How to run katago on rtx 3090

[drixs2050]

It appears that katago does not run on rtx 3090 while it does run on a rtx 2080s with the same setup.
Im not really sure what the problem really is
this is what I got After I run:

2020-09-30 16:07:01-0400: Loading model and initializing benchmark...
2020-09-30 16:07:01-0400: nnRandSeed0 = 13077320275990283880
2020-09-30 16:07:01-0400: After dedups: nnModelFile0 = .\g170-b40c256x2-s5095420928-d1229425124.bin.gz useFP16 auto useNHWC auto
2020-09-30 16:07:03-0400: Cuda backend thread 0: Found GPU GeForce RTX 3090 memory 25769803776 compute capability major 8 minor 6
2020-09-30 16:07:03-0400: Cuda backend thread 0: Model version 8 useFP16 = true useNHWC = true
2020-09-30 16:07:03-0400: Cuda backend thread 0: Model name: g170-b40c256x2-s5095420928-d1229425124

and the command line just got stuck here.
I suppose the problem could be the version of cuda used but I think katago only support cuda 10.2 so upgrade to cuda 11.1 does not appear to be possible solution.
Can someone help? thank you

[lightvector]

Whenever the next release is, it will likely support CUDA 11. Or alternatively, if you can compile KataGo yourself, the tip of the master branch on github already supports it. https://github.com/lightvector/KataGo

Otherwise, you should just use the OpenCL version. In fact, most people have reported that on RTX2080, the OpenCL version is *faster* than the CUDA version. Good chance the same is true for RTX30**.

So there's actually not a big reason to use CUDA - CUDA and especially CUDNN are more annoying to install, legally obtaining them they require you to agree to more annoying NVIDIA license terms, and they're not even clearly better than the custom hand-written OpenCL code.

[RobertJasiek]

What about OptiX? Could that further accelerate KataGo, does it already use it or do CUDA / OpenCL already use the cards optimally? I wonder because benchmarks are floating around showing that OptiX could be 3 or 6 times faster than CUDA / OpenCL for some non-go, non-3D-gaming softwares.