[DL使用指南2] 千年级3A大作 - Chiphell - 分享与交流用户体验

Ordered summary of results:

numSearchThreads =  5: 10 / 10 positions, visits/s = 768.55 nnEvals/s = 654.13 nnBatches/s = 262.88 avgBatchSize = 2.49 (10.5 secs) (EloDiff baseline)
numSearchThreads = 10: 10 / 10 positions, visits/s = 1370.05 nnEvals/s = 1143.12 nnBatches/s = 231.34 avgBatchSize = 4.94 (5.9 secs) (EloDiff +204)
numSearchThreads = 12: 10 / 10 positions, visits/s = 1520.07 nnEvals/s = 1265.04 nnBatches/s = 213.75 avgBatchSize = 5.92 (5.3 secs) (EloDiff +240)
numSearchThreads = 16: 10 / 10 positions, visits/s = 1478.80 nnEvals/s = 1283.18 nnBatches/s = 163.62 avgBatchSize = 7.84 (5.5 secs) (EloDiff +222)
numSearchThreads = 20: 10 / 10 positions, visits/s = 1668.26 nnEvals/s = 1449.97 nnBatches/s = 148.65 avgBatchSize = 9.75 (4.9 secs) (EloDiff +262)
numSearchThreads = 24: 10 / 10 positions, visits/s = 1649.51 nnEvals/s = 1453.94 nnBatches/s = 124.69 avgBatchSize = 11.66 (5.0 secs) (EloDiff +251)
numSearchThreads = 32: 10 / 10 positions, visits/s = 1835.77 nnEvals/s = 1658.07 nnBatches/s = 107.82 avgBatchSize = 15.38 (4.5 secs) (EloDiff +281)
numSearchThreads = 40: 10 / 10 positions, visits/s = 1879.08 nnEvals/s = 1722.75 nnBatches/s = 89.39 avgBatchSize = 19.27 (4.5 secs) (EloDiff +279)
numSearchThreads = 48: 10 / 10 positions, visits/s = 1885.12 nnEvals/s = 1761.32 nnBatches/s = 76.42 avgBatchSize = 23.05 (4.5 secs) (EloDiff +268)
numSearchThreads = 64: 10 / 10 positions, visits/s = 1931.60 nnEvals/s = 1849.76 nnBatches/s = 61.38 avgBatchSize = 30.14 (4.4 secs) (EloDiff +256)
numSearchThreads = 80: 10 / 10 positions, visits/s = 1938.04 nnEvals/s = 1898.51 nnBatches/s = 49.97 avgBatchSize = 38.00 (4.5 secs) (EloDiff +234)

Based on some test data, each speed doubling gains perhaps ~250 Elo by searching deeper.
Based on some test data, each thread costs perhaps 7 Elo if using 800 visits, and 2 Elo if using 5000 visits (by making MCTS worse).
So APPROXIMATELY based on this benchmark, if you intend to do a 5 second search:
numSearchThreads =  5: (baseline)
numSearchThreads = 10:  +204 Elo
numSearchThreads = 12:  +240 Elo
numSearchThreads = 16:  +222 Elo
numSearchThreads = 20:  +262 Elo
numSearchThreads = 24:  +251 Elo
numSearchThreads = 32:  +281 Elo (recommended)
numSearchThreads = 40:  +279 Elo
numSearchThreads = 48:  +268 Elo
numSearchThreads = 64:  +256 Elo
numSearchThreads = 80:  +234 Elo

If you care about performance, you may want to edit numSearchThreads in default_gtp.cfg based on the above results!
If you intend to do much longer searches, configure the seconds per game move you expect with the '-time' flag and benchmark again.
If you intend to do short or fixed-visit searches, use lower numSearchThreads for better strength, high threads will weaken strength.
If interested see also other notes about performance and mem usage in the top of default_gtp.cfg

2023-01-18 19:54:40+0800: GPU -1 finishing, processed 32333 rows 1243 batches