In plain words: A tool builds neural networks of many shapes and sizes and times how fast each trains on Google's TPU, NVIDIA's GPU, and an Intel CPU. No chip wins everywhere: each is fastest for certain kinds of models, so the choice depends on the network.
Abstract
Training deep learning models is compute-intensive and there is an industry-wide trend towards hardware specialization to improve performance. To systematically benchmark deep learning platforms, we introduce ParaDnn, a parameterized benchmark suite for deep learning that generates end-to-end models for fully connected (FC), convolutional (CNN), and recurrent (RNN) neural networks. Along with six real-world models, we benchmark Google's Cloud TPU v2/v3, NVIDIA's V100 GPU, and an Intel Skylake CPU platform. We take a deep dive into TPU architecture, reveal its bottlenecks, and highlight valuable lessons learned for future specialized system design. We also provide a thorough comparison of the platforms and find that each has unique strengths for some types of models. Finally, we quantify the rapid performance improvements that specialized software stacks provide for the TPU and GPU platforms.
Yu Emma Wang, Gu-Yeon Wei, David Brooks
arXiv:1907.10701 · cs.LG, cs.PF, stat.ML · submitted Jul 24, 2019 · updated Oct 22, 2019
abstract · pdf · html
If you're going to write a paper like this to share such results, do yourself a favor and make the major conclusions graphically large!
Every one of Figures 8-12 could take up 1/2 a page and not be inappropriate. It's arxiv - what's the benefit of economizing?