Nvidia claims 6x performance improvement with cudnn 7.2

[albu]

Hi, @u39kun!

Nvidia claims 6x performance improvement with recent cudnn 7.2 (https://developer.nvidia.com/cudnn) Could you please try it on Titan V?

Thank you!

[zhiyunl]

+1

[u39kun]

I don't have Titan V any more, but if someone could test and post results that would be great!

[u39kun]

I'm getting a 2080 Ti soon (like tomorrow), which has Tensor Cores. I will post results when I get to it.

[sp82]

Nvidia claims incredible performance out of all this tensorcores. I don't see this huge difference comparing 1080TI vs TitanV. Is Nvidia lying? from the benchmarks I feel tensor cores doesn't exist at all, I feel tensor cores is only an api on top of regular cuda cores. What you think?

EDIT: even in this tests: https://www.anandtech.com/show/12673/titan-v-deep-learning-deep-dive/9 sometimes the performance with tensorcores enabled of the TitanV is lower than tensorcores disabled.

[u39kun]

@sp82 I have written about it theoretical performance on Tensor cores vs real world here: https://medium.com/@u39kun/titan-v-vs-1080-ti-head-to-head-battle-of-the-best-desktop-gpus-on-cnns-d55a19866b7c and also https://medium.com/@u39kun/if-you-strictly-perform-tensor-core-operations-only-i-presume-that-you-would-get-the-performance-caaefc70e4e7 Let me know if that helps.

[sp82]

Thanks! Very interesting. I was thinking to get a RTX2060 for the tensorcores but for the training is better to go with a bigger GPU even without tensorcores at this point. Maybe tensorcores can make a good difference in inference thanks the optimizations done by TensorRT? Is TensorRT make use of tensorcores?