The last time parity benchmark I did was between KataGo v1.3.2+g170-20x256-1.9G (g170-b20c256x2-s1913382912-d435450331) and Leela Zero v0.17+#263. I tested with 2xRTX2080Ti, with 60 threads for KataGo and 20 threads for LZ. The results were:
KataGo (1600 po) vs Leela Zero (1000 po): KataGo won 279-145 (65.8%), Elo: +113.7 +- 31.4 KataGo (36000 po) vs Leela Zero (20000 po): KataGo won 91-39 (70.0%), Elo: +147.2 +- 55.0 KataGo (1 min/move) vs Leela Zero (1 min/move): KataGo won 55-19 (74.3%), Elo: +184.6 +- 69.6
Now there are new networks both for KataGo and Leela Zero so they need to be benchmarked further, but I expect no significant difference, due to the inflated nature of the self-Elo.
Its Elo evolution over time is on the GitHub page, though it is updated only when there is a new release. Networks and of course executables can be downloaded from the GitHub as well. For the time being, the latest 20b network (g170e-b20c256x2-s2430231552-d525879064) is considered the strongest one in time parity, at least until 10s/move or so.
|