CPU Scaling

[Leo]

I am wondering if there are any new developments on chess engines and how they scale. Is it still true that they still only scale to 16 cores? Why get a 128 core server for chess if it wont scale past 16 cores? Thanks for any input.

[mjlef]

I am wondering if there are any new developments on chess engines and how they scale. Is it still true that they still only scale to 16 cores? Why get a 128 core server for chess if it wont scale past 16 cores? Thanks for any input.

The limit to gains with more cores was more severe when program moslty you YBMW (https://chessprogramming.wikispaces.com ... 6704952c1e). WIth the switch to things like LazySMP, cores are no longer forced to wait, and stay active all the time. You can see this jump here on this page:

http://fastgm.de/threads5.html

Notice that Stockfish 5 gained only 4 elo doubling cores from 8 to 16. But Stockfish 7 which uses LazySMP jumped 48 elao from 8 to 16 cores.

Saying that, there is a clear trend of diminishing returns with more core doublings. For example: http://www.fastgm.de/schach/SMP-scaling-SF8-K10.4.pdf. There are some things that can improve this.

[noobpwnftw]

http://talkchess.com/forum/viewtopic.ph ... 263#759263

With more recent SMP approaches, basically you can get logarithmic improvements per core, or liner improvements per doubling cores till some hundred cores(tested). Beyond that, wild guess is it may be better than we thought.

[Leo]

I am wondering if there are any new developments on chess engines and how they scale. Is it still true that they still only scale to 16 cores? Why get a 128 core server for chess if it wont scale past 16 cores? Thanks for any input.

The limit to gains with more cores was more severe when program moslty you YBMW (https://chessprogramming.wikispaces.com ... 6704952c1e). WIth the switch to things like LazySMP, cores are no longer forced to wait, and stay active all the time. You can see this jump here on this page:

http://fastgm.de/threads5.html

Notice that Stockfish 5 gained only 4 elo doubling cores from 8 to 16. But Stockfish 7 which uses LazySMP jumped 48 elao from 8 to 16 cores.

Saying that, there is a clear trend of diminishing returns with more core doublings. For example: http://www.fastgm.de/schach/SMP-scaling-SF8-K10.4.pdf. There are some things that can improve this.

Thanks Mark. That answers my question. They still don't scale beyond 16 cores.