No, no, speed alone affects Elo-wise similarly to speed of regular engines. Similar Elo per doubling, higher at small time controls and weaker hardware, lower Elo per doubling to LTC and stronger hardware. Bigger nets are better at slower TC and stronger hardware. There was some speculation that at something like TCEC conditions, Lc0 scaling with TC and hardware is much worse than that of SF. It's possible, Lc0 seems to lose to SF in TCEC superfinal worse than expected for that Leela Ratio.yurikvelo wrote: ↑Wed Nov 04, 2020 7:43 pmI mean 2x faster hardware at any position or time control.
E.g. GPU #1 is 5 knps after 1 min at some position and GPU #2 is 10 knps after 1/2 min on the same position.
What will be ELO difference if run machine with GPU #1 vs machine with GPU #2
is traditional 40...60 ELO per doubling apply here?
When asking about networks (bigger vs smaller) for non-RTX GPU someone said bigger network is stronger and slow speed matters only if you are going to donate GPU for network learning, but do not matter for play. In other words - slow speed doesn't affect ELO much.
EDIT: More precisely, IIRC it was claimed that above 1 million nodes per move of Lc0 MCTS search, the scaling becomes very poor.