TalkChess.com

I installed CUDA 9.0 and cuDNN 7.12, with the libraries in CUDA lc0 folder. But lc0 crashes:

uciok
go
Found network file: C:/Users/Kai/Downloads/LC0_CUDA/weights.txt
Creating backend [cudnn]...
error CUDNN error: CUDNN_STATUS_ARCH_MISMATCH (C:/my/dev/leela-chess/lc0/src/neural/network_cudnn.cu:658)
====================================

My OS is Windows 8.1, CUDA 9.0 is for Win 8.1, cuDNN 7.12 is either for Win7 or for Win10 (tried both), there seems to be no WIndows 8.1 cuDNN. That indicated folder in the error message doesn't even exist on my PC.

Are you using the cuDNN for CUDA 9.0? The error says cuDNN mismatch, suggesting it is the wrong version.

Albert Silver wrote: ↑Sun May 13, 2018 4:16 pm Are you using the cuDNN for CUDA 9.0? The error says cuDNN mismatch, suggesting it is the wrong version.

Yes, it's cudnn-9.0-windows10-x64-v7.1 cuDNN package, and the win7 one was one also tried.

Laskos wrote: ↑Sun May 13, 2018 5:10 pm

Albert Silver wrote: ↑Sun May 13, 2018 4:16 pm Are you using the cuDNN for CUDA 9.0? The error says cuDNN mismatch, suggesting it is the wrong version.

Yes, it's cudnn-9.0-windows10-x64-v7.1 cuDNN package, and the win7 one was one also tried.

This is an error you typically get when your GPU has not the required compute capability.
If I'm not mistaken CUDA needs 3.0, maybe your GPU is too old.

Joost Buijs wrote: ↑Sun May 13, 2018 5:16 pm

Laskos wrote: ↑Sun May 13, 2018 5:10 pm

Albert Silver wrote: ↑Sun May 13, 2018 4:16 pm Are you using the cuDNN for CUDA 9.0? The error says cuDNN mismatch, suggesting it is the wrong version.

Yes, it's cudnn-9.0-windows10-x64-v7.1 cuDNN package, and the win7 one was one also tried.

This is an error you typically get when your GPU has not the required compute capability.
If I'm not mistaken CUDA needs 3.0, maybe your GPU is too old.

Ah, thanks, my GT 730 in not on the list. I installed everything in preparation for my GTX 1060, which will arrive soon. I hope this problem will disappear.

Laskos wrote: ↑Sun May 13, 2018 5:32 pm

Joost Buijs wrote: ↑Sun May 13, 2018 5:16 pm

Laskos wrote: ↑Sun May 13, 2018 5:10 pm
Yes, it's cudnn-9.0-windows10-x64-v7.1 cuDNN package, and the win7 one was one also tried.

This is an error you typically get when your GPU has not the required compute capability.
If I'm not mistaken CUDA needs 3.0, maybe your GPU is too old.

Ah, thanks, my GT 730 in not on the list. I installed everything in preparation for my GTX 1060, which will arrive soon. I hope this problem will disappear.

You must have DDR3 memory. With DDR3 memory GT 730 has only 2.1 compute capability which is insufficient for CUDA 9.0.
See under https://developer.nvidia.com/cuda-gpus.

Milos wrote: ↑Sun May 13, 2018 5:49 pm

Laskos wrote: ↑Sun May 13, 2018 5:32 pm

Joost Buijs wrote: ↑Sun May 13, 2018 5:16 pm

This is an error you typically get when your GPU has not the required compute capability.
If I'm not mistaken CUDA needs 3.0, maybe your GPU is too old.

Ah, thanks, my GT 730 in not on the list. I installed everything in preparation for my GTX 1060, which will arrive soon. I hope this problem will disappear.

You must have DDR3 memory. With DDR3 memory GT 730 has only 2.1 compute capability which is insufficient for CUDA 9.0.
See under https://developer.nvidia.com/cuda-gpus.

Thanks, that's correct.

Joost Buijs wrote: ↑Sun May 13, 2018 5:16 pm

Laskos wrote: ↑Sun May 13, 2018 5:10 pm

Albert Silver wrote: ↑Sun May 13, 2018 4:16 pm Are you using the cuDNN for CUDA 9.0? The error says cuDNN mismatch, suggesting it is the wrong version.

Yes, it's cudnn-9.0-windows10-x64-v7.1 cuDNN package, and the win7 one was one also tried.

This is an error you typically get when your GPU has not the required compute capability.
If I'm not mistaken CUDA needs 3.0, maybe your GPU is too old.

Just got GTX 1060 6GB GPU, performed some tests. NPS from starting position to depth 25-26 on ID305 with v0.10 GPU master is about 820. Latest CUDA LC0 from here https://crem.xyz/lc0/ is about 4000. People were saying that this build is buggy, and is barely performing above master non-CUDA v0.10 version. I found a huge advantage of CUDA version, almost to the level of its NPS advantage.

1'+1'' games, net ID305
100 games
LC0 GPU CUDA vs LC0 GPU v0.10
+44 -4 =52
+147 Elo points

At this level, the doubling is about 75 Elo points, so the CUDA build performance is about 4 times effective speed-up compared to master GPU non-CUDA build, or close to NPS advantage. I am pretty happy now with my GPU, from 170 NPS on 4 CPU threads I am suddenly now up to 4000 NPS on GPU. Huuuuge!

Laskos wrote: ↑Thu May 17, 2018 6:28 pm
Just got GTX 1060 6GB GPU, performed some tests. NPS from starting position to depth 25-26 on ID305 with v0.10 GPU master is about 820. Latest CUDA LC0 from here https://crem.xyz/lc0/ is about 4000. People were saying that this build is buggy, and is barely performing above master non-CUDA v0.10 version. I found a huge advantage of CUDA version, almost to the level of its NPS advantage.

So are you saying it has been fixed? Is there a document explaining how to change over (I also have a 1060)

Werewolf wrote: ↑Thu May 17, 2018 6:46 pm

Laskos wrote: ↑Thu May 17, 2018 6:28 pm
Just got GTX 1060 6GB GPU, performed some tests. NPS from starting position to depth 25-26 on ID305 with v0.10 GPU master is about 820. Latest CUDA LC0 from here https://crem.xyz/lc0/ is about 4000. People were saying that this build is buggy, and is barely performing above master non-CUDA v0.10 version. I found a huge advantage of CUDA version, almost to the level of its NPS advantage.

So are you saying it has been fixed? Is there a document explaining how to change over (I also have a 1060)

It seems no more buggy than the master itself v0.10 is . So, it seems its NPS are almost fair to compare to master NPS.
You have to decompress the weights file and put in the same folder with CUDA lc0 (renaiming it weights.txt). Then you have to install CUDA 9.0, after that cuDNN 7.1 for CUDA 9.0. After that, take the installed dll libraries and put them into lc0 folder. Should work.