Reference Guide

21 Dell EMC Ready Solutions for AI Deep Learning with NVIDIA | v1.0
(a) Horovod+TensorFlow across 32 GPUs using the ILSVRC2012 dataset with Resnet50
(b) MXNet across 32 GPUs using the ILSVRC2012 dataset with Resnet50
(c) Caffe2 across 32 GPUs for Resnet50 using the ILSVRC2012 dataset
Figure 9: The scaling performance of Deep Learning training on V100-PCIe
3.1.5 Storage Performance
The impact of different storage sub-system options for the Deep Learning Solution was evaluated next.
Experiments were designed to measure the performance of three different convolutional neural networks on
four types of storage systems. The TensorFlow framework was used and the three neural networks used were
AlexNet, ResNet50, and VGG16. The batch size used for benchmarking was 128 for VGG16, and 256 for both
AlexNet and ResNet50. All experiments were running in FP16 mode to stress the I/O operations in neural
network training. The four systems are described below.
Isilon F800