TE
TechEcho
Home24h TopNewestBestAskShowJobs
GitHubTwitter
Home

TechEcho

A tech news platform built with Next.js, providing global tech news and discussions.

GitHubTwitter

Home

HomeNewestBestAskShowJobs

Resources

HackerNews APIOriginal HackerNewsNext.js

© 2025 TechEcho. All rights reserved.

Cost Comparison of Deep Learning Hardware: Google TPUv2 vs. Nvidia Tesla V100

36 pointsby pulover 6 years ago

5 comments

zakover 6 years ago
Here is a larger-scale comparison of Cloud TPU and Google Cloud GPU performance and cost (focused on Cloud TPU Pods): <a href="https:&#x2F;&#x2F;cloud.google.com&#x2F;blog&#x2F;products&#x2F;ai-machine-learning&#x2F;now-you-can-train-ml-models-faster-and-lower-cost-cloud-tpu-pods" rel="nofollow">https:&#x2F;&#x2F;cloud.google.com&#x2F;blog&#x2F;products&#x2F;ai-machine-learning&#x2F;n...</a><p>All the code used in that comparison is open source, and there is a detailed methodology page with instructions that you can follow if you want to reproduce the results: <a href="https:&#x2F;&#x2F;github.com&#x2F;tensorflow&#x2F;tpu&#x2F;blob&#x2F;master&#x2F;benchmarks&#x2F;ResNet-50_v1.5_Performance_Comparison_TensorFlow_1.12_GCP.md" rel="nofollow">https:&#x2F;&#x2F;github.com&#x2F;tensorflow&#x2F;tpu&#x2F;blob&#x2F;master&#x2F;benchmarks&#x2F;Res...</a><p>Also, Cloud TPUs are available to everyone for free via Colab. Here is a sample Colab that shows how to train a Keras model on the Fashion MNIST dataset using the Adam optimizer: <a href="https:&#x2F;&#x2F;colab.research.google.com&#x2F;github&#x2F;tensorflow&#x2F;tpu&#x2F;blob&#x2F;master&#x2F;tools&#x2F;colab&#x2F;fashion_mnist.ipynb" rel="nofollow">https:&#x2F;&#x2F;colab.research.google.com&#x2F;github&#x2F;tensorflow&#x2F;tpu&#x2F;blob...</a><p>(I work on Cloud TPUs)
评论 #18964131 未加载
trishumeover 6 years ago
This doesn&#x27;t seem like a very informative benchmark to me. They don&#x27;t mention how&#x2F;whether they tuned the learning rates and batch sizes to optimize for each different device. Like they mention, they also use a very small network that isn&#x27;t something you need the power of a TPU to train quickly and may scale differently than a large network.<p>They also don&#x27;t post their code so I can&#x27;t check that their problems with ADAM aren&#x27;t due to using L2 regularization, which <a href="https:&#x2F;&#x2F;arxiv.org&#x2F;abs&#x2F;1711.05101" rel="nofollow">https:&#x2F;&#x2F;arxiv.org&#x2F;abs&#x2F;1711.05101</a> shows leads to worse performance than SGD and you should use weight decay instead.
评论 #18961813 未加载
评论 #18961659 未加载
twtwover 6 years ago
Looks like riseml has shut down and taken their comparison post down. I was hoping to compare the results.
评论 #18961039 未加载
deepnotderpover 6 years ago
Hmm interesting, is it possible that the batch size for the tpu is larger? I&#x27;m guessing they might be using some sort of large batches to populate their giant GEMM cores
andrewtbhamover 6 years ago
My understanding is that Teslas have scientific accuracy that is not needed for deep learning... my machine has nvidia 1080s.
评论 #18962018 未加载