hardware

[fern]

I wonder if you can organize a kind of SETI colaboration with hundreds or at least dozens of fans.

Fern

[mwyoung]

Your current box is dual 2687w which is 16 cores at 3.1 GHz base and 3.4 GHz turbo.
The new E5-2697 v2 will be 24 cores at 2.7 GHz base and 3 GHz turbo.

Carl, so with HT off we have
dual 2687w v2 16 cores at 3.4 for 54.4gig
dual 2697w v2 24 cores at 3 gig for 72gig

I wonder how much faster the dual 2697 v2 will be using a chess program compared to the dual 2687w v2?

thanks
kgburcham

It will be close, I have a felling that the 2687 v2 will be faster not in NPS, but because you will hit about the same NPS, but doing it with less cores. 8 vs 12 or 16 vs 24 in a dual configuration. Once you start running with less cores or single core with the 2687 v2. It is no contest at all, the 2687 v2 is better.

That is why I will must likely buy the 2687 v2 , it is more flexible in a work station because it can turbo to 4 Ghz.

Anything that scales very well, the 2697 v2 is the chip to go with.

I have only seen a few benchmarks, I saw a Fritz benchmark for the 2697 v2 using 12 cores, it hit 24,400,000+ NPS in Fritz benchmark test. It did not give the relative speed only NPS.

[mwyoung]

Your current box is dual 2687w which is 16 cores at 3.1 GHz base and 3.4 GHz turbo.
The new E5-2697 v2 will be 24 cores at 2.7 GHz base and 3 GHz turbo.

Carl, so with HT off we have
dual 2687w v2 16 cores at 3.4 for 54.4gig
dual 2697w v2 24 cores at 3 gig for 72gig

I wonder how much faster the dual 2697 v2 will be using a chess program compared to the dual 2687w v2?

thanks
kgburcham

Here is a fritz benchmark for the 2697 v2 vs the slower 2687W v1. I don't know how well the Fritz benchmark works at prediction the better processor for chess engines. But if this is correct the faster 2687 v2 should be as fast or faster. Since the 2687w v1 is only 1105 KNS slower, the 2687w v2 should more then make up that ground. If the v2 version is only 5% faster it should hit around 24115 KNs


But more testing will be publish to resolve this in the future.

[bob]

Bob, how about cache issues? Are there any?
Running 31 concurrent matches on a 32 core box means 62 instances of an engine is loaded. No problem?

I'm asking because on my 8-core box (AMD FX-8120 Bulldozer) I can only reliably run 4 concurrent games. But maybe this is due to Bulldozer's special architecture?

That's why I prefer a cluster-type testbed rather than a single box with a bunch of cores. Less cache sharing. A big multi-core box is best for a single program, obviously...

[bob]

I am not a chess programmer, but it seems to me it is not about who has more cores, but do you have enough cores to run your tests and ideas.

Even if we had 1000 cores it wouldn't be enough. We have more ideas than we can test reasonably and in fact we have to take shortcuts. Once shortcut we take is to reject an idea that doesn't immediately start hot. For example if we are down 5 ELO after 1000 games and have 10 more ideas in the queue we may simply move on to the next test and take a big chance on rejecting a good idea. But 5 ELO is way under the error margin and the idea still has an excellent chance to succeed.

Is Mr. Mansoor planning to hire also people and programmers to help with ideas for Houdini?

Will a two man team need this that much CPU power that Stockfish uses to test?

Stockfish:
Active - 34 machines 191 cores 1.29M nps (245.80M total nps) 346 games/minute

Testers 157
Developers 21
Active testers 22
Tests submitted 1454
Games played 26875732
CPU time 43.02 years

For the record, I think that's a bad idea. About half the time a new idea should "start off hot" but drop. The other half it will start "cold" and rise...

[Werewolf]

Your current box is dual 2687w which is 16 cores at 3.1 GHz base and 3.4 GHz turbo.
The new E5-2697 v2 will be 24 cores at 2.7 GHz base and 3 GHz turbo.

Carl, so with HT off we have
dual 2687w v2 16 cores at 3.4 for 54.4gig
dual 2697w v2 24 cores at 3 gig for 72gig

I wonder how much faster the dual 2697 v2 will be using a chess program compared to the dual 2687w v2?

thanks
kgburcham

The million dollar question.
The formula being quoted a lot (originates from the Rybka team?) is (no. Cores ^0.76) x clock speed x processor efficiency.

If this is correct then with the E5-2697 v2 we have:
(24^0.76) x 3 x 1 = 33.56
Which is higher than any other processor configuration.

My problem is, according to this, the E5-2687W (16 core setup @ 3.4 ghz) and 2690x (12 cores @ 4.2 ghz) ought to be about the same speed.

Perhaps for Rybka they are. But with Houdini my 12 core is noticeably faster at solving test suites etc.

[Vinvin]

A nice Cinebench 11.5 record (score : 32.50) on a Xeon E5-2697 v2 :
https://www.youtube.com/watch?v=T7Nr1tWT8Vo

[bob]

That is trivial to keep busy. My current cluster has 70 nodes, 8 cores per node, and I can keep it busy 24/7.

I run 30K game matches in under an hour on that box.

What is the TC you use? And what is the average nodes per second of each core ?

1. I use several. Fastest is 10sec +.1 sec increment. 60s+1s is also pretty common.

2. Primary box I use has 70 nodes, each node has 8 cores, crafty nps on one core = 3M using "bench" command.

To get the 30K in well under an hour, I use 20s +.1s. I don''t use that if I am testing search changes, however...

[mcostalba]

That is trivial to keep busy. My current cluster has 70 nodes, 8 cores per node, and I can keep it busy 24/7.

I run 30K game matches in under an hour on that box.

What is the TC you use? And what is the average nodes per second of each core ?

1. I use several. Fastest is 10sec +.1 sec increment. 60s+1s is also pretty common.

2. Primary box I use has 70 nodes, each node has 8 cores, crafty nps on one core = 3M using "bench" command.

To get the 30K in well under an hour, I use 20s +.1s. I don''t use that if I am testing search changes, however...

3Mnps at 20s+.1s is about 3M*2*(20+0.1*40) = 144 M nodes per game on average

This is a good quality game. Our std TC are 15+0.05 and 60+0.05 at normalized CPU speed of 1.4M nps so we have

47 M nodes per games (at 15 secs)
173 M nodes per games (at 60 secs)

This nodes per games figure IMHO is a good metric of overall game quality.

[Edsel Apostol]

That is trivial to keep busy. My current cluster has 70 nodes, 8 cores per node, and I can keep it busy 24/7.

I run 30K game matches in under an hour on that box.

What is the TC you use? And what is the average nodes per second of each core ?

1. I use several. Fastest is 10sec +.1 sec increment. 60s+1s is also pretty common.

2. Primary box I use has 70 nodes, each node has 8 cores, crafty nps on one core = 3M using "bench" command.

To get the 30K in well under an hour, I use 20s +.1s. I don''t use that if I am testing search changes, however...

3Mnps at 20s+.1s is about 3M*2*(20+0.1*40) = 144 M nodes per game on average

This is a good quality game. Our std TC are 15+0.05 and 60+0.05 at normalized CPU speed of 1.4M nps so we have

47 M nodes per games (at 15 secs)
173 M nodes per games (at 60 secs)

This nodes per games figure IMHO is a good metric of overall game quality.

Crafty NPS is different than SF NPS though. You should try Crafty bench on your testing framework to really compare accurately.