LC0 @ low power GPU

[corres]

My integrated card Radeon rx vega 10 runs at 10w and the bests 10x128 nets (i.e. 591226) are on par with single threaded Stockfish 11 on my Ryzen 7 3750h at blitz time control. Roughly, Lc0 reaches about 5000 nps (without tree reuse) while Stockfish about 1 500 000.

That seems pretty impressive to me !

Indeed! The onboard card is surprisingly strong, at least with 10x128 nets. I'll eventually try larger ones to see what happens then.

For the explanation see the above post.

[MMarco]

My integrated card Radeon rx vega 10 runs at 10w and the bests 10x128 nets (i.e. 591226) are on par with single threaded Stockfish 11 on my Ryzen 7 3750h at blitz time control. Roughly, Lc0 reaches about 5000 nps (without tree reuse) while Stockfish about 1 500 000.

That seems pretty impressive to me !

Indeed! The onboard card is surprisingly strong, at least with 10x128 nets. I'll eventually try larger ones to see what happens then.

For the explanation see the above post.

Thanks for sharing the information. I find it interesting (and somewhat surprising). I'll try to read to (vaguely at least) understand how NN works.

If I understood correctly what you said, networks will get stronger with GPU improvements even without software improvement in the binaries or changes in the way training runs are done (we would only need to rerun them on the new hardware to get an improved network). That would be good news to me as it would garantee future strenght increases on my current hardware.

[corres]

My integrated card Radeon rx vega 10 runs at 10w and the bests 10x128 nets (i.e. 591226) are on par with single threaded Stockfish 11 on my Ryzen 7 3750h at blitz time control. Roughly, Lc0 reaches about 5000 nps (without tree reuse) while Stockfish about 1 500 000.

That seems pretty impressive to me !

Indeed! The onboard card is surprisingly strong, at least with 10x128 nets. I'll eventually try larger ones to see what happens then.

For the explanation see the above post.

If I understood correctly what you said, networks will get stronger with GPU improvements even without software improvement in the binaries or changes in the way training runs are done (we would only need to rerun them on the new hardware to get an improved network). That would be good news to me as it would garantee future strenght increases on my current hardware.

One thing is the NN quality and another thing is the software quality, and the third thing is the power of your system, and the fourth thing the position what you analyze.
All of them determine the chess power of the engine used by you for analyze.
In the case of an engine-engine match there some other factor too, for e.g engine parameters (kiudee, etc) system parameters (HT, turbo usage, system temperature and cooling) the tournament time, opening positions, endgame tablebases and last but not least the chess power of the opposite engine.

[yurikvelo]

This is what I got with 75W GTX 1650
(Lc0 0.25.1 and latest networks from official site: 591226, 1541 and 63702)

[yurikvelo]

Lc0 0.25.1 CUDA 256x20-t40-1541 vs Stockfish_20060419_x64_modern
GPU=GTX 1650; benchmark=5419 nps, startpos=4486 nps
CPU: 2 Threads @ Ryzen 3400G, startpos=3084 MN/sec
Hash=2048
Syzygy = 6 man
Ponder = On
StartPGNEPD=8moves_v3.pgn
RepeatStartPosition=1
Contempt settings: default (bundled)
TC=2' + 1"

+18 -10 =73


3 MN/sec is not enough for SF-Dev, will try 3 cores

[yurikvelo]

256x20-t40-1541 vs Stockfish_20060419 4CPU roughly equal (at Ponder=On + 2' + 1" TC)

That is 4486 nps VS 5.45 MN/sec (startpos)

Large net 24x320-63702 at 3095 nps seem a bit weaker (maybe 3CPU is enough against GTX-1650)