Add a GPU support matrix in documentation

[fire]

Request description

Add a GPU support matrix table so that independent software vendors know what is expected to work.

What component(s) does this issue relate to?

Runtime

Additional context

About gpu compatibility with AMD and IREE?

Was curious about https://www.techpowerup.com/gpu-specs/radeon-rx-7900-xtx.c3941 which has 24GB and https://www.amd.com/en/products/graphics/workstations/radeon-pro/w7900.html which has 48GB of video ram.
and.. uh the steamdeck AMD gpu.

Scott Todd — Today at 13:51

What are you curious about specifically? General support for the hardware?

Jakub Kuderski — Today at 14:07
We support both 7900xtx and w7900.

Steam deck is based on rdan2 and should work, but there may be some sharp edges if nobody tried it before

Ben Vanik — Today at 14:12
I tried it ages ago, no clue if it still works though (both CPU and Vulkan)
last year around this time, apparently :P ⁠vulkan⁠

iFire — Today at 14:14
Will you claim LLMs run on the steamdeck 😄
Like in a table.

Ben Vanik — Today at 14:15
I claim nothing. YMMV :P

iFire — Today at 14:18

So both AMD 7900xtx and w7900 will have IREE organization support. That's great!

Scott Todd — Today at 14:20

We have CI workflows running unit tests and some model tests/benchmarks on w7900 GPUs using both the ROCm/HIP and Vulkan APIs. That's mostly what we had on hand in sufficient supply for automated testing, but we'd like to expand the test device matrix eventually.

Some performance optimization work is shared across all backends or at least across CDNA and RDNA (via the LLVMGPU / ROCm / HIP code generation pipelines).

iFire — Today at 14:22

Both windows and linux right?

Scott Todd — Today at 14:26

Yep. Related to that, we're evaluating HSA as a backend option for AMD hardware that has many nice features for the runtime to use, like on #18349, but that (I think?) is currently Linux-only. Definitely pushing to keep the core project fully supported on all operating systems.

Hmm that reminds me, I have a dual boot machine with some w7900s in it now. I should run through our docs etc. on Windows and check that things work out of the box.

[ScottTodd]

Agreed we can do more here.

Current relevant docs:

• https://iree.dev/#support-matrix (high level, no specifics)
• https://iree.dev/guides/deployment-configurations/gpu-vulkan/#support-matrix
• https://iree.dev/guides/deployment-configurations/gpu-vulkan/#compile-a-program (commonly used architectures)
• https://iree.dev/guides/deployment-configurations/gpu-cuda/#compile-a-program (commonly used architectures)
• https://iree.dev/guides/deployment-configurations/gpu-rocm/#compile-a-program (commonly used architectures)

edit: and https://iree.dev/guides/deployment-configurations/gpu-metal/ (when that has any content :P)

[fire]

There's a bug report about AMD 7900XTX card failure. nod-ai/SHARK-Studio#2165