Not logged inRybka Chess Community Forum
Forum RybkaChess Help Search Register Login
Up Topic The Rybka Lounge / Computer Chess / My 32 core machine.
1 2 3 Previous Next  
Parent - - By goon12 (*) [au] Date 2012-02-28 12:19
after patch getting same performance
Parent - By Werewolf (*****) [gb] Date 2012-02-28 12:26
I think the main issue chess players have with Bulldozer is its inefficiency. All your 32 cores are at 100%, yet they are not all performing equally - the 2nd core will be slower than the first per module.

Your final np/s, although high, isn't THAT high and with 32 cores adds massive inefficiency. Basically to maintain equality you have to get at least 1.3x (Robert H says 1.2 with Houdini) the np/s per doubling of cores.
So a Quad @ 10,000 kns = an 8 core at 12,000 kns.

However, it'd be interesting to run your machine under IDeA...
Parent - By Ray (****) Date 2012-02-28 13:39
I could be wrong, but the patch is only intended to adress light workloads, where not all the cores are busy, and is about making sure that the core use is optimal under that situation. When all cores are fuilly loaded, I thnk the patch probably has no effect.
Parent - - By Stonehenge (***) Date 2012-02-28 16:47

> after patch getting same performance


Can you post the results of Houdini's "autotune" command for 8, 16 and 32 threads?
See the Houdini manual for "Split Depth": http://www.cruxis.com/chess/manual/cores_and_threads_management.htm
Parent - By goon12 (*) [au] Date 2012-02-29 04:38
yeah i already tried it and it freezes after about 90 secs
Parent - By goon12 (*) [au] Date 2012-02-29 04:39
I'll try again
Parent - By goon12 (*) [au] Date 2012-02-29 04:42
it's probably because of ram, I'll try with 128mb of hash
Parent - By goon12 (*) [au] Date 2012-02-29 05:57
I'll get this to work, I've noticed build up time takes time, but once it gets going it's fine. I'm installing a new Version of windows to support all available ram and installing updates, I'll tell you how I go
Parent - By goon12 (*) [au] Date 2012-02-29 10:02
oki This is at 32 cores and 32 gig. All patched and tuned for large pages.


image hosting gif
Parent - By goon12 (*) [au] Date 2012-02-29 10:05
I'll do 16 cores shortly and 8 cores shortly
Parent - By goon12 (*) [au] Date 2012-02-29 10:22
16 cores and 32 gig.


upload images
Parent - - By goon12 (*) [au] Date 2012-02-29 10:39
8 cores and 32 gig.


adult image hosting
Parent - - By Stonehenge (***) Date 2012-02-29 14:30
Thank you, interesting results.
With the higher number of cores the best Split Depth appears to be 14.
From 16 cores to 32 cores the speed increases from 19,800 to 25,500 or by about 29% (close to the 30% value I mentioned earlier).
If you take into account the 20% efficiency loss when doubling the number of threads, the 16 cores will provide nearly the same strength as the full 32 cores.
Parent - - By goon12 (*) [au] Date 2012-03-01 01:55 Edited 2012-03-01 01:58
so it's a 3% gain going from 16 cores to 32 cores from these results. anything above 25% speed after doubling cores will give you a gain in performance.

19800 X 1.25 = 24750

24750 x 0.80 = 19800
Parent - - By goon12 (*) [au] Date 2012-03-01 02:08
Another thing I noticed was going from 8 cores to 16 was a 82% gain in speed while going from 16 to 32 was only 29%
Parent - - By Bouddha (****) [ch] Date 2012-03-01 05:27
So 16 cores is currently the optimum ?
Parent - - By goon12 (*) [au] Date 2012-03-01 07:54
depending on system, but for my system 32 cores will give me a 3% gain over 16. It seems as though houdini works good with 16. For IDeA 32 cores is amazing but for infinite analysis it's not the best.

My next system will be 2x xeon 2687. I'm hoping they can be overclocked. We'll just have to wait and see
Parent - By TheHug (Bronze) [us] Date 2012-03-01 07:59
It much be nice to have the money to buy though kind of toys. :grin:

Enjoy them, I bet they are a lot of fun. :smile:
Parent - By The Truth (**) [de] Date 2012-03-04 00:52
How many kns you get on a single core? I guess this is also analogous with bulldozer by amd.
Parent - - By Lukas Cimiotti (Bronze) [de] Date 2012-03-04 10:03

>If you take into account the 20% efficiency loss when doubling the number of threads


So Houdini's mp speedup is 1.6 per doubling of cores?
Parent - - By Werewolf (*****) [gb] Date 2012-03-04 12:57
Interestingly Bob H thought most programs lose 30% per doubling, only giving a 1.4x speedup!! I wonder if the figures they mean are 1.8x speed up (Houdini) and 1.7x speedup (Crafty & others)..?
Parent - By Lukas Cimiotti (Bronze) [de] Date 2012-03-04 13:06
Stonehenge's calculation only looks correct if it's really 1.6. That's why I asked him :wink:
- - By goon12 (*) [au] Date 2012-02-25 16:20
all parts arrived, will put together tomorrow.
Parent - - By Bouddha (****) [ch] Date 2012-02-25 16:45
great news !
Looking forward to see your results !

Are u planning to come over in plachess computer room ?
Parent - By David Evans (***) [gb] Date 2012-02-26 08:37
did you get my message Bouddha??
Parent - By GM (*) [is] Date 2012-03-05 14:06
Hi!

I need specs that I can send to computer stores and ask for an offer for 32 and 16 cores computer which will be used only for chess.
Have been thinking about Xeon or the 6 core i7 3930 (rather than the 4 core new i7 3820).
Would prefer quite a lot of RAM, maybe 32 GB.
The beauti will be running engines and also from time to time some database program.
Can you help me guys and post a detailed list?

Regards,
GM
Up Topic The Rybka Lounge / Computer Chess / My 32 core machine.
1 2 3 Previous Next  

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill