The P3 instances are the first widely and easily accessible machines that use the NVIDIA Tesla V100 GPUs. These GPUs are straight up scary in terms of firepower.
To give an understanding of the speed-up compared to the P2 instances for a research project of mine:
+ P2 (K80) with single GPU: ~95 seconds per epoch
+ P3 (V100) with single GPU: ~20 seconds per epoch
Admittedly this isn't exactly fair for either GPU - the K80 cards are straight up ancient now and the Volta isn't sitting at 100% GPU utilization as it burns through the data too quickly ([CUDA kernel, Python] overhead suddenly become major bottlenecks).
This gives you an indication of what a leap this is if you're using GPUs on AWS however.
Oh, and the V100 comes with 16GB of (faster) RAM compared to the K80's 12GB of RAM, so you win there too.
For anyone using the standard set of frameworks (Tensorflow, Keras, PyTorch, Chainer, MXNet, DyNet, DeepLearning4j, ...) this type of speed-up will likely require you to do nothing - except throw more money at the P3 instance :)
If you really want to get into the black magic of speed-ups, these cards also feature full FP16 support, which means you can double your TFLOPS by dropping to FP16 from FP32. You'll run into a million problems during training due to the lower precision but these aren't insurmountable and may well be worth the pain for the additional speed-up / better RAM usage.
Great write up as usual! Could you elaborate more on the python overhead a bit? We have fp16 support running in dl4j but I don't think we've really done much with volta yet beyond get it working. In practice, (especially when we do multi gpu async back round loading of data) we find gpus being data starved. I would love to compare support for what you're seeing with pytorch.
Honestly, I didn't spend enough time delving in to the Python overhead, especially in terms of the framework. Most of it would be an issue of my own causing however rather than the framework's. The original code I wrote was never written with data loading / saving in mind as the source for speed issues so I avoided what would have been premature optimization at the time.
Some of the slowdowns now just seem silly and aren't even listed in the per epoch timings: PyTorch doesn't have an asynchronous torch.save(). This means that if you save your model after each epoch, and the model save takes a few seconds, you're increasing your per epoch timings 5-10% just by saving the damn thing!
Regarding FP16, PyTorch supports, and there's even a pull request that updates the examples repo with FP16 support for language modeling and ImageNet. It's not likely to be merged as it greatly complicates a codebase that's meant primarily for teaching purposes but it's lovely to look at. I also think many of the FP16 issues will get a general wrapper and they'll become far more agnostic to the end user. For the most part they're all outlined in NVIDIA / Baidu's "Mixed Precision Training" paper. Might be useful for DeepLearning4j to go through the most common heavy throughput use cases and get them running (just as an example of how to work around issues really) if customers were using P100s/V100s?
I'm really interested in exploring the FP16 aspect as the QRNN, especially for single GPU, is sitting at basically 100% utilization, with almost all the time spent on matrix multiplications. FP16 is about the only way to speed it up at that stage. This gets a tad more complicated regardless as the CUDA kernel is not written in FP16 (and is not easy to do so) but even converting FP16->FP32->(QRNN element-wise CUDA kernel)->FP16 ("pseudo" FP16) should still be a crazy speedup. I tested that on the P100 and it took per epoch AWD-QRNN from ~28 seconds to ~18.
Nice comment. In regard to your reference to reducing precision to FP16 for performance gains, you might want to read a recently published paper by Baidu Research and NVIDIA teams on mixed precision training of deep learning models (link to the paper is at the end of the following relevant post): https://www.nextplatform.com/2017/10/11/baidu-sheds-precisio.... Enjoy! :-)
Genuinely curious: Given that Softlayer bare metal server prices start at 700$ per month is there even remote chance of this actually being profitable?
Nope, that would be hugely unprofitable. To give you an idea:
The P100 instances on Softlayer would cost around $2,000/mo, and would generate approximately $170/mo in ETH when fully optimized. One could probably build a DIY rig with the same hashing power for less than 2k total.
showing up for me on https://aws.amazon.com/ec2/pricing/ for each of On-Demand Instances, Reserved Instances, Spot Instances, and Dedicated Hosts pricing lists.
Are you selecting the regions where this is available - US East (N. Virginia), US West (Oregon), EU West (Ireland) and Asia Pacific (Tokyo)?
+ P2 (K80) with single GPU: ~95 seconds per epoch
+ P3 (V100) with single GPU: ~20 seconds per epoch
Admittedly this isn't exactly fair for either GPU - the K80 cards are straight up ancient now and the Volta isn't sitting at 100% GPU utilization as it burns through the data too quickly ([CUDA kernel, Python] overhead suddenly become major bottlenecks). This gives you an indication of what a leap this is if you're using GPUs on AWS however. Oh, and the V100 comes with 16GB of (faster) RAM compared to the K80's 12GB of RAM, so you win there too.
For anyone using the standard set of frameworks (Tensorflow, Keras, PyTorch, Chainer, MXNet, DyNet, DeepLearning4j, ...) this type of speed-up will likely require you to do nothing - except throw more money at the P3 instance :)
If you really want to get into the black magic of speed-ups, these cards also feature full FP16 support, which means you can double your TFLOPS by dropping to FP16 from FP32. You'll run into a million problems during training due to the lower precision but these aren't insurmountable and may well be worth the pain for the additional speed-up / better RAM usage.
- Good overview of Volta's advantages compared to event the recent P100: https://devblogs.nvidia.com/parallelforall/inside-volta/
- Simple table comparing V100 / P100 / K40 / M40: https://www.anandtech.com/show/11367/nvidia-volta-unveiled-g...
- NVIDIA's V100 GPU architecture white paper: http://www.nvidia.com/object/volta-architecture-whitepaper.h...
- The numbers above were using my PyTorch code at https://github.com/salesforce/awd-lstm-lm and the Quasi-Recurrent Neural Network (QRNN) at https://github.com/salesforce/pytorch-qrnn which features a custom CUDA kernel for speed