Looking for a CPU recommendation

[kasinp]

I am planning to purchase (or build) a new system. What would be a reasonably high-end CPU for chess these days? I'd like it to be significantly stronger than the old 4-core system it replaces.

Would love to hear your first-hand experiences and advice.

Thank you,
Peter

[Steve Maughan]

AMD's new Zen processor look awesome.

http://techfrag.com/2016/10/04/naples-cpu-benchmarks/

Systems will probably only show up early next year.

- Steve

[matthewlai]

I am planning to purchase (or build) a new system. What would be a reasonably high-end CPU for chess these days? I'd like it to be significantly stronger than the old 4-core system it replaces.

Would love to hear your first-hand experiences and advice.

Thank you,
Peter

If you can wait a bit, I would wait for the AMD Zen before deciding. That will take a few months, though.

If you want to build it now, I would get 2xE5-2670 from eBay, and something like this - http://www.ebay.com/itm/New-ASRock-Rack ... SwbsBXlsVU

Those chips are being sold for peanuts at the moment since Facebook decommissioned a bunch of them earlier this year, and completely flooded the market.

You get 16 cores. It will be faster than an i7-6900k, and the whole system will cost less than that chip alone.

[kasinp]

Thank you both. I was not aware of this new AMD chip; seems it may be worth waiting for it. I will need to read more to understand power/heat parameters though.

Peter

[jdart]

I do not recommend the older dual Xeons if what you care about is total performance of one engine on the box. They really aren't that fast per core and engines generally can't use all cores at 100%. But these boxes are good for testing and running matches because they have decent performance and a lot of cores.

For max performance I'd get the i7-6900k myself (assuming you can't afford a i7-6950k or a really high end modern Xeon). It uses faster RAM and can be overclocked.

--Jon

[matthewlai]

I do not recommend the older dual Xeons if what you care about is total performance of one engine on the box. They really aren't that fast per core and engines generally can't use all cores at 100%. But these boxes are good for testing and running matches because they have decent performance and a lot of cores.

For max performance I'd get the i7-6900k myself (assuming you can't afford a i7-6950k or a really high end modern Xeon). It uses faster RAM and can be overclocked.

--Jon

It does depend on how the application scales, but Sandy Bridge-E IPC is only about 15% lower than Broadwell E's. So the dual Xeons are still faster if the application scales well, and they are MUCH cheaper than the i7-6900k.

A 6900k + X99 motherboard is $1091 + ~$250 = $1341
2xE5-2670 + C602 motherboard is 2x$90 + ~$300 = $480

It's true that in absolute terms the 6900k will probably be better for chess, due to chess engines not scaling that well past 8 cores, but the difference won't be that much, and the price difference is almost 3x.

It does use faster RAM, but a dual Xeon setup would have 8 memory channels (vs 4), and twice the amount of cache. Though that's offset by NUMA inefficiencies, and it gets complicated.

6900k is a nice chip, but the pricing is just way out there in unreasonable land.

[jdart]

I happen to have both machines you mentioned. Plus a bunch of others.

Arasan on this position with 60 seconds/move:

4n3/2p5/1p2r2P/p1p2R2/P1N1k3/1PP4K/8/8 w - - bm Re5+; id "arasan19.43"; c0 "Anand-Landa, Bundesliga 0506 2005";

i7-6900k, 8 cores (12G hash): 6.7M nodes per second.

dual e5-2670, 16 cores (12G hash): 5.27M nodes per second.

So it is close but the i7 is faster. And this machine is not overclocked. You can crank the clock rate on a typical i7 a lot higher than I have it.

--Jon

[Vinvin]

I happen to have both machines you mentioned. Plus a bunch of others.

Arasan on this position with 60 seconds/move:

4n3/2p5/1p2r2P/p1p2R2/P1N1k3/1PP4K/8/8 w - - bm Re5+; id "arasan19.43"; c0 "Anand-Landa, Bundesliga 0506 2005";

i7-6900k, 8 cores (12G hash): 6.7M nodes per second.

dual e5-2670, 16 cores (12G hash): 5.27M nodes per second.

So it is close but the i7 is faster. And this machine is not overclocked. You can crank the clock rate on a typical i7 a lot higher than I have it.

--Jon

i7-4930K@4GHz, 6 cores, (12G hash) : 6.7M nodes per second.

Code: Select all

Arasanx-64-popcnt:

 35	00:51	 338 118k	6 512k	+0,64	1.Rh5 Nf6 2.h7 Nxh7 3.Rxh7 Kd3 4.Kg4 Kxc3 5.Kf5 Rc6 6.Rh3+ Kb4 7.Rf3 Rh6 8.Ne3 Ka3 9.Ke4 Rg6 10.Nc4+ Kb4 11.Ne5 Rd6 12.Kf5 Rh6 13.Rg3 Ka3 14.Nc4+ Kb4 15.Ne3 Rh5+ 16.Kg6 Rh2 17.Nd5+ Ka3 18.Nxc7 Rb2 19.Nb5+ Kb4 20.Nd6 Rxb3 21.Rg4+ Ka3 22.Kf7 Rf3+ 23.Ke6
 36	01:04	 435 370k	6 735k	+0,64	1.Rh5 Nf6 2.h7 Nxh7 3.Rxh7 Kd3 4.Kg4 Kxc3 5.Kf5 Rc6 6.Rh3+ Kb4 7.Rf3 Rh6 8.Ne3 Ka3 9.Ke4 Rg6 10.Nc4+ Kb4 11.Ne5 Rd6 12.Kf5 Rh6 13.Rg3 Ka3 14.Nc4+ Kb4 15.Ne3 Ka3 16.Nd5 Rc6 17.Ke5 Kb2 18.Rd3 Ka2 19.Nf4 Rd6 20.Rxd6 cxd6+ 21.Kxd6

[tpoppins]

I happen to have both machines you mentioned. Plus a bunch of others.

Arasan on this position with 60 seconds/move:

4n3/2p5/1p2r2P/p1p2R2/P1N1k3/1PP4K/8/8 w - - bm Re5+; id "arasan19.43"; c0 "Anand-Landa, Bundesliga 0506 2005";

i7-6900k, 8 cores (12G hash): 6.7M nodes per second.

dual e5-2670, 16 cores (12G hash): 5.27M nodes per second.

Something is wrong with either these numbers or Arasan's scalability. On that position after one minute with Arasan 19.1 POPCNT (12 GB hasj/TBs off) I get between 5.0 and 5.5 MN/s on my dual X5670 (12 cores and threads), which I understand you have as well among the bunch of others.

[matthewlai]

I happen to have both machines you mentioned. Plus a bunch of others.

Arasan on this position with 60 seconds/move:

4n3/2p5/1p2r2P/p1p2R2/P1N1k3/1PP4K/8/8 w - - bm Re5+; id "arasan19.43"; c0 "Anand-Landa, Bundesliga 0506 2005";

i7-6900k, 8 cores (12G hash): 6.7M nodes per second.

dual e5-2670, 16 cores (12G hash): 5.27M nodes per second.

So it is close but the i7 is faster. And this machine is not overclocked. You can crank the clock rate on a typical i7 a lot higher than I have it.

--Jon

I only get 800-1000% CPU utilization with Arasan using "cores 16", so if that's enough to get 5.27M vs 6.7M, with perfect scaling, the 2670s should be slightly faster.

I'm sure the i7 is faster for chess because of scaling, but still, that's not a lot of extra speed for almost 3x the price. You can overclock it, but overclocking a chip with a TDP of 140W at stock will require some very serious cooling, and that's not going to be cheap either. A low end all-in-one water cooler is not going to be enough. It will probably be another $200-$300 in cooling (custom loop).

It's much more power efficient (per core-Hz) to have more and slower cores than fewer and faster cores, because power consumption is proportional to core count, proportional to frequency, and proportional to voltage squared. At equal voltage they would be about even, but faster cores require higher voltage.