Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Hydr's KN/s

Author: Vincent Diepeveen

Date: 07:13:50 07/03/05

Go up one level in this thread


On July 03, 2005 at 08:02:49, Amir wrote:

>On July 03, 2005 at 07:21:55, Vincent Diepeveen wrote:
>
>>On July 03, 2005 at 06:57:42, Amir wrote:
>>
>>>On July 03, 2005 at 06:50:23, Vincent Diepeveen wrote:
>>>
>>>>On July 03, 2005 at 06:10:37, Amir wrote:
>>>>
>>>>>I have noticed that playchess.com software tells you the opponents KN/s after
>>>>>each game. Can anyone who has played Hydra (the new version -- Ares01) posts
>>>>>it's KN/s?
>>>>
>>>>It had about 220 million nps against us.
>>>>
>>>>Note that hardware nodes are not like software nodes.
>>>
>>>so 220 million nps is average for Hydra?
>>
>>it is hardware, so it gets in endgame not more nodes than it gets in opening.
>>arguably in endgame it gets less nps than opening because the treesize of 1
>>search of 2 to 3 ply is smaller.
>>
>>so yes 220 mln is average for 32 processors. please note i calculated theoretic
>>peek of it at 320 mln nps for 32 fpga chessprocessors. the machine has room for
>>64 fpga processors in total.
>>
>>1 fpga chessprocessor can get clocked at like 60-66Mhz easily.
>>
>>So I calculated peek at theoretic 60Mhz, that's very pessimistic taken.
>>Important runs they will clock it a tad higher to 66Mhz or so.
>>
>>So on paper the real theoretic peek is even higher in reality. More like 400 mln
>>nps, but you never get your peek in hardware of course.
>>
>>There is latency of pci bus, there is time needed to put all cpu's to work and
>>so on.
>>
>>By the way, my own experiments at supercomputers indicated that 16 cpu's is most
>>easy to get a good speedup from, relative seen. 32 cpu's it is a lot hardware,
>>but putting 32 cpu's to work is real easy. Yet the extra speedup you get from
>>moving from 16 to 32 cpu's is not so big. Not only speedup gets bigger, also the
>>worst case speedup still is utmost tiny.
>>
>>But the nodes per second out of 32 fpga processors being 220 million is very
>>constant. It hardly fluctuates. It's not that 1 run it's 50 million and the next
>>one 500 million. It's simply always around that 220 million.
>
>Since the cluster is 64 CPUs, what  would be the nps when they get 32 more FPGA
>cards?

the nodes per second is not so important. the speedup is. i expect not much
better speedup. but of course it will scale well when putting the hashtable
depth probe higher. so the nps will nearly increase lineairly.





This page took 0 seconds to execute

Last modified: Thu, 15 Apr 21 08:11:13 -0700

Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.