Page 1 of 1
Testmatch with Net 58573
Posted: Thu Dec 12, 2019 12:10 pm
by Werner
Hi,
I played 100 games against Stockfish 10 x64 1CPU under CEGT 40/20 conditions. Grafikcard is GTX 1650:
1 LCZero 0.23.1 Cuda (58573) +45 +23/=67/-10 56.50% 56.5/100
2 Stockfish 10.0 x64 1CPU -45 +10/=67/-23 43.50% 43.5/100
I ask me: are larger networks really better under these test conditions ?
Re: Testmatch with Net 58573
Posted: Thu Dec 12, 2019 3:01 pm
by Laskos
Werner wrote: ↑Thu Dec 12, 2019 12:10 pm
Hi,
I played 100 games against Stockfish 10 x64 1CPU under CEGT 40/20 conditions. Grafikcard is GTX 1650:
1 LCZero 0.23.1 Cuda (58573) +45 +23/=67/-10 56.50% 56.5/100
2 Stockfish 10.0 x64 1CPU -45 +10/=67/-23 43.50% 43.5/100
I ask me: are larger networks really better under these test conditions ?
Hard to say. Are these 40/20' real 20 minutes time control for 40 moves, or adjusted "20 minutes"? If they are real, it's quite a lot of time per move, and it might be that even on GTX 1650 with its fp-16 larger T40 would be better. A close call, I would think.
Re: Testmatch with Net 58573
Posted: Thu Dec 12, 2019 3:42 pm
by Werner
Thanks for the answer. Real time control was 40/10.
Re: Testmatch with Net 58573
Posted: Wed Dec 18, 2019 6:22 am
by bmp1974
Werner wrote: ↑Thu Dec 12, 2019 12:10 pm
Hi,
I played 100 games against Stockfish 10 x64 1CPU under CEGT 40/20 conditions. Grafikcard is GTX 1650:
1 LCZero 0.23.1 Cuda (58573) +45 +23/=67/-10 56.50% 56.5/100
2 Stockfish 10.0 x64 1CPU -45 +10/=67/-23 43.50% 43.5/100
I ask me: are larger networks really better under these test conditions ?
I feel 6 series network plays much stronger than 5 series. Eventually, 6 series will prevail over 4 series.