Next-Gen GPUs for LC0

[Werewolf]

Concerning speedup for Lc0:
With Nvidia Ampere on its way pretty soon I've been looking into leaked and (take with a pinch of salt etc etc) early reports suggest actually it's nothing special. Performance per watt looks surprisingly similar to Turing, with the new cards sucking 350W each rather than 250W from the 2080 Ti.

Good in winter...

Also, can someone please clarify if Lc0 on the GPU is using Cuda cores or Tensor cores? Since core count is now known and so is clock speed (pinch of salt again) we can begin estimating speedup.

[smatovic]

Concerning speedup for Lc0:
With Nvidia Ampere on its way pretty soon I've been looking into leaked and (take with a pinch of salt etc etc) early reports suggest actually it's nothing special. Performance per watt looks surprisingly similar to Turing, with the new cards sucking 350W each rather than 250W from the 2080 Ti.

Good in winter...

Also, can someone please clarify if Lc0 on the GPU is using Cuda cores or Tensor cores? Since core count is now known and so is clock speed (pinch of salt again) we can begin estimating speedup.

Milos' comment on Ampere:

http://talkchess.com/forum3/viewtopic.p ... 9&#p846617

--
Srdja

[Werewolf]

Depressing.

[Werewolf]

OK this looks BIG

https://www.anandtech.com/show/16057/nv ... 0-rtx-3090

[Werewolf]

Although it's not stated, surely that means FP16 on CUDA alone will be 71 TFLOPS, no? That's well over 2x performance increase for Lc0..

[Laskos]

Although it's not stated, surely that means FP16 on CUDA alone will be 71 TFLOPS, no? That's well over 2x performance increase for Lc0..

Yes, seems above 2x increase for both 3080 and 3070. In 2018 the best deal for NNs was 2070, better than 2080 or Ti. This year it seems the best deal will be 3080, not 3070 or 3090. But I might end up with 3070 (cheaper and consumes less electricity. I am afraid even about the my source with 3080's 320W).

[Ozymandias]

I don't know, I look at those 300W+ GPUs and I wonder what the utilities bill would look like (we have among the most expensive electricity in the world). I bought 3 1650 SUPERs back in January, and even those I have at 70% power.

[Vinvin]

Official announce of the NVIDIA GeForce RTX Série 3 :

[Werewolf]

I wonder how a head to head would go between the 3090 and 2x2080Ti for Lc0. I guess running off one card would be quite a big advantage

[smatovic]

Hmm, something is fishy with those RTX 30xx FP32 numbers, better wait for the first LC0 benchmarks imho, cos Turing added dedicated INT32 cores to the SM, these are still present in the whitepaper for the A100, but now with RTX 30xx Nvidia seems to count these INT32 cores as FP32 Cuda Cores, therefore the doubling of cores and FP32 throughput, but some tech sites claim it is yet unknown if the doubling of FP32 throughput per SM holds in real world applications....

--
Srdja