AMD Threadripper 2990x seems out what is the gain vs 1950x ?

[jpqy]

Here you have some benches and a compare :

http://www.ipmanchess.yolasite.com/amd- ... -bench.php

JP.

[Leo]

Here you have some benches and a compare :

http://www.ipmanchess.yolasite.com/amd- ... -bench.php

JP.

Thanks. Pretty impressive. About 8 years ago 30,000 nps was really good.

[Michael Sherwin]

To gain the best advantage of the Threadripper 2990wx in a bitboard engine avoid magic bitboards or any other memory intensive variation. Also a smaller hash might result in better node/second performance. The reason for this is that there is a memory access bottleneck imposed on the 16 additional cores that is so inefficient that 32 cores can in memory intensive programs run slower than 16 cores.

[Leo]

To gain the best advantage of the Threadripper 2990wx in a bitboard engine avoid magic bitboards or any other memory intensive variation. Also a smaller hash might result in better node/second performance. The reason for this is that there is a memory access bottleneck imposed on the 16 additional cores that is so inefficient that 32 cores can in memory intensive programs run slower than 16 cores.

Is Stockfish a bitboard engine? What is a magic bitboard? What size hash would you recommend?

[Werewolf]

The 32 core threadripper also has stunted memory access.

Presumably EPYC doesn't.

[Michael Sherwin]

To gain the best advantage of the Threadripper 2990wx in a bitboard engine avoid magic bitboards or any other memory intensive variation. Also a smaller hash might result in better node/second performance. The reason for this is that there is a memory access bottleneck imposed on the 16 additional cores that is so inefficient that 32 cores can in memory intensive programs run slower than 16 cores.

Is Stockfish a bitboard engine? What is a magic bitboard? What size hash would you recommend?

Is Stockfish a bitboard engine? Yes Although I have not investigated what kind.

What is a magic bitboard? I don't have that definition in memory anymore, LOL. Anyway Crafty uses magic bitboards and does terrible nodes/second compared to Stockfish which I believe is not magic bitboard.

What size hash for stockfish? Trial and error would be the answer there as I do not have a 2990wx system.

[mar]

Anyway Crafty uses magic bitboards and does terrible nodes/second compared to Stockfish

Do you have any data on that? I remember Crafty actually had probably the highest nps I ever saw (not so long ago).
Anyway - comparing nps of two different engines is irrelevant, also it doesn't correlate well with elo,
an engine with half/quarter the nps of another can still be hundreds of elo stronger.

[Michael Sherwin]

Anyway Crafty uses magic bitboards and does terrible nodes/second compared to Stockfish

Do you have any data on that? I remember Crafty actually had probably the highest nps I ever saw (not so long ago).
Anyway - comparing nps of two different engines is irrelevant, also it doesn't correlate well with elo,
an engine with half/quarter the nps of another can still be hundreds of elo stronger.

Crafty on a 32 core TR 2990wx is not faster than on a 16 core TR 2950x while SF gains about 65% more nodes per second moving from the TR 2950 to the TR 2990wx. The reason that SF does not double it's node rate from the 50 to the 90 is because the hashtable still weighs on the memory bottleneck. So to try to be clear there is no comparison being done directly between SF and Crafty. The comparison is between Crafty and Crafty and between SF and SF on the different processors.

[Dann Corbit]

The 32 core threadripper also has stunted memory access.

Presumably EPYC doesn't.

Correct. Epyc has twice the memory bandwidth.