Author: Robert Hyatt
Date: 10:02:09 01/23/00
Go up one level in this thread
On January 23, 2000 at 03:44:56, Vincent Diepeveen wrote: >On January 21, 2000 at 16:23:47, Tom Kerrigan wrote: > >>On January 21, 2000 at 15:10:32, Robert Hyatt wrote: >> >>>After creating "cray blitz" I found it difficult to think about trying to >>>write a program for a Micro. And it took a lot of time/effort to do so. After >>>building DB, it would take Hsu a lot of time/effort to try to write a program >>>for a Cray, and then more time/effort to think about a PC-based program. >> >>Well, presumably he wasn't an idiot and wrote the SP program in something >>portable. Then all he would have to do is write a quiescence search function and >>port the DB evaluation function. After designing the function in hardware, >>remaking it in software should seem positively trivial. He already knows all the >>terms anyway. How long could it possibly take? Definitely less than a week for a >>"first draft," I would guess. >> >>Here's another thing I was just thinking about. DB had a "fast eval" that took 3 >>cycles, and the full eval took something like 11 cycles. Most of the time, the >>fast eval was good enough. Presumably the 40,000 instructions that he reported >>was for the full-blown, 11 cycle eval. >> >>Here is my best guess at how fast a DB program would run on a PIII. Assume 75% >>of the time it takes 11k instructions to eval, and 25% of the time, it takes >>40k. So that's an average of 18k. Now, figuring that the PIII almost always >>retires 1.5 instructions per clock cycle, it takes 12k clock cycles per node. >>Now assume you're running at 800MHz. That's 66000 NPS, and it's still being >>fairly conservative (with the 75%). That strikes me as a perfectly reasonable >>speed; if I'm not mistaken, some strong micro programs run that fast on the same >>hardware. >> >>As for putting in the effort to make such a program, I think that's a no-brainer >>too. Imagine how much money he could make off of selling the DB program for PCs. >>A million people would want a copy, the first day it's announced. And it >>wouldn't even matter how strong it is. He could just write on the back of the >>box, "This program runs 3000 times faster on the official DB hardware!" and >>everybody will think it's terrific. >> >>Maybe you can think of a reason why he hasn't done this already... > >ASML sells machines to produce 0.18 micron for 13 million dutch guilder, >roughly 6 million US$. > >Then assuming it has same speed as DB, Hsu gets confronted with the >fact that he gotta do at least a 100 times less calls a second >to the hardware, so he needs to do roughly 2.5 ply extra in hardware, >say at 6 to 7 ply. > >This automatically means that less gets done in the search of the general >purpose processor, so hash can not get used there. Hash at the 0.18 >chessprocessor can't be done either of course, so that's an immense problem >when trying to search above say 9 ply, knowing that deep blue in crucial >positions got to 10 ply. > >Vincent > >>-Tom Vincent: Deep blue did _not_ get to 10 plies in critical positions. I have said this for 3 years now. The output logs clearly show this. They got to 10-11 plies in _software_. Plus another 4 plies in hardware. 10+4 does not equal 10.
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.