Author: Dann Corbit
Date: 20:05:27 01/20/00
Go up one level in this thread
On January 20, 2000 at 22:25:42, Peter Kappler wrote:
>On January 20, 2000 at 19:41:12, Dann Corbit wrote:
>
>>On January 20, 2000 at 12:15:33, Peter Kappler wrote:
>>[snip]
>>>I'd expect DB2 to win, but I think it would be closer than people expect. It
>>>certainly wouldn't blow the micros off the board.
>>
>>But we are talking about old technology verses new.
>>
>>Consider {back to the present}:
>>The fastest incarnations of RS/6000 do 3+ teraflops. That's 3,000 gigaflops or
>>3,000,000 megaflops.
>>
>>The new Hsu chips are much faster than the predecessors.
>>
>>Take a new RS/6000 ("fully loaded") and add as many Hsu/Campbell processor
>>systems as will fit into it. It is potentially one thousand times as powerful
>>as the old system.
>
>1000 times? How did you get to this number?
The old machine had 480 chips.
I am assuming 64,000 of the new chips (256 clusters of 256 chips per cluster).
64,000 times 30 M NPS per new chip = 1920000000000 / 200000000 = 9600x as fast
for hardware node calculations. Figure 90% wasted energy because of
communications drawbacks gives 1000 times faster (roughly).
>> Why not? Now, add in a full complement of 8 piece tablebase
>>files (calculated by that machine).
>>
>
>Huh? 8 piece? Any idea how many terabytes of disk this would require? Not to
>mention the months(years?) it would take to generate the tablebases. We're just
>now starting to see the first 6 piece tables, and I don't believe anybody has
>ever generated a 7-piecer. Maybe a TB expert will chime in and tell us what it
>would take to get to 8.
IBM could use Eugene's algorithm to generate them in a few months (wild guess -
but each of the 256 boards would work on a separate problems with (perhaps 32
gig of memory at their disposal). -- I did say "fully loaded."
>
>>If it were their goal, IBM, Hsu and company could make a chess machine that
>>would crush a team of Kasparov + Old-Deep-Blue [Just conjecture, but I think
>>that they really could do it]. Perhaps an ELO of 3500.
>>
>
>What you're describing is a machine that would score ~99% against Kasparov. I
>think computer chess will *never* reach that level.
I don't think the improvement would be just the horsepower boost of a few extra
plies. That sort of reasoning would assume that we just use the exponential
increase and brute force forward (though that alone would be pretty awesome).
Rather, I am assuming that they calculate brute force out to 18 ply or so, and
then use extensions to get out to 60 ply or so by extrapolation of favorable
possibilities.
>This goes back to the old debate about what ELO rating a computer would attain
>if it played chess perfectly. I have always argued that this number is finite,
>because I believe that the drawing margin in chess is sufficiently large to
>allow one side to make a few mistkes and still draw against a perfect opponent.
I don't know nearly enough about this even to speculate. I suppose if Kasparov
played for a draw as white and conceded the loss as black, such a strategy might
minimize the damage. ELO is relative to a large pool of players. It would be
difficult to get in enough games to get an accurate
>But that's a whole different discussion... :-)
>
>
>>A machine like that would simply be unbeatable. It could make strategic plans.
>>It could analyze every move that has ever been played to a depth of 20 plies and
>>store it in a database. Such a computing device would be invincible.
>
>
>I think that more speed is not the answer, because at 15+ plies, you're already
>faced with with seriously diminishing returns per extra ply. Any tactics that
>exist in the position (or that a human would ever have a prayer of finding) have
>long since been discovered, so the computer's positional evaluation starts to
>become the limiting factor. And computer chess still has a long, long way to go
>in that area. As you've stated yourself in other threads, the GMs are getting
>better at exploiting these weaknesses.
Why assume that all of that horsepower must be applied in the age-old method of
pounding through ply after ply. I don't think it needs to work that way at all.
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.