Author: Vincent Diepeveen
Date: 03:59:00 08/04/05
Go up one level in this thread
On August 04, 2005 at 02:37:20, Mark jones wrote: >Can you imagine how would Junior,shredder,fritz would have played if they where >deployed on A super computer like this: >http://www.top500.org/sublist/System.php?id=7605 > >If this were possible not only it would kill all the humans I think it would >have crushed Hydra to... >What do you thisk about it? And was there an attempt to deploy an pc program >on asuper computer? Diep ran on a 1024 processor supercomputer in 2003. That supercomputer entered the top500.org list as #6 or so around the year 2000. The biggest partition supports 512 processors and can run jobs from 500 processors. That partition was used for the benchmark at top500, therefore it shows up as 512 processors there where in reality it's 1024 processors. They changed that aspect though. Todays top500 supercomputers are more bandwidth oriented than supercomputers from a few years ago. That means that no chessprogram will ever work at any of the 'big' supercomputers mentionned at the top500.org. Just getting a single message to 32768 processors already takes half an hour. I hope you realize that doesn't work realtime very well. Those 512, say at most 2048 processor partitions (see the 10240 itanium2 NASA supercomputer), are about the maximum you could potentially use for a chessprogram. In theory of course everything is possible, in practice the realtime and quicker levels that get played in world champs just make it practically impossible to join with a supercomputer. With 3 minutes a move, a supercomputer can join. At 60 in 2 + 30 minutes all, that means that if a game lasts more than 60 moves, that the supercomputer is going to forfeit 100% sure. Please note hydra is not a supercomputer. It's a cheap pc cluster of just 32 nodes with 32 fpga cards. You can compare that very well to a quad opteron dual core 2.2Ghz. Practically same effective combined power it delivers. Please realize that a 512 processor 500Mhz supercomputer where diep run at, it is hard to get a good speedup. When having 3 minutes or more a move, i measured some peeks of 20% speedup here and there, but the reality is that many moves get played at faster limits. You will realize that not every move can eat 3 minutes. So the reality was that predicted moves it searched real deep, up to 20 ply, whereas unpredicted moves it reached 14 ply. First few rounds i had some scaling problems in software, i fixed that during the tournament. So effectively it's better to run on a quad opteron dual core. What most people do not realize is that supercomputers have very low clocked processors always. It's far easier in computerchess to have a single 2.2Ghz opteron processor, than it is to have a 500Mhz MIPS (washing machine) processor. The big IBM machines at this top500 list feature 700Mhz power processor. They deliver 2.8 gflop, but in branchy integer code they suck major ass, to say it VERY POLITE. So you need already a 100 processors or so just to get *near* the same speed that a quad opteron has directly. Because just do the math. Worst case it's about 5% speedup. 5% speedup from 700Mhz * n processors compare it with the quad opteron dual core which is 8 x 2.2ghz x 1.5 (higher ipc) ipc = instructions pro cycle At what N is the supercomputer similar to quad opteron using this worst case performance? Do you see the problem? The speed a single cpu of such a supercomputer delivers is the problem. In the past the supercomputers had faster processors than pc processors were. Now the opposite is the case. Then additional the only program that works at supercomputers is Diep. No other strong program scales to 100+ processors. Hydra will need years before they get it to work at 100+ processors. Money is not a problem there as you can imagine. Just the programmer is.
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.