Auf TCEC wird zur Analyse auch das 4096 Netz herangezogen.
Es rechnet tiefer als das Standart Netz.

Hier einige Informationen auf Discord:

**Format:**  Double gauntlet, same as at SuFi 27 binaries: Stockfish and Redfish vs. Lc0 BT4-it332 and BT4-1740 (no same engine games).
**Hardware:**  R7 2700X (8c/16t) | RTX 3080 10GB (SM=68)  | Syzygy-6
**Stockfish:**  SuFi 27 version Stockfish_24120821, abrok avx2; and SuFi 27 kibitzer "Redfish" a recent Stockfish with 4096 net, https://github.com/xu-shawn/Stockfish/tree/4096net  avx2 compile, bench ok
**Stockfish settings:**  12 threads, 1GB hash, Sy.50mr off, m.o. 0
**Lc0:**  https://github.com/uwuplant/lc0/tree/daniel-update-contempt-fix  compiled w/ Cutlass
**Lc0 settings:**  Same as at SuFi 27, see SuFi27-config-note.txt
**Lc0 nets:**  BT4-1024x15x32h-swa-6147500-policytune-332.pb.gz , BT4-1740.pb.gz
**Openings:**  UHO_Lichess_4852_v1.epd random rev., different opening per game pair
**Time control:  Stockfish/Redfish:** G/8s + .1s , **Lc0:** G/16s + .2s
**cutechess-cli:**  draw 35. 4 .20, resign 3 5.00 twosided, timemargin=200 restart=on policy=round -srand 2024122420241231 -wait 1000   **Ordo 1.2.6:**  s=2400```                    RATING  ERROR  POINTS  PLAYED   (%)  CFS(%)     W     D     L  D(%)
1 Stockfish_SuFi27 :  33.8    6.2  2321.0    4000  58.0     100  1376  1890   734  47.3
2 Redfish_4096net  :  13.9    6.3  2211.0    4000  55.3     100  1255  1912   833  47.8
3 Lc0.it332_SuFi27 : -20.4    6.1  1753.0    4000  43.8      89   779  1948  1273  48.7
4 Lc0.sufi27_1740  : -27.3    6.4  1715.0    4000  42.9     ---   788  1854  1358  46.4```**Comments:**  See SuFi-config-note.txt, the nets' difference kept declining roughly as much as error keeping CFS low, here 1740 ~11% chance of being better.
In the Stockfish Discord search `has:4096` for recent posts on Stockfish with 4096 net.