New chess computer

[beachknight]

I am assisting my chess friend to build a desktop computer for
chess operations. We have gotten 2 configs so far and I would
like to hear your opinions, esp. about fritz bench marks or
houdini bench marks plus mainboards etc. Or a third alternative...

first one is i7 3960x (4 cores, 8 with hyperthreading I suppose) 3.5 GHz
second one is ci7 3930x extreme (6 cores, 12 threads) 3.3 GHz

I am checking the specs of first one right now.

TIA,

PS: One more thing: What about SANDISK 240 gb extreme at 12k rpm?
I never heard it before.

[beachknight]

oops. first one is i7 3770K.

[Zhu.Jianzhao]

from my experiences on playchess.com's engine room, the average speed of houdini 3 running on i7 3930x is about 20000knps, which is much stronger than the houdini 3 running at 12 cores unknown cpu with average speed of 30000knps. i don't know why.., but normally the fact i found is that if getting the same knps, houdini 3 at 4 cores cpu seems to be stronger than the one running on eight cores.

[Houdini]

from my experiences on playchess.com's engine room, the average speed of houdini 3 running on i7 3930x is about 20000knps, which is much stronger than the houdini 3 running at 12 cores unknown cpu with average speed of 30000knps. i don't know why.., but normally the fact i found is that if getting the same knps, houdini 3 at 4 cores cpu seems to be stronger than the one running on eight cores.

Running more threads increases the inefficiency of the parallel alpha-beta search.
You need a 20% higher node speed to compensate for a doubling of the number of threads.
Example: a 6-core running at 20,000 kN/s will be about as strong as a 12-core running at 24,000 kN/s, or a 16-core running at 26,000 kN/s.

That's why an overclocked Core i7-3930K is such a great bargain, you get great performance at a reasonable price.

Robert

[beachknight]

from my experiences on playchess.com's engine room, the average speed of houdini 3 running on i7 3930x is about 20000knps, which is much stronger than the houdini 3 running at 12 cores unknown cpu with average speed of 30000knps. i don't know why.., but normally the fact i found is that if getting the same knps, houdini 3 at 4 cores cpu seems to be stronger than the one running on eight cores.

Running more threads increases the inefficiency of the parallel alpha-beta search.
You need a 20% higher node speed to compensate for a doubling of the number of threads.
Example: a 6-core running at 20,000 kN/s will be about as strong as a 12-core running at 24,000 kN/s, or a 16-core running at 26,000 kN/s.

That's why an overclocked Core i7-3930K is such a great bargain, you get great performance at a reasonable price.

Robert

i7-3930K 3.2 GHz seems to be very good alternative
instead of i7-3960X 3.3 GHz extreme. Also without overclocking.
I'd like to know what would be the Houdini 3 figures on i7-3930K
3.2 GHz. I have checked few pages. I could not find an appropriate info.

Best,

[beachknight]

I found this page:

http://www.evga.com/forums/tm.aspx?m=15 ... print=true

[Zhu.Jianzhao]

Good! Thanks for your instructions

[beachknight]

Good! Thanks for your instructions

Thanks Zhu. I am still looking for advices on other hardware parts,
esp. mainboards and cooling parts plus knps figures of H3 (or H15).
I suppose H3 on this config (without OC) makes 16 mnps mostly.
Best,

[shrapnel]

[quote="]

Running more threads increases the inefficiency of the parallel alpha-beta search.
You need a 20% higher node speed to compensate for a doubling of the number of threads.
Example: a 6-core running at 20,000 kN/s will be about as strong as a 12-core running at 24,000 kN/s, or a 16-core running at 26,000 kN/s.

That's why an overclocked Core i7-3930K is such a great bargain, you get great performance at a reasonable price.

Robert[/quote]
Hmm.... interesting figures. Do you mind explaining how you arrived at those exact figures ?
Also, you seem to imply that increasing number of cores actually increases the inefficiency of the Engine, irrespective of the fact whether the threads being considered are actual physical cores or hyper-threaded ones !.
I'm running an i7 3930 k o/c to 4.7 Ghz btw.
Also, another reason your statement interests me, is that I've being running some experiments with HT actually turned ON ( contrary to popular wisdom) on my i7 3930 k and have found some interesting results in my Engine-Engine matches against online opponents.
I have found that utilising some of the hyper-threads ( not necessarily all ) and changing some of the settings in the default Houdini 3 ( which I'm not prepared to reveal at present ), I'm getting excellent results against strong Engine players.
Hyper-threading is in NO WAY reducing the strength of Houdini 3, provided I reach a much higher kn/s speed than the average 20,000 kn/s of the i7 3930k.
Yes, if only a slight increase in kn/s with HT enabled, yes, HT on doesn't give any better results.
But. if enough hyper-threads are enabled and kn/s markedly improves, the Engine DEFINITELY becomes stronger !
This is where your figures interest me, because I only have a rough estimate of the increase in kn/s required to get more power out of the engine.
Also, read somewhere that Window 7 64 bit and Windows 8 even more so, handle HT much better.
So, I think its time that (sacred) truth " HT is bad for chess" needs to be looked into afresh !

[Laszlo Gaspar]

Hi all!

My opinion:

Don't buy a hyperthreading CPU for premium price for chess!

My calculations:

When you switch on HT, the OS sceduler balances the thread loads between the virtual processors. You will see 5/8 speed on every virtual processor, because the hyperthread has only 1/4 speed of the real core and it is averaged. You have now doubled cores on 5/8 speed.
5/8 speed means that 8/5 (=1.6) speedup would reach the original, so as every % means around 1 ELO difference, the virtual processors has 60 ELO less . But you have doubled cores, this adds +40ELO. And finally you have a -20ELO on the HT enabled machine which asks for a 20% speedup to be on par with the non-HT machine.
So Robert is right until the HT technology is not better and paralel scaling is not better. Worth to mention that Deep Junior scales very well, reaching +70ELO with thread doubling on the CCRL list, and the theoretical maximum is around 100-120ELO.

To go back to the original question the 3770K with less cores is around 20 ELO weaker but it is much-much cheaper, so it is better, but even better is a simple 4 core overclockable CPU without HT as you would not use it for chess and even much cheaper (3570K?).