Tensorflow 1.5 about deep-learning-benchmark

[u39kun]

I agree. Not sure if you've read my article or not, but I've written about Titan V vs 1080 Ti and cost benefit analysis:
https://medium.com/@u39kun/titan-v-vs-1080-ti-head-to-head-battle-of-the-best-desktop-gpus-on-cnns-d55a19866b7c

I was really hopeful about Titan V, but after running the benchmark, I've returned Titan V's to NVIDIA within their 30-day return window for a refund. NVIDIA was very nice about it and gave me a full refund.

[anatolix]

Looks like TF 1.4 was compiled with new cuda but never used CUBLAS_TENSOR_OP_MATH.
Also I am not really sure they still have full support, becuase they mentioned GEMM and not mentioned Convolution in release notes (Tensor cores could do two types of operations GEMM and Convolution)

I'm pretty sure that Tensor Cores were being used

For my perception looks like they not. We got near 2x improvement in speed which could be effect of fp16. Nvidia test says it should be 10x https://devblogs.nvidia.com/programming-tensor-cores-cuda-9/

There is no reason to buy 3000$ card for 2x improvement, on my opinion.

[u39kun]

@anatolix Thanks for letting me know about the TensorFlow 1.5.0 release. Great to hear that TensorFlow 1.5.0 with pre-built CUDA 9 / CUDNN 7 binaries are officially out.

If someone can run benchmarking on the official TF 1.5.0 that would be great!
I'm myself very curious about the new numbers with TensorFlow 1.5.0.
BTW I no longer own the Titan V; I may re-run the numbers on AWS p3.2xlarge (V100) when I get around to it. Or I hope someone beats me to it!

BTW, the version of Tensorflow 1.4.0 that was used for benchmarking is a modified, optimized version put out by NVIDIA built with CUDA 9 and CuDNN 7 support (available via docker pull nvcr.io/nvidia/tensorflow:17.12 though this requires you to have an account with NVIDIA GPU Cloud: https://ngc.nvidia.com.) Seeing the speed up, especially in comparison to PyTorch, I'm pretty sure that Tensor Cores were being used, though perhaps not in certain scenarios.

[anatolix]

Guys reports volta gets additional 20-30% improvement on TF 1.5. But still long way to go to 10x improvement which nvidia promised in their tests.

[u39kun]

@anatolix FYI, the README has been updated with numbers from TensorFlow 1.5.0 thanks to @melgor

[neil-119]

Would've liked to see NVCaffe (nvidia's official fork) benchmarked, as it's supposed to have better support for tensor core ops.