Author: Robert Hyatt
Date: 15:04:58 07/26/00
Go up one level in this thread
On July 26, 2000 at 17:22:38, Tom Kerrigan wrote: >On July 26, 2000 at 14:55:04, Robert Hyatt wrote: > >>My argument won't "go south". 480 processors is the right number. 2.2M nps >>per processor (average) is the right number. 480 * 2.2M is > 1B. Which is >>a right number. 200M is his 'effective number' which no other parallel >>program is reporting, they _all_ report the total nodes / total time, to >>get a raw NPS. Hsu's RAW nps is around 700M, based on his statement in several >>papers that he can drive the chess chips at about 70% duty cycle using the >>current SP2 configuration. Or factoring in his 30% efficiency, he claims 200M >>which is a very realistic (and conservative when compared with every other >>parallel program's NPS) number. >> >>Unlike yourself, I try to not manufacture data, nor twist it to suit my own >>purposes. I simply report _factual_ numbers for DB, numbers you could easily >>verify _if_ you really wanted to. But that would take the fun out of the >>constant arguing, wouldn't it??? > >Okay, let's verify. I just went to IEEE's web site and searched through the >abstracts of all the DB articles. This is where you suggested to look, right? >Here are some direct quotes: > >"Because Deep Blue can evaluate 200 million moves per second, one observer said >it had won by packing 380 years of human thought into three minutes." > >"The IBM Deep Blue supercomputer that defeated World Chess Champion Garry >Kasparov in the 1997 historic match had 480 custom chess chips in the system." > >"At 2 to 2.5 million chess positions per second, one chess chip is equivalent to >a 100 billion instructions/sec supercomputer." > >"Now it has the ability to calculate 50 to 100 billion moves within three >minutes." (Deeper Blue) > >"Deep Blue is a 32-node IBM Power Parallel SP2 high performance computer. Each >node of the SP2 employs a single microchannel card containing eight dedicated >VLSI chess processors, for a total of 256 processors working in tandem." (seems >to imply Deeper Blue) That is mixed up of course. The SP2 was DB2, the original SP was used for DB1. So 1/2 of the above (SP2) is from DB, the other half (256 nodes) was from DB1. Quite a mix of numbers... 1/2 right... > >So there seems to be some confusion about how many processors were in Deeper >Blue, but one article says 480, so you may be right. That is the only number I have seen. Note that I am referring _only_ to articles written by Hsu/Campbell, such as the IEEE article on DB2. Not the fluff written by others such as newspapers and so forth that come from places like the IBM web site that is obviously in disrepair.. > >But the point of main contention here, namely 1 billion nodes per second, is not >supported anywhere. Certainly it is. Reread the above text that you quoted.. " >"The IBM Deep Blue supercomputer that defeated World Chess Champion Garry >Kasparov in the 1997 historic match had 480 custom chess chips in the system." > >"At 2 to 2.5 million chess positions per second, one chess chip is equivalent to > If you take 2 to 2.5 (actually 2 to 2.4 according to Hsu's numbers) and multiply by 480 chips, what do you get? an average of 2.2M nodes per second per chip, multiplied by 480 chips... ??? > >In fact, it's directly contradicted by two articles. One article says 200M NPS. >The other article says 277M to 555M NPS. Even the IBM web site says 100M, and >you'd think it would have the most optimistic number. It has the DB1 number, for reasons unknown. IE it appears (to me) that perhaps someone started to update it, but never finished, leaving old and new numbers mixed up horribly. > >So who is manufacturing data? It sure doesn't look like Chris. Does it look like me? is 480 * 2.2M pretty close to 1B (remember that I said this was the theoretical _peak_ NPS for DB. PEAK. Not typical. I then gave the 70% duty cycle which drops this to 700M, and then the 30% efficiency mentioned in Hsu's thesis, which drops it to 200M Did _I_ manufacture anything? Please point out what... > >Remember not long ago when I was quoting Hsu's estimate of how many general >purpose instructions it would take to search a DB node? And you told me that Hsu >obviously didn't know what he was talking about and the estimate was worthless? >Well, the estimate was published in the IEEE journal by the man who built the >chip. It was staring you right in the face, and against all common sense you >chose to ignore it. So who is being academic here? It doesn't look like you. I said that it is very hard to decide how many GP instructions it would take to emulate the DB hardware. As the GP instruction set for the pentiums has changed significantly... > >-Tom
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.