Author: Robert Hyatt
Date: 22:00:23 01/27/00
Go up one level in this thread
On January 27, 2000 at 22:38:40, Tom Kerrigan wrote: >On January 27, 2000 at 21:14:25, Robert Hyatt wrote: > >>No... what I said was that I knew what Cray Blitz did internally, and how it >>worked. And I was _still_ surprised by how much slower the PC would execute > >Okay, if your only point was that you were surprised going from A to B, then >that's fine. > >The reason for my "confusion" is a comment that you made a few posts ago: > >>The point is on "what" architecture? 40K on a sparc == 20K on an X86. > >And then you begin talking about how much you can do with Cray instructions. I'm >sure you can see why I might think your point was something else. Perhaps just a bad explanation on my part. The point I was trying to make was that even though I know the cray as well as anyone, and I know the X86 pretty well, it is _very_ hard to say that a program on the cray translates into N instructions on the X86. Somethings that are simply trivial to do on a good vector architecture suddenly become hairy on a scalar architecture.. > >>What is the difference between 40K and 400K instructions? Significant enough > >I remember that MChess was running at ~2,500 NPS on a 133MHz Pentium in 1995. >133M / 40k = 3,300 3-4K seems to be a good number. Frans has this down to 2K. A program like Hiarcs seems to be at 10K or so, although it is impossible to say what he is doing, lots of eval or lots of computation trying to figure out which parts of the tree to extend further... >Hsu's estimate implies that DB would run a little faster than MChess. If the >estimate is off by a factor of 2 either way, I don't think it makes a huge >difference. But if the estimate is "optimistic" by a factor of 10, that means DB >would search 250 NPS, which I don't consider reasonable for a chess program. > >>>It doesn't matter HOW he thought he could compute the term, only only matters >>>that it's POSSIBLE. >>But it _does_ matter. Otherwise how do you do the computation? And are you >>thinking in terms of what is easy in hardware? Have you ever written a piece >>of code, thinking when you started that it would be 2K lines of code, but ended >>up 20K? I would never have guessed that Crafty would end up at nearly 50K and >>still growing. > >But that argument goes both ways. I doubt that he would consistently >underestimate how much work it takes to evaluate every single term. > If he were guessing, he would likely underestimate, rather than having to retract a bigger guess later. But again I am guessing about him making a guess. But I am fairly sure that the 40K was not a number that had hours of thought behind it... I would hate to guess how many instructions my eval turns into, and I wrote it... >>Anything is possible. I simply take that "number" as a "SWAG" to make a point >>that a _lot_ of things are going on inside DB's processor. If _I_ had written >>that article, I would probably have guessed "low" to be conservative, since >>everyone seems to want to jump all over any significant claims he makes... > >But there is another opposite argument here: he wants to make his chip seem as >cool as possible. That isn't Hsu. You have to know him to see how this doesn't fit. At an ACM event, sitting around a table talking, you would _never_ guess who he was unless you knew or asked. He is a long way from a Berliner, for example, if you read the Fredkin announcement about DT winning the intermediate prize... Hitech was mentioned more than once, for no reason. Hsu is way more soft-spoken. > >Until we know more, I don't see why we can't just assume that Hsu is right. >You're coming up with every reason for his estimate to be low, but I could be >arguing that his estimate is high just as aggressively. The point is, neither of >us seems to know anything besides the actual estimate, so I don't see why we >can't just take it on faith for now. > >-Tom OK.. assume 40K is _the_ answer for an X86 program to search one node. What do we know? Eval instructions? search instructions? IE we _still_ don't know what that means.
This page took 0.01 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.