T1 4000000 wird immer noch bei TCEC und CCC Turnieren eingesetzt.
Das Training bei T81 verläuft - wie gewohnt - schleppend.
https://discord.com/channels/425419482568196106/539960268982059008/1136101972794151004Hier der letzte Test im Verhältnis zu Stockfish:
**Comment:** Here observed relative Sf-Lc0 strength (score %'s) nearly the same as:
CCC20 Blitz: Final (3|2) 500 games: Sf **58.6%-41.4%** Lc0_BT2-4150k.
**Hardware:** Ryzen 2700X | RTX 3080 10GB
**Lc0:**
https://github.com/Ergodice/lc0/commits/master Jul 22 2023 042efe2, defaults except Syzygy-6 MoveOverheadMs=0
**Lc0 nets:** t1-768x15x24h-swa-4000000.pb.gz , BT2-768x15-swa-4150000.pb.gz
**Stockfish_23071919:** abrok 64x avx2, defaults except Threads=14 Hash=8192 Syzygy-6 SyPrLim=6 Sy50MoRu=off Move Overhead=0
**Time control:** Game 1 min. + 1 sec.
**Gauntlet:** same openings per Lc0
**Openings:** TCEC SuFi rev., five Seasons: 20 + 21 + 22 + 23 + 24 (250)
**cutechess-cli:** policy=round, draw 35. 2 0.20, resign 4.00 2 two-sided, timemargin=200, restart=on ``` RATING ERROR POINTS PLAYED (%) CFS(%) W D L D(%)
1 Stockfish_23071919 : 68.3 17.9 585.0 1000 58.5 100 277 616 107 61.6
2 Lc0.31dagE042.t1-4000k : 0.0 ---- 207.5 500 41.5 50 54 307 139 61.4
3 Lc0.31dagE042.BT2-4510 : 0.0 24.9 207.5 500 41.5 --- 53 309 138 61.8```
Die Engines haben mit v0.30.0 und v031.0 Rev. neue Updates erhalten.
Letzte Rev. Version spielt derzeit bei CCC.
Interessant wäre noch die Entwicklung bei BT3 Netzen:
« Big transformer 3. Tentative new network arch using smolgen-augmented self-attention from BT2. The plan is to have embedding size 1024, ffn projection size 1024, 32 heads per layer, and 15 total layers with mish activation. Experiments are currently in progress to quantize the dense layers to INT8 or FP8 precision to improve speed, though the gains only start to show at embedding size 2048+ (there is a 50% speedup at 2048). There are also cuda optimizations available, which should reduce latency by 10 to 15%. More info at
https://github.com/Ergodice/lczero-training/blob/attention-net-body/README.mdLast updated: 7/17/23 «