Author: Robert Hyatt
Date: 14:37:28 01/27/00
Go up one level in this thread
On January 27, 2000 at 16:26:46, Tom Kerrigan wrote: >On January 27, 2000 at 13:17:32, Robert Hyatt wrote: > >>The point is on "what" architecture? 40K on a sparc == 20K on an X86. > >Not really. You have been arguing for a long time that the DB evaluation >function could not be implemented on a PC simply because it is too expensive. My >goal is simply to "prove" that it's not too expensive, and by extension, some PC >programs probably have equal (possibly better) functions. This "proof" doesn't >require the exact number of instructions. FHH's estimate could be off by a >factor of 2, and the function would still not be too expensive. OK. We can't really analyze deep blue. But I can analyze a program I wrote years ago. Cray Blitz. We did a _lot_ of clever programming using hardware that was unique to the cray, from vector merges, to vector selection, to vector reduction, to gather/scatter memory references, etc. On a Cray C90 single CPU, running at effectively 250mhz, we search about 150K nodes per second. I am currently fiddling with the Portland Compiler Group's FORTRAN compiler as a physicist wants to use it on our 8-way cluster. I tried this last week just to see how it would run as that was one of the more complex FORTRAN programs I had access to. Amazingly, it ran at just over 1K nodes per second on my 400mhz xeon. So scale that to about 600 nodes per second at 250mhz. The cray is a GP machine. The Xeon is a GP machine. At equal clock periods, the Cray is roughly 200 times faster. If you had asked me beforehand, I would not have guessed that... The point here is that my CB eval is _absolutely_ non-viable for a PC. Whether it could be written more efficiently is another issue. It would be a 100% re- design to do so. A good comparison: how many instructions would it take to write a simular for a reasonable CPU chip like the 386? That ought to be fairly close to the number of instructions to do the same things that the DB chip does... > >>Even more important, is deriving that 40K estimate. It would take a _lot_ >>of thought to come up with a real number, because hardware design doesn't >>translate to "N instructions" trivially. IE I suspect that the number 40K >>is just a big number that was used to illustrate how much stuff DB is doing >>in the hardware chips using parallel circuits. > >I think this theory is deeply insulting to FHH. He has published his estimate in >a well-regarded and widely-read journal, and he didn't give any HINT of a >warning that the estimate has any error at all. By saying that 40k "is just a >big number" you are attacking his professional integrity. Not at all. I don't think "the number" is important. I think the "bigness" is... > >Let's say you're trying to sell your house. You know that your house is pretty >big, so you place an advertisement in the newspaper saying that it's 8 million >square feet. Somebody comes to look at the house and points out that it's >nowhere near 8 million square feet. Would you say, "Oh, yeah, it actually takes >some effort to measure the size of a house, so I just figured I would convey the >impression that it's big." No, you wouldn't say that, because it's lying. And >neither would Hsu. Not the same thing. That is directly computable. Hsu would have to actually _write_ the code to know precisely how many instructions are required... and I know he didn't do that from his book. > >I think it would actually be fairly easy to come up with an estimate like this. >You just go through all the terms you have, and imagine how you would implement >them in software. Guess at how many instructions it would take and add >everything up. For example, if you know that finding a doubled pawn takes about >5 instructions, and you think that you could find an isolated pawn the same way, >you can guess that it's another 5 instructions. And so on. > >-Tom To an extent. But I can answer that with one AND. How many for a non-bitboard program? Wouldn't the non-bitboarder and I come up with _different_ estimates for the same thing. And wouldn't we _both_ be right? And wrong?
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.