Note: All images in this directory, unless specified otherwise, are licensed under CC BY-NC 4.0.
| Figure number | Description | Notes |
|---|---|---|
| 6-1 | GPU starvation, while waiting for CPU to finish preparing the data | |
| 6-2 | Terminal output of nvidia-smi highlighting the GPU utilization | |
| 6-3 | Profiler’s timeline in TensorBoard shows an idle GPU while the CPU is processing as well as CPU idling while the GPU is processing | |
| 6-4 | The NVIDIA DALI pipeline | |
| 6-5 | Effect of varying batch size on time per epoch (seconds) as well as on percentage GPU utilization (Log scales have been used for both X- and Y-axes.) | |
| 6-6 | A graph showing the change in loss as the learning rate is increased | |
| 6-7 | A graph showing the rate of change in loss as the learning rate is increased | |
| 6-8 | A matrix multiplication for A x B operation with one of the multiplications highlighted | |
| 6-9 | The $400,000 NVIDIA DGX-2 deep learning system | |
| 6-10 | Comparing different models for size, accuracy, and operations per second (adapted from “An Analysis of Deep Neural Network Models for Practical Applications” by Alfredo Canziani, Adam Paszke, and Eugenio Culurciello) | Page 2, Figure 2 |
| 6-11 | Comparison of several mobile-friendly models in the paper “MixNet: Mixed Depthwise Convolution Kernels” by Mingxing Tan and Quoc V. Le | Page 9, Figure 7 |
| 6-12 | Quantizing from a 0 to 1 32-bit floating-point range down to an 8-bit integer range for reduced storage space |
