Author: Robert Hyatt
Date: 10:02:51 06/27/98
Go up one level in this thread
On June 27, 1998 at 10:50:38, Don Dailey wrote: >On June 27, 1998 at 02:08:03, Ed Schröder wrote: > >>>Posted by Keith Ian Price on June 26, 1998 at 21:40:50: >> >>>Hello, Roberto, >> >>>When I talked to Hsu about this at the end of April, he said that IBM had no >>>intention of selling the chess processors as a PC product, but that he was >>>negotiating with them for the rights to the chips (1997 version), so that he >>>could possibly market such a product. There's a lot of "ifs" there, though. I >>>asked him some hypotheticals, such as: If there were a market for 1 million of >>>these boards, at what price point do you think you could sell them? He answered >>>$200. With a market of 10,000 that would double to $400. I agree with you that >>>this would be a great thing to experiment with, and I believe even Vincent would >>>shell out $400 real quick to test it. Hsu also said that his short eval took >>>only one cycle, and his long eval took 8 cycles. Move generation took 4 cycles. >>>The long eval was necessary in only 20% of the cases. >> >>It's my believe if things are done in the right way (have some ideas) Hsu >>can easily sell 10,000 pieces, more likely 25,000 pieces if not 50,000. >> >>I do not sell competitive software myself but I surely will make an >>exception if Hsu manage to enter the market with his chip. Everybody >>should simply have it and the better sold the lower the end-user-price! >> >>- Ed - >> >> >>>kp > >But my question would be, what are you actually getting? Are you >getting a chip that does n ply searches with deep blue's evaluation >function and little control over the search, extensions and evaluation >function. Maybe it would make it possible to have a bunch of super >strong programs but would they all be deep blue clones? That doesn't >sound very exciting to me. > >- Don remember that there are really three "pieces" of the puzzle in DB: 1. the host software search.. here is where they do singular extensions, for example, and where *we* would probably do null move r=2. SO that part of the search would be *hugely* different for DB vs US, but we would still go fast as hell with the hardware. 2. the chss processors are pretty well "cast into stone" and can search N nodes per second each (N=2.4 M at present, but I'd bet Hsu could make that 10X faster using todays .25 micron fab). So we'd get a plain-vanilla N ply search out of the hardware, plus the capture search, at a fantastic speed. 3. the downloadable parameters that affect how the chess chip evaluates the positions. Here we could vary wildly from them, within the framework of what their recognizers can extract from a position. But the weights and interactions would be of our choosing. So we wouldn't be DB "clones" exactly, but we'd all be pretty well tactically similar, speed-wise, but with different basic search algorithms and evaluation weights... which is pretty much like it is right now when you think about it. You could even cop-out and not use their eval, just use the chip to maintain the position, generate moves, and so forth, although it would be a shame to give up the hardware alpha/beta search. the hardest part would be portability across widely varying platforms, because your host processor can search N nodes per second (lets say 300K for Crafty on the 4 cpu ALR). Their chess processor can search 2.4M nodes per second. TO keep my ALR busy, *and* keep the chess processors busy, there is exactly *one* way to divide the tree up. You find the point, somewhere between the root and the tips, where 10% of the nodes are on the root-side, and 90% of the nodes are on the tip-side, and split the work there... then your host can search and barely keep the chess processor fed. But when you change the host, you change this "split point". That would be a pain trying to be prepared to answer the question "where do I set the host split point on my AMD K6-2/266, overclocked to 366 with a 112mhz bus speed?"
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.