coprocessor xeon phi

[bupalo]

Is it useful for chess engines?

[reflectionofpower]

Is it useful for chess engines?

Unless I read it wrong it's 60 cores at 1.053 GHz a pop = 63 GHz. It would look much sweeter if it ran at say 2.7 GHz. For instance say a Dual 12 core setup with 2.7 Ghz a core, almost 65 GHz but at Turbo it's 84 GHz so if you could get that xeon phi up to 2 GHz it would look very appealing at 120 GHz.

[zullil]

Is it useful for chess engines?

My guess is no, but there's some basic information here: https://software.intel.com/en-us/articl ... ght-for-me

[Laskos]

Is it useful for chess engines?

Unless I read it wrong it's 60 cores at 1.053 GHz a pop = 63 GHz. It would look much sweeter if it ran at say 2.7 GHz. For instance say a Dual 12 core setup with 2.7 Ghz a core, almost 65 GHz but at Turbo it's 84 GHz so if you could get that xeon phi up to 2 GHz it would look very appealing at 120 GHz.

One has to consider that chess engines do not scale well to 60 cores. So, even an OC to 4.5 GHz 6 core i7 (total of 27 GHz) might overperform as strength goes that 60 core 1 GHz thing.

[bupalo]

Thanks ti all the replies .probably the true is that is so expensive that is better to buy some more cores with that price

[Werewolf]

A much more interesting question is whether Intel's plans to include an FPGA device on the CPU can be exploited by chess programmers.

That might result in serious speed up.

[ZirconiumX]

Is it useful for chess engines?

Unless I read it wrong it's 60 cores at 1.053 GHz a pop = 63 GHz. It would look much sweeter if it ran at say 2.7 GHz. For instance say a Dual 12 core setup with 2.7 Ghz a core, almost 65 GHz but at Turbo it's 84 GHz so if you could get that xeon phi up to 2 GHz it would look very appealing at 120 GHz.

One has to consider that chess engines do not scale well to 60 cores. So, even an OC to 4.5 GHz 6 core i7 (total of 27 GHz) might overperform as strength goes that 60 core 1 GHz thing.

At that many cores, something like Monte Carlo Tree Search starts to become a potentially powerful idea, as AB at that amount would have horrible scaling, where as MCTS usually has great scaling (practically linear), but is rather slow due to playing out entire chess games at random. I reckon the memory requirements could be a problem though, if you implement classical tree based MCTS.

Matthew:out

[Bloodbane]

I reckon the memory requirements could be a problem though, if you implement classical tree based MCTS.

If memory runs out just prune all nodes with visit count below some threshold. If not enough memory is freed increase the limit and try again. Another alternative is to stop expanding the tree and just run playouts but I don't like that.