KataGo icon indicating copy to clipboard operation
KataGo copied to clipboard

Katago not working with cuda 11.5 and cudnn 8.3.0

Open kiatchun opened this issue 3 years ago • 5 comments

Katago Cuda was working before I upgraded to cuda 11.5 and cudnn 8.3.0 . Is katago incompatible with cuda 11.5 or did I just forget to do something.

kiatchun avatar Nov 28 '21 07:11 kiatchun

Screenshot 2021-11-28 154444 Screenshot 2021-11-28 154629

kiatchun avatar Nov 28 '21 07:11 kiatchun

KataGo has never been tested with CUDA 11.5. As you can see from the release notes (https://github.com/lightvector/KataGo/releases), officially speaking, the windows executables were compiled with only CUDA 11.2.

Please don't rush to upgrade CUDA versions every time a new CUDA version comes out. The performance of different CUDA versions is often extremely similar, so upgrading commonly has very little benefit. However CUDA versions may have incompatibilities, and rarely upgrading will even cause a performance regression or a new bug, so at the same time as there is very little benefit, there are also risks to upgrading.

If you look at the CUDA version history and release notes, you can see every time CUDA releases a new version, a few months later they release "update 1" and "update 2" and so on that fix bugs and performance regressions that were caused by the new release. Generally I would not bother updating CUDA except very very infrequently and only if I need it for the software I'm running. And I would only ever update to a version that I know has been out for many many months and is highly tested and stable, because the very latest version may more commonly have problems.

lightvector avatar Nov 28 '21 12:11 lightvector

https://stackoverflow.com/questions/69879188/could-not-load-library-cudnn-cnn-infer64-8-dll-error-code-126

mega-optimus avatar Nov 29 '21 06:11 mega-optimus

So there is a fix....

kiatchun avatar Nov 30 '21 09:11 kiatchun

I compiled Katago with cuda 11.6 + cudnn 8.3.2 + TensorRT 8.2 GA Update 1 and it's working fine.

ActiveIce avatar Jan 21 '22 06:01 ActiveIce

My result is that, using cuda 11.6 + cudnn 8.2.4.15 (11.4) is fine.

Using higher version of cudnn, it can still finish compiling without error, but it would caused an exception in cudabackend.cpp when the program runs.

sbbdms avatar Mar 15 '22 09:03 sbbdms