) OP (16, 16, 16) 50.3 μs ± 0.497% (49.6 … 51.0 μs) 3.39 μs ± 3.87% (3.1 … 3.81 μs) 6.25
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (256, 64, 64) Warps (4, 4) OP (16, 16, 16) 10.3 μs ± 1.24% (10.0 … 10.7 μs) 3.38 μs ± 3.33% (3.1 … 3.81 μs) 31.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 256, 64) Warps (4, 1) OP (16, 16, 16) 1.53 ms ± 0.175% (1.52 … 1.54 ms) 3.47 μs ± 3.91% (3.1 … 4.05 μs) 0.203
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 29.6 μs ± 0.733% (29.1 … 30.3 μs) 3.28 μs ± 3.98% (2.86 … 3.81 μs) 9.84
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (256, 64, 32) Warps (4, 2) OP (16, 16, 16) 9.96 μs ± 1.09% (9.54 … 10.3 μs) 3.37 μs ± 3.46% (2.86 … 3.81 μs) 30.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 64, 32) Warps (1, 4) OP (16, 16, 16) 3.97 μs ± 3.91% (3.81 … 4.29 μs) 3.39 μs ± 3.71% (3.1 … 3.81 μs) 81.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (128, 64, 64) Warps (1, 4) OP (16, 16, 16) 8.21 μs ± 1.67% (8.11 … 8.58 μs) 3.21 μs ± 4.68% (2.86 … 3.34 μs) 35.3
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 256, 32) Warps (4, 1) OP (16, 16, 16) 68.0 μs ± 0.446% (67.0 … 68.9 μs) 3.3 μs ± 3.74% (2.86 … 3.81 μs) 4.27
WMMA GEMM Float16*Float16+Float32=Float32 (128×256) · (256×128) (NN) Block (128, 128, 64) Warps (4, 2) OP (8, 32, 16) 30.0 μs ± 0.634% (29.6 … 30.5 μs) 2.69 μs ± 5.94% (2.38 … 2.86 μs) 8.06
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 64, 32) Warps (4, 1) OP (16, 16, 16) 3.62 μs ± 2.84% (3.58 … 4.05 μs) 3.29 μs ± 4.23% (2.86 … 3.81 μs) 80.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 256, 16) Warps (1, 4) OP (16, 16, 16) 12.9 μs ± 1.28% (12.6 … 13.1 μs) 3.22 μs ± 4.34% (2.86 … 3.34 μs) 22.6
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 64, 64) Warps (4, 1) OP (16, 16, 16) 3.82 μs ± 2.8% (3.58 … 4.05 μs) 3.47 μs ± 4.11% (3.1 … 4.05 μs) 86.7
WMMA GEMM Float16*Float16+Float16=Float16 (512×512) · (512×512) (TT) Block (128, 128, 64) Warps (4, 2) OP (32, 8, 16) 11.3 μs ± 1.47% (11.0 … 11.7 μs) 6.5 μs ± 2.23% (6.2 … 6.68 μs) 56.5
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TN) OP (4, 16, 1), base shape (4, 8, 1) 44.7 μs ± 0.621% (43.9 … 45.8 μs) 3.81 μs ± 3.1% (3.58 … 4.05 μs) 8.15
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 128, 64) Warps (4, 4) OP (16, 16, 16) 5.86 μs ± 2.38% (5.48 … 6.2 μs) 3.39 μs ± 3.77% (3.1 … 3.81 μs) 56.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 64, 32) Warps (1, 4) OP (16, 16, 16) 3.66 μs ± 3.53% (3.58 … 4.05 μs) 3.39 μs ± 3.36% (3.1 … 3.81 μs) 86.7
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 64, 32) Warps (2, 4) OP (16, 16, 16) 3.62 μs ± 2.95% (3.34 … 4.05 μs) 3.38 μs ± 4.13% (2.86 … 3.81 μs) 85.7
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 64, 64) Warps (2, 2) OP (16, 16, 16) 3.52 μs ± 3.23% (3.1 … 3.81 μs) 3.39 μs ± 4.08% (2.86 … 3.81 μs) 92.3
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 256, 64) Warps (1, 4) OP (16, 16, 16) 19.2 μs ± 0.823% (18.8 … 19.6 μs) 3.31 μs ± 2.92% (2.86 … 3.58 μs) 15.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 64, 16) Warps (4, 1) OP (16, 16, 16) 5.62 μs ± 2.27% (5.25 … 5.72 μs) 3.56 μs ± 3.25% (3.1 … 4.05 μs) 59.1
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 64, 64) Warps (1, 4) OP (16, 16, 16) 3.45 μs ± 3.86% (3.1 … 3.81 μs) 3.38 μs ± 3.03% (3.1 … 3.81 μs) 100.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 128, 32) Warps (1, 4) OP (16, 16, 16) 5.91 μs ± 2.2% (5.72 … 6.2 μs) 3.38 μs ± 3.79% (2.86 … 3.81 μs) 50.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 128, 16) Warps (1, 4) OP (16, 16, 16) 7.42 μs ± 1.42% (7.15 … 7.87 μs) 3.35 μs ± 2.62% (2.86 … 3.81 μs) 40.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (128, 128, 16) Warps (4, 1) OP (16, 16, 16) 15.0 μs ± 1.13% (14.8 … 15.3 μs) 3.39 μs ± 3.36% (3.1 … 3.81 μs) 21.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 64, 64) Warps (2, 2) OP (16, 16, 16) 3.72 μs ± 3.75% (3.58 … 4.05 μs) 3.37 μs ± 3.85% (2.86 … 3.81 μs) 80.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (256, 64, 16) Warps (4, 2) OP (16, 16, 16) 15.2 μs ± 1.11% (15.0 … 15.5 μs) 3.34 μs ± 3.59% (2.86 … 3.81 μs) 19.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 256, 16) Warps (4, 2) OP (16, 16, 16) 14.7 μs ± 1.11% (14.5 … 15.0 μs) 3.22 μs ± 4.4% (2.86 … 3.34 μs) 19.7
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (256, 64, 16) Warps (2, 4) OP (16, 16, 16) 14.0 μs ± 0.727% (13.6 … 14.5 μs) 3.56 μs ± 3.34% (3.1 … 4.05 μs) 22.8
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 256, 32) Warps (4, 4) OP (16, 16, 16) 10.9 μs ± 1.39% (10.7 … 11.2 μs) 3.31 μs ± 2.81% (2.86 … 3.58 μs) 26.7
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (128, 128, 32) Warps (4, 2) OP (16, 16, 16) 9.58 μs ± 1.71% (9.3 … 9.78 μs) 3.23 μs ± 4.34% (2.86 … 3.34 μs) 30.8
WMMA GEMM Float16*Float16+Float16=Float16 (512×256) · (256×512) (NN) Block (128, 128, 64) Warps (4, 2) OP (32, 8, 16) 8.02 μs ± 2.04% (7.63 … 8.34 μs) 4.34 μs ± 3.01% (4.05 … 4.77 μs) 53.1
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NT) OP (16, 16, 1), base shape (4, 8, 1) 40.4 μs ± 0.423% (40.1 … 40.8 μs) 3.72 μs ± 3.84% (3.58 … 4.05 μs) 8.93
WMMA GEMM Float16*Float16+Float16=Float16 (512×256) · (256×512) (TN) Block (128, 128, 64) Warps (4, 2) OP (32, 8, 16) 7.93 μs ± 2.01% (7.63 … 8.34 μs) 4.46 μs ± 3.63% (4.29 … 4.77 μs) 56.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 128, 32) Warps (1, 4) OP (16, 16, 16) 5.96 μs ± 2.66% (5.72 … 6.2 μs) 3.46 μs ± 3.96% (3.1 … 3.81 μs) 54.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 64, 16) Warps (4, 1) OP (16, 16, 16) 6.79 μs ± 1.97% (6.44 … 6.91 μs) 3.38 μs ± 3.39% (3.1 … 3.81 μs) 48.1
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 128, 16) Warps (4, 1) OP (16, 16, 16) 7.83 μs ± 2.13% (7.63 … 8.11 μs) 3.35 μs ± 3.58% (2.86 … 3.81 μs) 37.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 64, 16) Warps (1, 4) OP (16, 16, 16) 5.89 μs ± 2.01% (5.48 … 6.2 μs) 3.34 μs ± 2.57% (2.86 … 3.81 μs) 52.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (128, 128, 32) Warps (1, 4) OP (16, 16, 16) 10.9 μs ± 1.4% (10.7 … 11.2 μs) 3.3 μs ± 3.83% (2.86 … 3.81 μs) 26.7
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (128, 64, 64) Warps (4, 2) OP (16, 16, 16) 5.56 μs ± 2.97% (5.25 … 5.72 μs) 3.22 μs ± 4.21% (2.86 … 3.34 μs) 54.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 64, 32) Warps (2, 4) OP (16, 16, 16) 4.02 μs ± 3.64% (3.81 … 4.29 μs) 3.37 μs ± 3.9% (2.86 … 3.81 μs) 75.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (128, 64, 16) Warps (4, 4) OP (16, 16, 16) 7.94 μs ± 1.47% (7.63 … 8.34 μs) 3.48 μs ± 3.94% (3.1 … 3.81 μs) 40.6
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (128, 128, 64) Warps (2, 4) OP (16, 16, 16) 29.5 μs ± 0.599% (29.1 … 30.0 μs) 3.22 μs ± 4.31% (2.86 … 3.34 μs) 9.84
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 128, 16) Warps (1, 4) OP (16, 16, 16) 7.18 μs ± 1.15% (6.91 … 7.63 μs) 3.3 μs ± 3.92% (2.86 … 3.81 μs) 41.4
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (128, 64, 16) Warps (2, 4) OP (16, 16, 16) 7.31 μs ± 1.99% (7.15 … 7.63 μs) 3.31 μs ± 3.77% (2.86 … 3.81 μs) 40.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (128, 128, 32) Warps (4, 2) OP (16, 16, 16) 9.58 μs ± 1.67% (9.3 … 9.78 μs) 3.4 μs ± 3.91% (2.86 … 3.81 μs) 30.8
WMMA GEMM Float16*Float16+Float16=Float16 (2048×2048) · (2048×2048) (TT) Block (128, 128, 64) Warps (4, 2) OP (8, 32, 16) 2.99 ms ± 0.329% (2.97 … 3.04 ms) 178.0 μs ± 0.5% (176.0 … 183.0 μs) 5.93
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 64, 64) Warps (2, 4) OP (16, 16, 16) 3.53 μs ± 3.2% (3.1 … 3.81 μs) 3.38 μs ± 3.25% (3.1 … 3.81 μs) 100.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 128, 16) Warps (4, 4) OP (16, 16, 16) 6.92 μs ± 1.05% (6.68 … 7.39 μs) 3.22 μs ± 4.27% (2.86 … 3.34 μs) 42.9
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 128, 64) Warps (2, 2) OP (16, 16, 16) 32.4 μs ± 0.817% (31.7 … 33.4 μs) 3.38 μs ± 3.78% (3.1 … 3.81 μs) 9.77
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (128, 64, 64) Warps (4, 2) OP (16, 16, 16) 5.56 μs ± 2.88% (5.25 … 5.72 μs) 3.38 μs ± 3.14% (3.1 … 3.81 μs) 59.1
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (128, 64, 32) Warps (2, 4) OP (16, 16, 16) 5.77 μs ± 2.95% (5.48 … 5.96 μs) 3.22 μs ± 4.42% (2.86 … 3.34 μs) 52.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 128, 64) Warps (4, 1) OP (16, 16, 16) 8.11 μs ± 2.04% (7.87 … 8.58 μs) 3.39 μs ± 3.81% (2.86 … 3.81 μs) 36.4
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 64, 64) Warps (1, 4) OP (16, 16, 16) 3.59 μs ± 1.92% (3.34 … 4.05 μs) 3.28 μs ± 4.22% (2.86 … 3.58 μs) 85.7
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (128, 128, 64) Warps (4, 4) OP (16, 16, 16) 10.7 μs ± 1.03% (10.5 … 11.2 μs) 3.48 μs ± 4.03% (3.1 … 4.05 μs) 29.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (128, 128, 32) Warps (2, 4) OP (16, 16, 16) 10.4 μs ± 1.24% (10.0 … 10.7 μs) 3.47 μs ± 4.15% (3.1 … 4.05 μs) 31.0
WMMA GEMM Float16*Float16+Float32=Float32 (2048×2048) · (2048×2048) (NN) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 3.17 ms ± 0.192% (3.16 … 3.19 ms) 289.0 μs ± 0.164% (289.0 … 292.0 μs) 9.15
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (128, 128, 16) Warps (4, 2) OP (16, 16, 16) 12.2 μs ± 1.43% (11.9 … 12.4 μs) 3.55 μs ± 3.61% (3.1 … 4.05 μs) 26.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 256, 64) Warps (2, 4) OP (16, 16, 16) 43.7 μs ± 0.633% (42.9 … 44.6 μs) 3.38 μs ± 3.2% (3.1 … 3.81 μs) 7.22
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (128, 64, 64) Warps (2, 4) OP (16, 16, 16) 6.14 μs ± 1.77% (5.72 … 6.44 μs) 3.38 μs ± 4.0% (2.86 … 3.81 μs) 50.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 64, 64) Warps (2, 2) OP (16, 16, 16) 3.73 μs ± 3.8% (3.58 … 4.05 μs) 3.47 μs ± 3.91% (3.1 … 3.81 μs) 86.7
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 128, 16) Warps (4, 4) OP (16, 16, 16) 7.72 μs ± 1.75% (7.63 … 8.11 μs) 3.56 μs ± 3.25% (3.1 … 4.05 μs) 40.6
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 64, 16) Warps (4, 2) OP (16, 16, 16) 4.98 μs ± 3.54% (4.77 … 5.25 μs) 3.55 μs ± 3.44% (3.1 … 4.05 μs) 65.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 256, 32) Warps (4, 1) OP (16, 16, 16) 65.7 μs ± 0.359% (65.1 … 66.8 μs) 3.47 μs ± 3.98% (3.1 … 3.81 μs) 4.76
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (128, 64, 64) Warps (2, 2) OP (16, 16, 16) 38.5 μs ± 0.716% (37.4 … 39.3 μs) 3.47 μs ± 3.9% (3.1 … 3.81 μs) 8.28
FPU GEMM Float16*Float16=Float32 (2048×2048) · (2048×2048) (NT) OP (8, 16, 2), base shape (4, 8, 1) 15.6 ms ± 0.105% (15.5 … 15.7 ms) 289.0 μs ± 0.0862% (289.0 … 292.0 μs) 1.86
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 256, 32) Warps (2, 4) OP (16, 16, 16) 10.5 μs ± 0.804% (10.3 … 11.0 μs) 3.3 μs ± 3.14% (2.86 … 3.58 μs) 27.9
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (256, 64, 16) Warps (4, 1) OP (16, 16, 16) 14.2 μs ± 1.04% (14.1 … 14.5 μs) 3.31 μs ± 3.87% (2.86 … 3.81 μs) 20.3
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 64, 16) Warps (4, 4) OP (16, 16, 16) 5.03 μs ± 3.38% (4.77 … 5.25 μs) 3.47 μs ± 3.98% (3.1 … 4.05 μs) 65.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (128, 64, 32) Warps (2, 2) OP (16, 16, 16) 5.45 μs ± 3.18% (5.25 … 5.72 μs) 3.38 μs ± 3.83% (2.86 … 3.81 μs) 54.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 256, 64) Warps (2, 4) OP (16, 16, 16) 50.0 μs ± 0.576% (49.4 … 51.0 μs) 3.3 μs ± 3.71% (2.86 … 3.58 μs) 5.8
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (128, 64, 16) Warps (4, 1) OP (16, 16, 16) 8.04 μs ± 2.09% (7.87 … 8.34 μs) 3.34 μs ± 3.83% (2.86 … 3.81 μs) 36.4
WMMA GEMM Float16*Float16+Float16=Float16 (512×512) · (512×512) (NT) Block (128, 128, 64) Warps (4, 2) OP (8, 32, 16) 54.3 μs ± 0.427% (53.6 … 54.8 μs) 6.41 μs ± 1.78% (5.96 … 6.68 μs) 11.1
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (128, 128, 64) Warps (2, 4) OP (16, 16, 16) 30.2 μs ± 0.665% (29.6 … 31.0 μs) 3.47 μs ± 3.87% (3.1 … 3.81 μs) 10.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (128, 64, 64) Warps (4, 2) OP (16, 16, 16) 6.11 μs ± 2.21% (5.72 … 6.2 μs) 3.47 μs ± 3.85% (3.1 … 3.81 μs) 54.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 64, 16) Warps (1, 4) OP (16, 16, 16) 5.65 μs ± 2.24% (5.48 … 5.96 μs) 3.22 μs ± 4.23% (2.86 … 3.34 μs) 52.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 128, 32) Warps (2, 2) OP (16, 16, 16) 5.87 μs ± 2.16% (5.48 … 6.2 μs) 3.38 μs ± 3.85% (2.86 … 3.81 μs) 52.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 30.0 μs ± 0.608% (29.6 … 30.5 μs) 3.39 μs ± 3.83% (2.86 … 3.81 μs) 9.68
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (256, 64, 64) Warps (2, 2) OP (16, 16, 16) 22.6 μs ± 0.865% (22.2 … 23.1 μs) 3.22 μs ± 4.42% (2.86 … 3.34 μs) 12.9
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 256, 64) Warps (4, 4) OP (16, 16, 16) 11.0 μs ± 1.09% (10.7 … 11.4 μs) 3.4 μs ± 4.1% (2.86 … 3.81 μs) 26.7
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (256, 64, 64) Warps (4, 2) OP (16, 16, 16) 56.2 μs ± 0.449% (55.6 … 57.0 μs) 3.46 μs ± 4.18% (3.1 … 3.81 μs) 5.58
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 128, 32) Warps (2, 4) OP (16, 16, 16) 6.04 μs ± 2.72% (5.72 … 6.2 μs) 3.31 μs ± 2.92% (2.86 … 3.58 μs) 50.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (128, 128, 32) Warps (2, 4) OP (16, 16, 16) 9.57 μs ± 1.67% (9.3 … 9.78 μs) 3.4 μs ± 3.92% (2.86 … 3.81 μs) 30.8
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (128, 128, 16) Warps (4, 4) OP (16, 16, 16) 11.2 μs ± 1.18% (11.0 … 11.4 μs) 3.34 μs ± 2.85% (2.86 … 3.81 μs) 26.1
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (256, 64, 32) Warps (1, 4) OP (16, 16, 16) 71.3 μs ± 0.453% (70.3 … 72.2 μs) 3.21 μs ± 4.56% (2.86 … 3.34 μs) 4.07
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 64, 16) Warps (4, 4) OP (16, 16, 16) 5.24 μs ± 3.28% (5.01 … 5.48 μs) 3.34 μs ± 3.56% (2.86 … 3.81 μs) 57.1
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (256, 64, 64) Warps (4, 2) OP (16, 16, 16) 54.1 μs ± 0.488% (53.4 … 54.8 μs) 3.38 μs ± 3.44% (3.1 … 3.81 μs) 5.8
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (128, 128, 16) Warps (1, 4) OP (16, 16, 16) 13.1 μs ± 1.16% (12.9 … 13.4 μs) 3.34 μs ± 3.45% (2.86 … 3.81 μs) 22.2
FPU GEMM Float16*Float16=Float32 (128×128) · (128×128) (NT) OP (8, 16, 2), base shape (4, 8, 1) 17.4 μs ± 0.894% (16.9 … 17.9 μs) 1.99 μs ± 7.9% (1.67 … 2.15 μs) 9.86
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (128, 64, 64) Warps (4, 2) OP (16, 16, 16) 5.6 μs ± 2.6% (5.25 … 5.72 μs) 3.29 μs ± 3.52% (2.86 … 3.58 μs) 54.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (128, 128, 64) Warps (4, 1) OP (16, 16, 16) 19.3 μs ± 0.986% (18.8 … 19.8 μs) 3.46 μs ± 4.03% (3.1 … 3.81 μs) 16.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (256, 64, 64) Warps (2, 2) OP (16, 16, 16) 27.9 μs ± 0.708% (27.2 … 28.6 μs) 3.38 μs ± 3.89% (2.86 … 3.81 μs) 10.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 128, 64) Warps (4, 4) OP (16, 16, 16) 6.18 μs ± 2.34% (5.72 … 6.44 μs) 3.38 μs ± 3.95% (2.86 … 3.81 μs) 50.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 128, 32) Warps (4, 2) OP (16, 16, 16) 5.92 μs ± 2.39% (5.72 … 6.2 μs) 3.46 μs ± 3.8% (3.1 … 3.81 μs) 54.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 64, 64) Warps (4, 2) OP (16, 16, 16) 3.52 μs ± 3.4% (3.1 … 3.81 μs) 3.28 μs ± 4.0% (2.86 … 3.81 μs) 92.3
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (128, 128, 16) Warps (2, 4) OP (16, 16, 16) 12.4 μs ± 1.39% (12.2 … 12.6 μs) 3.49 μs ± 3.71% (3.1 … 3.81 μs) 25.5
WMMA GEMM Float16*Float16+Float32=Float32 (128×128) · (128×128) (TN) Block (128, 128, 64) Warps (4, 2) OP (32, 8, 16) 7.62 μs ± 1.87% (7.39 … 7.87 μs) 2.08 μs ± 6.42% (1.91 … 2.38 μs) 25.8
WMMA GEMM Float16*Float16+Float16=Float16 (512×512) · (512×512) (TN) Block (128, 128, 64) Warps (4, 2) OP (32, 8, 16) 11.3 μs ± 1.46% (11.0 … 11.7 μs) 6.56 μs ± 2.2% (6.2 … 6.68 μs) 56.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 128, 32) Warps (2, 2) OP (16, 16, 16) 5.9 μs ± 2.03% (5.72 … 6.2 μs) 3.47 μs ± 3.88% (3.1 … 3.81 μs) 54.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 256, 16) Warps (2, 4) OP (16, 16, 16) 12.7 μs ± 1.36% (12.4 … 12.9 μs) 3.38 μs ± 3.29% (3.1 … 3.81 μs) 25.0
FPU GEMM Float64*Float64=Float64 (2048×2048) · (2048×2048) (NN) OP (8, 16, 2), base shape (4, 8, 1) 59.8 ms ± 0.195% (59.7 … 60.1 ms) 41.0 ms ± 0.0136% (41.0 … 41.0 ms) 68.8
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 64, 32) Warps (4, 2) OP (16, 16, 16) 3.56 μs ± 2.39% (3.1 … 3.81 μs) 3.27 μs ± 4.4% (2.86 … 3.58 μs) 92.3
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (256, 64, 32) Warps (1, 4) OP (16, 16, 16) 75.5 μs ± 0.407% (74.9 … 77.0 μs) 3.39 μs ± 3.78% (2.86 … 3.81 μs) 3.82
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 256, 16) Warps (2, 2) OP (16, 16, 16) 15.4 μs ± 0.906% (15.0 … 15.7 μs) 3.34 μs ± 2.64% (2.86 … 3.81 μs) 19.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (128, 64, 64) Warps (4, 4) OP (16, 16, 16) 5.74 μs ± 2.88% (5.48 … 5.96 μs) 3.39 μs ± 4.1% (2.86 … 3.81 μs) 52.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (128, 128, 64) Warps (4, 2) OP (8, 32, 16) 30.2 μs ± 0.73% (29.6 … 31.0 μs) 3.49 μs ± 3.82% (3.1 … 3.81 μs) 10.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 64, 64) Warps (4, 4) OP (16, 16, 16) 3.71 μs ± 3.81% (3.58 … 4.05 μs) 3.38 μs ± 4.03% (2.86 … 3.81 μs) 80.0
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TN) OP (16, 16, 1), base shape (4, 8, 1) 40.4 μs ± 0.383% (40.1 … 40.8 μs) 3.8 μs ± 3.09% (3.58 … 4.05 μs) 8.93
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (128, 64, 16) Warps (4, 2) OP (16, 16, 16) 7.66 μs ± 1.63% (7.39 … 7.87 μs) 3.49 μs ± 3.75% (3.1 … 3.81 μs) 41.9
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NT) OP (4, 32, 1), base shape (1, 32, 1) 14.2 μs ± 0.981% (14.1 … 14.5 μs) 3.73 μs ± 3.83% (3.58 … 4.29 μs) 25.4
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 256, 64) Warps (4, 1) OP (16, 16, 16) 1.36 ms ± 0.185% (1.35 … 1.37 ms) 3.22 μs ± 4.37% (2.86 … 3.34 μs) 0.211
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (128, 128, 64) Warps (2, 2) OP (16, 16, 16) 16.3 μs ± 1.04% (16.0 … 16.7 μs) 3.3 μs ± 3.81% (2.86 … 3.58 μs) 17.9
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 256, 64) Warps (2, 2) OP (16, 16, 16) 21.2 μs ± 0.819% (20.7 … 21.7 μs) 3.22 μs ± 4.35% (2.86 … 3.34 μs) 13.8
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (128, 128, 32) Warps (4, 2) OP (16, 16, 16) 10.4 μs ± 1.27% (10.0 … 10.7 μs) 3.38 μs ± 4.0% (2.86 … 3.81 μs) 28.6
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (256, 64, 64) Warps (2, 2) OP (16, 16, 16) 26.7 μs ± 0.687% (26.2 … 27.2 μs) 3.3 μs ± 3.18% (2.86 … 3.81 μs) 10.9
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (128, 64, 16) Warps (2, 2) OP (16, 16, 16) 7.43 μs ± 1.41% (7.15 … 7.87 μs) 3.36 μs ± 3.93% (2.86 … 3.81 μs) 40.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (256, 64, 32) Warps (4, 1) OP (16, 16, 16) 17.8 μs ± 0.836% (17.6 … 18.1 μs) 3.28 μs ± 4.01% (2.86 … 3.58 μs) 16.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (128, 128, 32) Warps (2, 2) OP (16, 16, 16) 10.6 μs ± 1.21% (10.5 … 11.0 μs) 3.47 μs ± 3.95% (3.1 … 4.05 μs) 29.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (128, 128, 64) Warps (1, 4) OP (16, 16, 16) 19.4 μs ± 0.961% (19.1 … 20.0 μs) 3.47 μs ± 3.83% (3.1 … 3.81 μs) 16.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 64, 16) Warps (2, 2) OP (16, 16, 16) 4.8 μs ± 3.39% (4.53 … 5.01 μs) 3.22 μs ± 4.44% (2.86 … 3.34 μs) 63.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 256, 16) Warps (1, 4) OP (16, 16, 16) 12.2 μs ± 1.4% (11.9 … 12.4 μs) 3.48 μs ± 3.83% (3.1 … 3.81 μs) 26.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 256, 16) Warps (4, 1) OP (16, 16, 16) 15.8 μs ± 1.11% (15.5 … 16.2 μs) 3.37 μs ± 3.37% (3.1 … 3.81 μs) 20.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (128, 64, 16) Warps (1, 4) OP (16, 16, 16) 7.99 μs ± 1.85% (7.87 … 8.34 μs) 3.35 μs ± 2.79% (2.86 … 3.81 μs) 36.4
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 64, 32) Warps (4, 2) OP (16, 16, 16) 3.68 μs ± 3.68% (3.58 … 4.05 μs) 3.22 μs ± 4.5% (2.86 … 3.34 μs) 80.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 256, 64) Warps (4, 1) OP (16, 16, 16) 1.36 ms ± 0.177% (1.36 … 1.37 ms) 3.38 μs ± 3.24% (3.1 … 3.81 μs) 0.228
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 64, 64) Warps (2, 2) OP (16, 16, 16) 3.51 μs ± 3.45% (3.1 … 3.81 μs) 3.28 μs ± 4.12% (2.86 … 3.81 μs) 92.3
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 256, 32) Warps (1, 4) OP (16, 16, 16) 17.1 μs ± 0.767% (16.7 … 17.4 μs) 3.38 μs ± 3.31% (3.1 … 3.81 μs) 18.6
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 64, 16) Warps (2, 4) OP (16, 16, 16) 5.46 μs ± 3.06% (5.25 … 5.72 μs) 3.38 μs ± 3.3% (3.1 … 3.81 μs) 59.1
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (256, 64, 32) Warps (2, 4) OP (16, 16, 16) 10.1 μs ± 1.36% (9.78 … 10.3 μs) 3.4 μs ± 3.84% (2.86 … 3.81 μs) 29.3
WMMA GEMM Float16*Float16+Float32=Float32 (128×256) · (256×128) (TN) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 30.4 μs ± 0.626% (30.0 … 31.0 μs) 2.51 μs ± 5.95% (2.15 … 2.62 μs) 7.14
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (128, 64, 64) Warps (4, 1) OP (16, 16, 16) 33.4 μs ± 0.83% (32.7 … 34.3 μs) 3.47 μs ± 4.05% (3.1 … 4.05 μs) 9.49
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NT) OP (8, 16, 2), base shape (4, 8, 1) 45.5 μs ± 0.524% (44.8 … 46.5 μs) 3.71 μs ± 3.8% (3.58 … 4.05 μs) 7.98
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (128, 128, 16) Warps (4, 4) OP (16, 16, 16) 12.1 μs ± 1.43% (11.9 … 12.4 μs) 3.38 μs ± 3.38% (3.1 … 3.81 μs) 26.0
WMMA GEMM Float16*Float16+Float16=Float16 (2048×2048) · (2048×2048) (TT) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 3.01 ms ± 0.468% (2.97 … 3.06 ms) 178.0 μs ± 0.478% (176.0 … 182.0 μs) 5.94
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (256, 64, 32) Warps (2, 4) OP (16, 16, 16) 10.1 μs ± 1.41% (9.78 … 10.3 μs) 3.38 μs ± 3.35% (3.1 … 3.81 μs) 31.7
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 128, 64) Warps (4, 1) OP (16, 16, 16) 8.37 μs ± 2.01% (8.11 … 8.58 μs) 3.3 μs ± 3.11% (2.86 … 3.81 μs) 35.3
WMMA GEMM Float16*Float16+Float16=Float16 (256×512) · (512×256) (NN) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 51.7 μs ± 0.381% (51.3 … 52.2 μs) 4.79 μs ± 3.76% (4.29 … 5.25 μs) 8.37
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (256, 64, 32) Warps (2, 4) OP (16, 16, 16) 12.3 μs ± 1.27% (11.9 … 12.6 μs) 3.37 μs ± 3.86% (2.86 … 3.81 μs) 24.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 64, 16) Warps (4, 2) OP (16, 16, 16) 4.93 μs ± 3.32% (4.77 … 5.25 μs) 3.34 μs ± 2.42% (3.1 … 3.58 μs) 65.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 128, 64) Warps (2, 2) OP (16, 16, 16) 32.1 μs ± 0.781% (31.5 … 32.9 μs) 3.29 μs ± 3.82% (2.86 … 3.81 μs) 9.09
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (256, 64, 32) Warps (2, 4) OP (16, 16, 16) 12.0 μs ± 1.17% (11.7 … 12.4 μs) 3.48 μs ± 3.96% (3.1 … 4.05 μs) 26.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (128, 64, 16) Warps (2, 4) OP (16, 16, 16) 6.92 μs ± 1.0% (6.44 … 7.39 μs) 3.34 μs ± 2.49% (2.86 … 3.81 μs) 44.4
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (256, 64, 16) Warps (2, 2) OP (16, 16, 16) 17.2 μs ± 0.656% (16.9 … 17.4 μs) 3.49 μs ± 3.8% (3.1 … 3.81 μs) 18.3
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 128, 16) Warps (1, 4) OP (16, 16, 16) 7.52 μs ± 2.01% (7.39 … 7.87 μs) 3.48 μs ± 3.93% (3.1 … 4.05 μs) 41.9
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (128, 64, 32) Warps (4, 1) OP (16, 16, 16) 5.91 μs ± 2.22% (5.72 … 6.2 μs) 3.47 μs ± 4.01% (3.1 … 4.05 μs) 54.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (128, 128, 32) Warps (2, 2) OP (16, 16, 16) 10.6 μs ± 1.29% (10.5 … 11.0 μs) 3.3 μs ± 3.27% (2.86 … 3.81 μs) 27.3
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 64, 32) Warps (4, 2) OP (16, 16, 16) 3.98 μs ± 3.83% (3.81 … 4.29 μs) 3.3 μs ± 2.93% (2.86 … 3.81 μs) 75.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 30.7 μs ± 0.685% (30.0 … 31.2 μs) 3.47 μs ± 3.88% (3.1 … 3.81 μs) 10.3
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (128, 128, 16) Warps (4, 1) OP (16, 16, 16) 13.3 μs ± 0.863% (12.9 … 13.6 μs) 3.34 μs ± 3.39% (2.86 … 3.58 μs) 22.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 64, 16) Warps (1, 4) OP (16, 16, 16) 6.1 μs ± 2.28% (5.72 … 6.2 μs) 3.38 μs ± 3.3% (3.1 … 3.81 μs) 54.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 128, 32) Warps (4, 1) OP (16, 16, 16) 5.41 μs ± 2.62% (5.25 … 5.72 μs) 3.21 μs ± 4.5% (2.86 … 3.34 μs) 54.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (256, 64, 32) Warps (1, 4) OP (16, 16, 16) 77.0 μs ± 0.399% (76.3 … 78.2 μs) 3.47 μs ± 3.95% (3.1 … 3.81 μs) 4.06
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (128, 64, 32) Warps (4, 4) OP (16, 16, 16) 5.75 μs ± 2.98% (5.48 … 5.96 μs) 3.39 μs ± 4.05% (2.86 … 3.81 μs) 52.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 256, 16) Warps (1, 4) OP (16, 16, 16) 13.0 μs ± 1.13% (12.6 … 13.1 μs) 3.35 μs ± 3.59% (2.86 … 3.81 μs) 22.6
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 128, 16) Warps (4, 2) OP (16, 16, 16) 8.66 μs ± 1.84% (8.34 … 9.06 μs) 3.38 μs ± 3.41% (3.1 … 3.81 μs) 37.1
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (128, 64, 32) Warps (2, 2) OP (16, 16, 16) 5.41 μs ± 2.59% (5.25 … 5.72 μs) 3.3 μs ± 3.73% (2.86 … 3.81 μs) 54.5
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TT) OP (4, 16, 1), base shape (4, 8, 1) 48.8 μs ± 0.792% (47.7 … 50.3 μs) 3.56 μs ± 2.55% (3.1 … 3.81 μs) 6.5
WMMA GEMM Float16*Float16+Float32=Float32 (2048×2048) · (2048×2048) (TN) Block (128, 128, 64) Warps (4, 2) OP (8, 32, 16) 3.16 ms ± 0.334% (3.14 … 3.19 ms) 290.0 μs ± 0.195% (288.0 … 294.0 μs) 9.18
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 64, 64) Warps (1, 4) OP (16, 16, 16) 3.81 μs ± 2.9% (3.58 … 4.05 μs) 3.38 μs ± 3.97% (2.86 … 3.81 μs) 80.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (128, 64, 32) Warps (4, 4) OP (16, 16, 16) 6.41 μs ± 1.95% (5.96 … 6.68 μs) 3.29 μs ± 3.86% (2.86 … 3.81 μs) 48.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (256, 64, 64) Warps (2, 2) OP (16, 16, 16) 30.5 μs ± 0.66% (29.8 … 31.0 μs) 3.37 μs ± 3.06% (3.1 … 3.81 μs) 10.4
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 128, 32) Warps (4, 1) OP (16, 16, 16) 5.89 μs ± 1.97% (5.48 … 6.2 μs) 3.3 μs ± 2.94% (2.86 … 3.81 μs) 52.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 256, 16) Warps (2, 4) OP (16, 16, 16) 13.2 μs ± 1.21% (12.9 … 13.4 μs) 3.48 μs ± 4.15% (3.1 … 4.05 μs) 24.1
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 128, 16) Warps (4, 2) OP (16, 16, 16) 7.77 μs ± 2.06% (7.63 … 8.11 μs) 3.34 μs ± 2.57% (2.86 … 3.81 μs) 37.5
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TN) OP (32, 4, 1), base shape (32, 1, 1) 108.0 μs ± 0.494% (107.0 … 111.0 μs) 3.79 μs ± 3.25% (3.58 … 4.05 μs) 3.36
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (128, 128, 16) Warps (4, 4) OP (16, 16, 16) 11.2 μs ± 1.35% (11.0 … 11.4 μs) 3.31 μs ± 3.61% (2.86 … 3.81 μs) 26.1
WMMA GEMM Float16*Float16+Float16=Float16 (256×256) · (256×256) (NT) Block (128, 128, 64) Warps (4, 2) OP (32, 8, 16) 7.76 μs ± 2.18% (7.39 … 8.11 μs) 3.33 μs ± 3.15% (2.86 … 3.81 μs) 38.7
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 256, 64) Warps (4, 4) OP (16, 16, 16) 11.2 μs ± 1.42% (11.0 … 11.4 μs) 3.47 μs ± 4.04% (3.1 … 3.81 μs) 28.3
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 128, 64) Warps (2, 2) OP (16, 16, 16) 36.4 μs ± 0.805% (35.5 … 37.2 μs) 3.47 μs ± 3.92% (3.1 … 3.81 μs) 8.72
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (256, 64, 32) Warps (4, 4) OP (16, 16, 16) 10.1 μs ± 1.41% (9.78 … 10.3 μs) 3.29 μs ± 3.91% (2.86 … 3.81 μs) 29.3
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 64, 32) Warps (1, 4) OP (16, 16, 16) 3.96 μs ± 3.81% (3.81 … 4.29 μs) 3.3 μs ± 2.97% (2.86 … 3.58 μs) 75.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 256, 32) Warps (4, 1) OP (16, 16, 16) 64.6 μs ± 0.427% (63.9 … 65.6 μs) 3.38 μs ± 3.33% (3.1 … 3.81 μs) 4.85
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 64, 16) Warps (4, 2) OP (16, 16, 16) 4.87 μs ± 2.65% (4.53 … 5.25 μs) 3.3 μs ± 4.0% (2.86 … 3.81 μs) 63.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 256, 16) Warps (4, 2) OP (16, 16, 16) 12.8 μs ± 1.13% (12.4 … 13.1 μs) 3.35 μs ± 2.91% (2.86 … 3.81 μs) 23.1
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 128, 64) Warps (4, 2) OP (16, 16, 16) 6.11 μs ± 2.18% (5.72 … 6.2 μs) 3.47 μs ± 3.84% (3.1 … 3.81 μs) 54.2
WMMA GEMM Float16*Float16+Float32=Float32 (128×256) · (256×128) (TN) Block (128, 128, 64) Warps (4, 2) OP (32, 8, 16) 10.6 μs ± 1.29% (10.5 … 11.0 μs) 2.49 μs ± 6.35% (2.15 … 2.62 μs) 20.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 128, 64) Warps (1, 4) OP (16, 16, 16) 38.2 μs ± 0.762% (37.2 … 39.3 μs) 3.48 μs ± 3.91% (3.1 … 4.05 μs) 8.33
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (128, 64, 64) Warps (1, 4) OP (16, 16, 16) 8.79 μs ± 1.93% (8.58 … 9.06 μs) 3.48 μs ± 3.78% (3.1 … 3.81 μs) 36.1
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (128, 128, 32) Warps (1, 4) OP (16, 16, 16) 11.8 μs ± 1.26% (11.4 … 12.2 μs) 3.47 μs ± 3.96% (3.1 … 3.81 μs) 27.1
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (64, 64, 32) Warps (4, 2) OP (16, 16, 16) 3.88 μs ± 3.05% (3.81 … 4.29 μs) 3.47 μs ± 3.83% (3.1 … 3.81 μs) 81.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 64, 64) Warps (4, 4) OP (16, 16, 16) 3.76 μs ± 3.7% (3.58 … 4.05 μs) 3.38 μs ± 3.42% (3.1 … 3.81 μs) 86.7
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (256, 64, 16) Warps (4, 4) OP (16, 16, 16) 13.7 μs ± 1.03% (13.4 … 14.1 μs) 3.35 μs ± 3.7% (2.86 … 3.81 μs) 21.4
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (128, 64, 16) Warps (4, 4) OP (16, 16, 16) 6.82 μs ± 1.93% (6.44 … 6.91 μs) 3.21 μs ± 4.34% (2.86 … 3.34 μs) 44.4
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (128, 128, 32) Warps (4, 4) OP (16, 16, 16) 9.74 μs ± 1.09% (9.54 … 10.0 μs) 3.28 μs ± 4.06% (2.86 … 3.81 μs) 30.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (256, 64, 32) Warps (4, 1) OP (16, 16, 16) 18.5 μs ± 0.784% (18.1 … 18.8 μs) 3.22 μs ± 4.31% (2.86 … 3.34 μs) 15.8
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 64, 64) Warps (1, 4) OP (16, 16, 16) 3.61 μs ± 2.48% (3.34 … 4.05 μs) 3.22 μs ± 4.52% (2.86 … 3.34 μs) 85.7
WMMA GEMM Float16*Float16+Float32=Float32 (256×128) · (128×256) (TN) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 16.5 μs ± 1.01% (16.2 … 16.9 μs) 2.47 μs ± 6.53% (2.15 … 2.62 μs) 13.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 128, 16) Warps (4, 4) OP (16, 16, 16) 7.57 μs ± 2.11% (7.39 … 7.87 μs) 3.48 μs ± 3.89% (3.1 … 3.81 μs) 41.9
WMMA GEMM Float16*Float16+Float32=Float32 (256×128) · (128×256) (TT) Block (128, 128, 64) Warps (4, 2) OP (8, 32, 16) 17.8 μs ± 0.84% (17.4 … 18.1 μs) 2.52 μs ± 5.56% (2.15 … 2.62 μs) 12.3
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (128, 64, 16) Warps (4, 4) OP (16, 16, 16) 7.64 μs ± 1.75% (7.39 … 7.87 μs) 3.38 μs ± 3.51% (3.1 … 3.81 μs) 41.9
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (64, 128, 64) Warps (2, 2) OP (16, 16, 16) 33.6 μs ± 0.824% (32.9 … 34.3 μs) 3.48 μs ± 3.91% (3.1 … 3.81 μs) 9.42
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 64, 16) Warps (2, 4) OP (16, 16, 16) 5.25 μs ± 3.46% (5.01 … 5.48 μs) 3.2 μs ± 4.53% (2.86 … 3.34 μs) 57.1
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (256, 64, 64) Warps (4, 2) OP (16, 16, 16) 55.9 μs ± 0.419% (55.3 … 56.7 μs) 3.48 μs ± 3.94% (3.1 … 3.81 μs) 5.6
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (64, 128, 16) Warps (4, 4) OP (16, 16, 16) 7.36 μs ± 1.85% (7.15 … 7.63 μs) 3.38 μs ± 3.34% (3.1 … 3.81 μs) 43.3
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 64, 64) Warps (4, 1) OP (16, 16, 16) 3.5 μs ± 3.68% (3.1 … 3.81 μs) 3.28 μs ± 3.96% (2.86 … 3.81 μs) 92.3
WMMA GEMM Float16*Float16+Float16=Float16 (256×256) · (256×256) (NN) Block (128, 128, 64) Warps (4, 2) OP (32, 8, 16) 7.76 μs ± 2.04% (7.63 … 8.11 μs) 3.37 μs ± 3.09% (2.86 … 3.81 μs) 37.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (128, 64, 64) Warps (4, 2) OP (16, 16, 16) 5.95 μs ± 2.63% (5.72 … 6.2 μs) 3.38 μs ± 3.8% (2.86 … 3.81 μs) 50.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 256, 16) Warps (2, 2) OP (16, 16, 16) 13.2 μs ± 1.2% (12.9 … 13.4 μs) 3.35 μs ± 3.47% (2.86 … 3.81 μs) 22.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (128, 64, 64) Warps (4, 4) OP (16, 16, 16) 5.67 μs ± 2.48% (5.48 … 5.96 μs) 3.22 μs ± 4.5% (2.86 … 3.34 μs) 52.2
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (128, 128, 16) Warps (4, 1) OP (16, 16, 16) 12.8 μs ± 0.95% (12.6 … 13.1 μs) 3.55 μs ± 3.49% (3.1 … 4.05 μs) 24.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (128, 64, 16) Warps (4, 4) OP (16, 16, 16) 6.72 μs ± 1.96% (6.44 … 6.91 μs) 3.32 μs ± 3.93% (2.86 … 3.58 μs) 44.4
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (128, 64, 16) Warps (2, 4) OP (16, 16, 16) 8.59 μs ± 2.04% (8.34 … 8.82 μs) 3.56 μs ± 3.22% (3.1 … 4.05 μs) 37.1
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 128, 16) Warps (2, 2) OP (16, 16, 16) 8.14 μs ± 1.92% (7.87 … 8.34 μs) 3.48 μs ± 3.89% (3.1 … 4.05 μs) 39.4
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (256, 64, 64) Warps (1, 4) OP (16, 16, 16) 1.71 ms ± 0.15% (1.7 … 1.72 ms) 3.22 μs ± 4.4% (2.86 … 3.34 μs) 0.168
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (256, 64, 32) Warps (2, 2) OP (16, 16, 16) 26.8 μs ± 0.55% (26.5 … 27.2 μs) 3.47 μs ± 3.98% (3.1 … 3.81 μs) 11.7
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (64, 128, 32) Warps (4, 1) OP (16, 16, 16) 5.41 μs ± 2.79% (5.25 … 5.72 μs) 3.39 μs ± 3.86% (2.86 … 3.81 μs) 54.5
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (128, 128, 32) Warps (2, 4) OP (16, 16, 16) 10.4 μs ± 1.17% (10.0 … 10.7 μs) 3.38 μs ± 3.96% (2.86 … 3.81 μs) 28.6
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (256, 64, 16) Warps (1, 4) OP (16, 16, 16) 17.2 μs ± 0.686% (16.7 … 17.6 μs) 3.34 μs ± 2.87% (2.86 … 3.81 μs) 17.1
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 64, 16) Warps (4, 1) OP (16, 16, 16) 5.15 μs ± 2.73% (5.01 … 5.48 μs) 3.31 μs ± 3.9% (2.86 … 3.81 μs) 57.1
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (NN) OP (8, 8, 1), base shape (4, 8, 1) 46.4 μs ± 0.884% (45.1 … 47.9 μs) 3.62 μs ± 2.88% (3.58 … 4.05 μs) 7.94
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (256, 64, 64) Warps (2, 4) OP (16, 16, 16) 15.2 μs ± 1.14% (14.8 … 15.7 μs) 3.38 μs ± 3.44% (3.1 … 3.81 μs) 21.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (128, 128, 16) Warps (4, 1) OP (16, 16, 16) 12.0 μs ± 1.35% (11.7 … 12.2 μs) 3.48 μs ± 3.96% (3.1 … 4.05 μs) 26.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (128, 64, 32) Warps (2, 2) OP (16, 16, 16) 5.44 μs ± 2.93% (5.25 … 5.72 μs) 3.22 μs ± 4.29% (2.86 … 3.34 μs) 54.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 128, 16) Warps (4, 1) OP (16, 16, 16) 7.53 μs ± 2.06% (7.39 … 7.87 μs) 3.31 μs ± 3.75% (2.86 … 3.81 μs) 38.7
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (64, 256, 16) Warps (4, 4) OP (16, 16, 16) 11.8 μs ± 1.41% (11.4 … 12.2 μs) 3.34 μs ± 3.79% (2.86 … 3.81 μs) 25.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TN) Block (128, 64, 32) Warps (2, 2) OP (16, 16, 16) 5.93 μs ± 2.46% (5.72 … 6.2 μs) 3.3 μs ± 2.96% (2.86 … 3.58 μs) 50.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (128, 128, 32) Warps (2, 2) OP (16, 16, 16) 10.7 μs ± 1.3% (10.5 … 11.0 μs) 3.47 μs ± 3.79% (3.1 … 3.81 μs) 29.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (128, 64, 64) Warps (2, 4) OP (16, 16, 16) 5.68 μs ± 2.68% (5.48 … 5.96 μs) 3.29 μs ± 3.87% (2.86 … 3.81 μs) 52.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 256, 64) Warps (4, 4) OP (16, 16, 16) 10.4 μs ± 1.04% (10.0 … 10.7 μs) 3.3 μs ± 3.83% (2.86 … 3.81 μs) 28.6
FPU GEMM Float32*Float32=Float32 (2048×2048) · (2048×2048) (NT) OP (8, 16, 2), base shape (4, 8, 1) 30.0 ms ± 0.0246% (30.0 … 30.0 ms) 776.0 μs ± 0.244% (759.0 … 779.0 μs) 2.53
FPU GEMM Float64*Float64=Float64 (256×256) · (256×256) (TT) OP (8, 16, 2), base shape (4, 8, 1) 275.0 μs ± 0.0864% (275.0 … 276.0 μs) 109.0 μs ± 0.16% (109.0 … 110.0 μs) 39.7
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 256, 32) Warps (4, 4) OP (16, 16, 16) 10.2 μs ± 1.14% (9.78 … 10.5 μs) 3.22 μs ± 4.34% (2.86 … 3.34 μs) 29.3
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (128, 128, 16) Warps (4, 4) OP (16, 16, 16) 12.1 μs ± 1.31% (11.9 … 12.4 μs) 3.35 μs ± 3.37% (2.86 … 3.81 μs) 24.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (128, 64, 64) Warps (2, 2) OP (16, 16, 16) 38.0 μs ± 0.705% (37.2 … 38.9 μs) 3.38 μs ± 3.26% (3.1 … 3.81 μs) 8.33
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (64, 128, 32) Warps (2, 2) OP (16, 16, 16) 5.44 μs ± 3.01% (5.25 … 5.72 μs) 3.37 μs ± 3.19% (3.1 … 3.81 μs) 59.1
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (256, 64, 32) Warps (2, 2) OP (16, 16, 16) 23.5 μs ± 0.682% (23.1 … 23.8 μs) 3.38 μs ± 3.31% (3.1 … 3.81 μs) 13.4
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (128, 64, 64) Warps (4, 1) OP (16, 16, 16) 35.3 μs ± 0.697% (34.8 … 36.0 μs) 3.29 μs ± 4.03% (2.86 … 3.58 μs) 8.22
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (256, 64, 64) Warps (2, 4) OP (16, 16, 16) 15.5 μs ± 1.11% (15.3 … 15.7 μs) 3.38 μs ± 3.88% (2.86 … 3.81 μs) 18.8
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (128, 128, 64) Warps (2, 2) OP (16, 16, 16) 17.1 μs ± 0.789% (16.7 … 17.4 μs) 3.47 μs ± 3.81% (3.1 … 3.81 μs) 18.6
FPU GEMM Float32*Float32=Float32 (128×128) · (128×128) (TN) OP (8, 8, 1), base shape (4, 8, 1) 44.4 μs ± 0.867% (43.2 … 46.0 μs) 3.8 μs ± 2.93% (3.58 … 4.05 μs) 8.29
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (128, 64, 32) Warps (2, 4) OP (16, 16, 16) 5.7 μs ± 2.85% (5.48 … 5.96 μs) 3.38 μs ± 3.5% (3.1 … 3.81 μs) 56.5
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (64, 256, 16) Warps (4, 1) OP (16, 16, 16) 14.8 μs ± 0.987% (14.5 … 15.3 μs) 3.22 μs ± 4.29% (2.86 … 3.34 μs) 19.7
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NT) Block (128, 128, 32) Warps (1, 4) OP (16, 16, 16) 12.8 μs ± 1.15% (12.4 … 13.1 μs) 3.39 μs ± 4.05% (2.86 … 3.81 μs) 23.1
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (TT) Block (256, 64, 32) Warps (4, 4) OP (16, 16, 16) 11.7 μs ± 1.45% (11.4 … 11.9 μs) 3.37 μs ± 3.77% (2.86 … 3.81 μs) 25.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NT) Block (128, 128, 16) Warps (2, 2) OP (16, 16, 16) 12.7 μs ± 1.35% (12.4 … 12.9 μs) 3.56 μs ± 3.48% (3.1 … 4.05 μs) 25.0
WMMA GEMM Float16*Float16+Float32=Float32 (256×256) · (256×256) (NN) Block (128, 64, 32) Warps (2, 4) OP (16, 16, 16) 6.17 μs ± 2.0% (5.96 … 6.44 μs) 3.47 μs ± 3.82% (3.1 … 3.81 μs) 52.0
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (128, 64, 64) Warps (4, 4) OP (16, 16, 16) 5.68 μs ± 2.75% (5.48 … 5.96 μs) 3.29 μs ± 4.02% (2.86 … 3.58 μs) 52.2
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TN) Block (128, 128, 64) Warps (4, 2) OP (16, 16, 16) 30.2 μs ± 0.69% (29.8 … 31.0 μs) 3.22 μs ± 4.37% (2.86 … 3.34 μs) 9.6
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (TT) Block (64, 128, 16) Warps (2, 2) OP (16, 16, 16) 7.16 μs ± 0.948% (6.91 … 7.39 μs) 3.32 μs ± 3.98% (2.86 … 3.81 μs) 41.4
WMMA GEMM Float16*Float16=Float32 (256×256) · (256×256) (NN) Block (256, 64, 32) Warps (4, 1) OP (16, 16, 16) 20.4 μs ± 0.747% (20.0 … 20.5 μs) 3.38 μs ± 3.22% (3.1 … 3.81 μs) 15.5

@codecov Codecov
Copy link

codecov bot commented Feb 1, 2024 •
edited
Loading

Codecov Report

All modified and coverable lines are covered by tests ✅

Comparison is base (ef7f0a1) 32.04% compared to head (6174240) 32.04%.
Report is 1 commits behind head on master.

❗ Current head 6174240 differs from pull request most recent head 158ce73. Consider uploading reports for the commit 158ce73 to get more accurate results

Additional details and impacted files 
@@           Coverage Diff           @@
##           master     #190   +/-   ##
=======================================
  Coverage   32.04%   32.04%           
=======================================
  Files          11       11           
  Lines         958      958           
=======================================
  Hits          307      307           
  Misses        651      651

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

@maleadt maleadt merged commit c9aba86 into master Feb 14, 2024
3 of 5 checks passed
@maleadt maleadt deleted the tb/improve_tuning branch February 14, 2024 15:30
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers
No reviews
Assignees
No one assigned
Labels
None yet
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
1 participant
@maleadt