Lc0 28.0 + 610k vs Stockfish 14 (various threads)

MMarco · Post by **MMarco** » Sat Oct 02, 2021 2:42 am

Out of curioty, I ran Lc0 28.0 with a recent mainrun net (id. 610000, 384x30) against threaded versions of Stockfish 14 (1 thread, 4 threads, 8 threads). Here are the specs used.

Time control: 100s +1s
Hardware: 3070 and 10700 (all threads locked at 4.0 Ghz)
Openings: TCEC Superfinals 16 to 20 (250 positions)
Lc0 28.0: --backend-opts=multi_stream=true, --nncache=10000000 --minibatch-size=80
Syzygy: 3-4-5 mens
Adjudication: win if both engine score +500 cp for 5 moves consecutively, draw when both engines scores no greater than 5 cp for 5 moves consecutively.
Benchmark: 10s from the start position (bmi2 for SF, cuda-fp16 for Lc0)
Lc0 28.0 + 610k = 18 200 nps
SF14 (1T) = 1 400 000 nps
SF14 (4T) = 5 600 000 nps
SF14 (8T) = 10 100 000 nps

From 1 to 8 threads, Stockfish gained a bit more than 100 elo here. It seems than doubling the numbers of threads gives around +35 elo under these test conditions.

Code: Select all

# PLAYER              :  RATING  ERROR  PLAYED  (%)   CFS    W   D    L    D(%)
1 Stockfish 14 (8T)   :    30.1   16.6    500  54.00  100   95  350   55  70.00
2 Lc0 28.0 + 610k     :     0.0   ----   1500  52.27   67  285  998  217  66.53
3 Stockfish 14 (4T)   :    -3.8   16.7    500  49.50  100   81  333   86  66.60
4 Stockfish 14 (1T)   :   -78.5   17.2    500  39.70  ---   41  315  144  63.00

White advantage = 93.47 +/- 5.10
Draw rate (equal opponents) = 77.79 % +/- 1.43

Depth statistics

Code: Select all

Engine             sec/mov  Depth    MIDG   EARLY   ENDG    LATE
Lc0 28.0 + 610k      2.22   12.75   14.65 | 14.97 | 11.06 |  8.69
Stockfish 14 (1T)    2.25   29.46   26.10 | 26.27 | 29.88 | 43.02
Stockfish 14 (4T)    2.17   31.81   27.96 | 28.41 | 33.25 | 44.66
Stockfish 14 (8T)    2.18   32.75   28.73 | 28.53 | 34.21 | 46.64