Shahrad wrote: ↑Tue Mar 30, 2021 1:00 pm
Many thanks for all the replies.
There is a weird problem: lc0 cudnn is 10kN/s faster than the cuda. As said it has a 2060 RTX.
That is the case for small networks (10x128) like the one included (703810) in Lc0 official release for many graphic card.
NPS will be much slower than for a small networks, but Lc0 will play stronger due to added "knowledge". For large networks, Cuda will be faster than CudNN. Also make sure you have the latest drivers installed for your card. If those are old, they might not use recent optimizations in the Lc0 binary.
Good to know, I installed the J94-100, thank you. Now it is running at 4kn/s and the question is at which depth the evaluation is somewhat reliable? I am using it only for analysis and not for engine tournaments.
It is unclear to me what you mean by "somewhat reliable". Lc0 is tactically weaker than Stockfish and will miss more tactics if the two engines are given the same thinking time. But in some positions it will do better. For example, in a comp-comp competition last year on big harware (88 physical / 176 threads for Stockfish, 4x V100 for Lc0) Lc0 found a truly deep sacrifice after a Stockfish mistake: http://talkchess.com/forum3/viewtopic.p ... on#p848855 (maybe newer version of Stockfish won't fall into it thought. I didn't verify).
At 4 kn/s, Lc0 27.0 + J94-100 will certainly beat easily Magnus Carlsen at 1 sec/move. At this time control, I would expect Lc0 to miss some tactics, but not to drop pieces or pawns due to a tactical shot a human would see.
Shahrad wrote: ↑Tue Mar 30, 2021 1:00 pm
Many thanks for all the replies.
There is a weird problem: lc0 cudnn is 10kN/s faster than the cuda. As said it has a 2060 RTX.
That is the case for small networks (10x128) like the one included (703810) in Lc0 official release for many graphic card.
NPS will be much slower than for a small networks, but Lc0 will play stronger due to added "knowledge". For large networks, Cuda will be faster than CudNN. Also make sure you have the latest drivers installed for your card. If those are old, they might not use recent optimizations in the Lc0 binary.
Good to know, I installed the J94-100, thank you. Now it is running at 4kn/s and the question is at which depth the evaluation is somewhat reliable? I am using it only for analysis and not for engine tournaments.
I forgot to say that in the endgame, the fewer pieces on the board left, the stronger will be Stockfish relative to Lc0 as Stockfish will be able to calculate almost perfectly till the end of the game. Lc0 relative strenght resides more in complex openings/middle game/early endgame positions. In queen's ending Stockfish is vastly superior, Lc0 is having trouble to get all transpositions and gets often tricked by SF in comp-comp games I saw. However in my opinion, Lc0 is superior in rook endings as her extra knowledge pays off (i.e. where to put the rook according to the pawn structure, is able to recognize drawn rook+h+f pawns vs rook etc. and many other patterns) and transpositions aren't as bad as in queen's endings.
Shahrad wrote: ↑Tue Mar 30, 2021 1:00 pm
Many thanks for all the replies.
There is a weird problem: lc0 cudnn is 10kN/s faster than the cuda. As said it has a 2060 RTX.
That is the case for small networks (10x128) like the one included (703810) in Lc0 official release for many graphic card.
NPS will be much slower than for a small networks, but Lc0 will play stronger due to added "knowledge". For large networks, Cuda will be faster than CudNN. Also make sure you have the latest drivers installed for your card. If those are old, they might not use recent optimizations in the Lc0 binary.
Good to know, I installed the J94-100, thank you. Now it is running at 4kn/s and the question is at which depth the evaluation is somewhat reliable? I am using it only for analysis and not for engine tournaments.
It is unclear to me what you mean by "somewhat reliable". Lc0 is tactically weaker than Stockfish and will miss more tactics if the two engines are given the same thinking time. But in some positions it will do better. For example, in a comp-comp competition last year on big harware (88 physical / 176 threads for Stockfish, 4x V100 for Lc0) Lc0 found a truly deep sacrifice after a Stockfish mistake: http://talkchess.com/forum3/viewtopic.p ... on#p848855 (maybe newer version of Stockfish won't fall into it thought. I didn't verify).
At 4 kn/s, Lc0 27.0 + J94-100 will certainly beat easily Magnus Carlsen at 1 sec/move. At this time control, I would expect Lc0 to miss some tactics, but not to drop pieces or pawns due to a tactical shot a human would see.
With somewhat reliable, I mean the move is not a move of a 2600 elo Grandmaster but a 2900 elo engine (lol).