[computer-go] December KGS Computer Go tournament: full boards, fast

[terry mcintyre]

From: Nick Wedd <nick at maproom.co.uk>


In message <20071127143430.GS14890 at machine.or.cz>, Petr Baudis 
<pasky at ucw.cz> writes

>>  are there any limits (set by either rules or ethiquette) on power of
>>the machines running the bots? Or noone cares? I wonder if it's ok to
>>use a 16-core opteron-packed machine to run the bot or something of
 the
>>scale of a reasonable modern desktop would be more appropriate.

>The rules say, there are no limits.  Etiquette says that, if you do use
 
>a 16-core opteron-packed machine or whatever, you should tell me, so 
>that I can mention it in my report on the event.

>This used not to be a significant issue.  Before the advent of UCT, 
>doubling or even 32-folding the processor power did not produce a big 
>increase in playing strength.  Now it does, and so I should perhaps 
>reconsider this.

Deep Blue, which beat Kasparov, was a purpose-built supercomputer. I
think we should be delighted that computer Go programs now make use of
some heavy iron. If the first shodan-strength Go computer requires a sheikh's ransom, average Joes will
purchase the same capability a decade later. In 2007, almost all computers are dual-core; quads are quite affordable; the 8-way Mac Pro is "a reasonable modern desktop" for some; a 16-way rackmount server is about $7-8k, not at all excessive for a research project. The SC072 is available for $15k. Following Moore's Law, these prices will plummet further. 16 cores may be "a reasonable modern desktop" in 2008, certainly by 2009.

I can understand some concern about a "level playing field", but there are likely to be a lot of 8-way competitors this year, and 16-way the next, and so forth - you'd have to keep raising the bar every year. This sort of hardware is on hand for many go enthusiasts. Larger clusters should be welcomed. If Sicortex were to enter a program optimized for their SC5832, that would surely spur competitors to eliminate wasted cycles and improve search efficiency! The power of the SC5832 is about 8 doublings from the desktops of well-heeled hobbyists - perhaps 12 years away. 

It would be interesting to test several conjectures. Do current algorithms scale to many (hundreds or thousands of cores)? Do elo ratings scale with the number of cores? The present course of computer hardware evolution suggests that manycore computers are our future; the answers to such questions will drive the programs of tomorrow.

 




      ____________________________________________________________________________________
Be a better sports nut!  Let your teams follow you 
with Yahoo Mobile. Try it now.  http://mobile.yahoo.com/sports;_ylt=At9_qDKvtAbMuh1G1SQtBI7ntAcJ
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://computer-go.org/pipermail/computer-go/attachments/20071127/8d3ddc20/attachment.htm