The GPUs were tested by BioHPC on DAWNBench, a Stanford deep-learning benchmark that prioritizes both training speed and test accuracy. DAWNBench sets a rigorous goal of 94% image classification accuracy for the CIFAR10 dataset, which itself is composed of 50,000 32x32 pixel training images and 10,000 test images. Deep-learning architectures like ResNet are suitable for this task, and a ResNet-18 model was selected for the benchmark.
Training and testing was done using PyTorch 0.41 on single GPUs (green) and on nodes with two GPUs (blue) when applicable. Lower training times are desired. The results show the newly added Tesla P40 cards are slightly faster than the Tesla P100. The P4 cards train at approximately half the rate of the P100. The most exciting results are the very low training times on the single and double Tesla V100 node, slashing the training times compared to the P100 by a factor of two.