Lc0 in GTX video card

[Jouni]

Thanks for all replies. Lc0 page has misleading info about fastest drivers, but I used correct ones. And cudnn-fp16 is used automatically:

Creating backend [cudnn-auto]...
Switching to [cudnn-fp16]...
CUDA Runtime version: 10.0.0
Cudnn version: 7.4.2
Latest version of CUDA supported by the driver: 11.1.0

BTW in mwyoung test 25.12. Lc0 in RTX2080TI loses badly to Stockfish 16 cores!

[Ferdy]

Run the full benchmark command, just for reference for this particular hardware.

CPU: i7-2600K, 3.4Ghz 4cores/8threads
GPU: GTX 1650 super
Cuda: V11.2.67
Name: Lc0 v0.26.3, backend: cudnn-fp16
Command: ./lc0 benchmark -w nn_name

Code: Select all

+-----------+-----------+-------+
|   nn_name | nn_size   |   nps |
|-----------+-----------+-------|
|    703810 | 10x128    | 44998 |
|     42850 | 20x256    |  8502 |
|     66680 | 30x384    |  2520 |
+-----------+-----------+-------+

[MMarco]

Here's my results:

CPU: Ryzen-7 3750H, 2.3Ghz 4cores/8threads
GPU: GTX 1660 Ti mobile
Cuda drivers: 11.2.0
Name: Lc0 v0.26.3, backend: cudnn-fp16
Command: ./lc0 benchmark -w nn_name

Code: Select all

+-----------+-----------+-------+
|   nn_name | nn_size   |   nps |
|-----------+-----------+-------|
|    703810 | 10x128    | 57157 |
|     42850 | 20x256    |  9825 |
|     66680 | 30x384    |  2933 |
+-----------+-----------+-------+

[Jouni]

How I run benchmark? https://lczero.org/play/configuration/flags/ says nothing about it.

[MMarco]

How I run benchmark? https://lczero.org/play/configuration/flags/ says nothing about it.

In windows I open a command prompt in Lc0 folder. Then I use the following command: ./lc0 benchmark

This is the default bench at 10s per positions on 34 positions. You can also add extra commands like:

./lc0 benchmark --weights=path/filebane.pb --threads=2 --minibatch-size=256 --movetime=1000 (the last one being time per positions in milliseconds, or --nodes=100000) etc. as shown in the link you gave.

[Jouni]

Don't work here:

Microsoft Windows [Version 10.0.18363.1256]
(c) 2019 Microsoft Corporation.

c:\Engines\Lc0>./lc0 benchmark
'.' is not recognized as an internal or external command,
operable program or batch file.

c:\Engines\Lc0>

[jshriver]

Don't work here:

Microsoft Windows [Version 10.0.18363.1256]
(c) 2019 Microsoft Corporation.

c:\Engines\Lc0>./lc0 benchmark
'.' is not recognized as an internal or external command,
operable program or batch file.

c:\Engines\Lc0>

Windows is different try .\lc0.exe benchmark

The ./ for linux or .\ for windows just means run this program in the current directory.

[Jouni]

That works, thanks.

[Ferdy]

Don't work here:

Microsoft Windows [Version 10.0.18363.1256]
(c) 2019 Microsoft Corporation.

c:\Engines\Lc0>./lc0 benchmark
'.' is not recognized as an internal or external command,
operable program or batch file.

c:\Engines\Lc0>

Windows 10 has power shell. Both forward and back slash would work.

[Werner]

I tested the 2 versions on my GTX 1650 Super under the same conditions with go movetime 60000; NN 10x128

LC0 Cuda 0.26.3 backend cuda nps 19247
LC0 Cuda 0.26.3 backend cuda-fp16 nps 6634

LC0 CUDNN 0.26.3 backend cudnn nps 22990
LC0 CUDNN 0.26.3 backend cudnn-fp16 nps 37701

LC0 CUDNN 0.26.3 backend cuda nps 23732
LC0 CUDNN 0.26.3 backend cuda-fp16 nps 37573

So I think, the older cudnn version maybe the strongest with this card.