Lc0 CUDA 10.2 compile (0.25.1) is faster

[Eduard]

Lc0 0.25.1 CUDA 10.2 compile is fast.

You can download this lc0 binary for Windows 10 at Discord under test-discuss, Mai 1.

Or here, with all needed Nvidia 10.2 drivers (340 MB):

https://filehorst.de/d/dyhbuIiz

Short test CUDA 10.2 compile under Win 10 on my slow GTX 1050 Ti, after 60s:
Sergio 3200, new 950 nps (old 650 nps)
Sergio 1810, new 3000 nps (old 2100 nps)
701820, new 15 kns (old 14 kns)
Fat Fritz 1.1, new 2900 nps (old 2500 nps)

old=lc0 0.25.1 with old Nvidia drivers, new=Lc0 0.25.1_10.2 compile with Nvidia drivers 10.2.

Great work, thanks!

[MMarco]

Thank you for the info.

Unfortunately there's something wrong with the new combo drivers+compile on with GTX 1660 ti : with 20x256 nets the nps drastically dropped from 9000 to a bare 1000 nps... Does someone have the same problem?

[corres]

As I remember well CUDA 10.2 needs VS 2019 compiler and newer NVIDIA GPU driver.

[Werner]

Thank you for the info.

Unfortunately there's something wrong with the new combo drivers+compile on with GTX 1660 ti : with 20x256 nets the nps drastically dropped from 9000 to a bare 1000 nps... Does someone have the same problem?

For Gtx 16...
set `--backend-opts=custom_winograd=false` when using official v0.25 build.

[corres]

As I remember well CUDA 10.2 needs VS 2019 compiler and newer NVIDIA GPU driver.

NVIDIA Windows 10 64 bits and Windows 7 64 bits driver ver.445.87 is applicable to CUDA 10.2 but some user marked some issue with it.
Any experience using for Leela?

[Laskos]

As I remember well CUDA 10.2 needs VS 2019 compiler and newer NVIDIA GPU driver.

NVIDIA Windows 10 64 bits and Windows 7 64 bits driver ver.445.87 is applicable to CUDA 10.2 but some user marked some issue with it.
Any experience using for Leela?

I am using it, seems fine, both this compile and the older v25.1 compile. Only that I am not getting any speed gain with this CUDA 10.2 build. Maybe only fp32 gets big gains.

[corres]

As I remember well CUDA 10.2 needs VS 2019 compiler and newer NVIDIA GPU driver.

NVIDIA Windows 10 64 bits and Windows 7 64 bits driver ver.445.87 is applicable to CUDA 10.2 but some user marked some issue with it.
Any experience using for Leela?

I am using it, seems fine, both this compile and the older v25.1 compile. Only that I am not getting any speed gain with this CUDA 10.2 build. Maybe only fp32 gets big gains.

Thanks for the info but if I am right you use DX12 backend and not cudnn-fp16 backend.

[Laskos]

As I remember well CUDA 10.2 needs VS 2019 compiler and newer NVIDIA GPU driver.

NVIDIA Windows 10 64 bits and Windows 7 64 bits driver ver.445.87 is applicable to CUDA 10.2 but some user marked some issue with it.
Any experience using for Leela?

I am using it, seems fine, both this compile and the older v25.1 compile. Only that I am not getting any speed gain with this CUDA 10.2 build. Maybe only fp32 gets big gains.

Thanks for the info but if I am right you use DX12 backend and not cudnn-fp16 backend.

No, I benchmarked that T...160 20x256 JHorthos net using cudnn-fp16 backend. DX12 is good with larger nets, but here the issue was this CUDA 10.2 compile and libraries.

[corres]

As I remember well CUDA 10.2 needs VS 2019 compiler and newer NVIDIA GPU driver.

NVIDIA Windows 10 64 bits and Windows 7 64 bits driver ver.445.87 is applicable to CUDA 10.2 but some user marked some issue with it.
Any experience using for Leela?

I am using it, seems fine, both this compile and the older v25.1 compile. Only that I am not getting any speed gain with this CUDA 10.2 build. Maybe only fp32 gets big gains.

Thanks for the info but if I am right you use DX12 backend and not cudnn-fp16 backend.

No, I benchmarked that T...160 20x256 JHorthos net using cudnn-fp16 backend. DX12 is good with larger nets, but here the issue was this CUDA 10.2 compile and libraries.

I see.

[pohl4711]

I am using it, seems fine, both this compile and the older v25.1 compile. Only that I am not getting any speed gain with this CUDA 10.2 build. Maybe only fp32 gets big gains.

Same here on my RTX 2060 (mobile). No speed gain compared to offcial 0.25.1 release with cudnn-fp16.