Author: Robert Hyatt
Date: 19:06:14 01/26/00
Go up one level in this thread
On January 26, 2000 at 18:33:41, Tom Kerrigan wrote: >On January 26, 2000 at 18:12:10, Jeremiah Penery wrote: > >>On January 26, 2000 at 18:06:45, Tom Kerrigan wrote: >> >>>No, I assure you that the general-purpose instruction vs. DB instruction >>>situation is very clear, regardless of background. If you don't understand it, I >>>will be happy to explain it via e-mail, but I don't think it needs to be >>>re-hashed here. >> >>Unfortunately, it's not so clear. But I don't plan on rehashing it here, >>either, until I get some more information on it, hopefully from the source. > >No, it is quite obvious. It takes very little computer expertise to understand, >too. Here is a direct quote from the abstract of FHH's IEEE article: > >"On a general-purpose computer, the computation performed by the chess chip for >one chess position is estimated to require up to 40,000 general-purpose >instructions." > >Now we can make the following assumptions: >* one hertz = one instruction (good enough) >* Bob is correct -> FHH meant 40,000 DB instructions >* DB chip searched at least 2M NPS > >Here is the simple arithmetic based on these assumptions: > >(2M nodes/sec) * (40k instructions/node) = 80G > >In other words, the DB chip would have to run at 80 GIGAhertz to search 2M NPS, >which the obviously did not do. > >-Tom That doesn't compute. Hardware doesn't work like that... If I need to do 10 operations that are unrelated, I don't take 10 cycles, I take one. If I need to do 100 things that are unrelated, I take one. If I have to combine things, I use a tree which means that logN() adds have to be done. But I can also do several adds in one cycle. And I can probably do 100 things, that then cause 100 more things to be done, which cause 100 more things to be done, all in one cycle. It is _very_ hard to compare a specific hardware solution to a general purpose one. Otherwise it would be hard to explain how their full eval takes 10 clocks, and a _full_ node takes 10 clocks, and doing _both_ takes 10 clocks. Obviously everything gets overlapped, ie doing the first clock of the fast eval, doing the first find-victim cycle, doing the second eval clock, the first find-aggressor cycle, etc. A bunch of things going on asynchronously, yet they get synchronized just in time to advance to the next node as needed. Take a simple adder circuit to get an idea of how this flies...
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.