Author: Dann Corbit
Date: 16:54:34 02/17/04
Go up one level in this thread
On February 17, 2004 at 19:50:59, Bob Durrett wrote: >On February 17, 2004 at 18:43:20, Dann Corbit wrote: > >>On February 17, 2004 at 18:14:28, Bob Durrett wrote: >> >>>On February 17, 2004 at 17:47:56, Slater Wold wrote: >>> >>>>About $50,000 and 2 years. >>> >>>It might be worth it! think marketing! >> >>In two years, the 20M NPS that crafty can now do in software will be 80M NPS. >>In two more years, the 80M NPS will be 320M NPS (faster than Deep Blue). >>In two more years, the 320M NPS will be 1.3B NPS (absurdly fast). >> >>How would anyone ever get their money back? >> >>Keep in mind that Dr. Hyatt's program (which smashed all comers in the last >>CCT-6 contest) was not present at Paderborn in all its resplendent glory. >> >>With an 8-way box running at full tilt, I expect that the sofware crafty is >>already more than a match for Hydra. > >I disagree with none of the above. > >But think about this: > >Someone here recently said that there are two thousand lines of code in the >source code for Crafty's position evaluation. Admittedly, all two thousand >lines would not have to be executed [after compilation] but nevertheless, it's >easy to imagine that it would take quite a few clock cycles for this code to >produce an "evaluation." > >I do not know what percentage of the processor's time is used for position >evaluation, but for the sake of discussion, let's say it's 50%. > >Replacing that code with hardware which would produce the desire outputs in a >few clock cycles would essentially halve the amount of time required to achieve >some prechosen depth of search. [Same depth, half the time.] > >Now suppose another block of code were identified which accounted for 50% or the >clock cycles [assuming the evaluation hardware is being used.] Then having that >block of code "hardware-ized" too would produce a factor of four improvement. >[Same depth, one fourth the time.] > >We have two to the nth. If there are nine such blocks reduced to fast hardware, >then we have two to the ninth or 512. That means it takes only one 512th the >time to achieve the same depth. > >If most of the code were replaced with fast hardware, the amount of time >required to reach a predetermined depth might approach zero. This means that >the "hardware-ized" Crafty would be enormously powerful. It might take a very >long time for the speed improvement in PCs to catch up and don't forget that as >PC hardware gets faster, the hardware in the "hardware-ized" Crafty may be >getting faster AT THE SAME RATE! > >The more I write the more I'm convinced that this experiment should be >performed. Incidentally, $50,000 is probably far too little for research. If you do a profile of crafty, it is very even handed. If you remove a bottleneck, the next routine will be the bottleneck. I think 50K is way, way underpriced, unless it is only move generation and some rudimentary evaluation terms. To transform all of crafty would surely be several million dollars.
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.