Hmm.. 1 thread is better than 2 threads with Lc0 and large 30x384 nets?

[Laskos]

Almost accidentally got this result with LittleBlitzer checking the in-game NPS of Lc0 with J92-190 net on 1 thread and 2 threads (GPU is RTX 2070). 2-threaded is not faster than 1-threaded and the result in 100 games was advantageous to 1 threaded (15s+0.25s games).

Code: Select all

Games Completed = 100 of 100 (Avg game length = 57.919 sec)
Settings = Gauntlet/128MB/15000ms+250ms/M 700cp for 3 moves, D 80 moves/EPD:C:\LittleBlitzer\2moves_80_100.epd(1749)
Time = 6388 sec elapsed, 0 sec remaining
 1.  Lc0 J92-190 Threads=2    	45.5/100	9-18-73  	(L: m=0 t=0 i=0 a=18)	(D: r=31 i=16 f=0 s=0 a=26)	(tpm=465.7 d=8.64 nps=9527)
 2.  Lc0 J92-190 Threads=1    	54.5/100	18-9-73  	(L: m=0 t=0 i=0 a=9)	(D: r=31 i=16 f=0 s=0 a=26)	(tpm=465.4 d=8.63 nps=9616)

To check more thoroughly, I used then Cutechess to play 1000 such games:

Code: Select all

Score of Lc0_J92-190 1 thread vs Lc0_J92-190 2 threads: 305 - 250 - 445 [0.527]
...      Lc0_J92-190 1 thread playing White: 264 - 23 - 213  [0.741] 500
...      Lc0_J92-190 1 thread playing Black: 41 - 227 - 232  [0.314] 500
...      White vs Black: 491 - 64 - 445  [0.714] 1000
Elo difference: 19.1 +/- 16.0, LOS: 99.0 %, DrawRatio: 44.5 %
1000 of 1000 games finished.

LOS = 99.0% without using the correct pentonamial variance which would give a LOS > 99.5%. Decisive result and I wonder whether it can be generalized to RTX GPUs and large nets.