Author: Angrim
Date: 02:06:34 12/16/99
Go up one level in this thread
On December 15, 1999 at 14:39:18, Greg Lindahl wrote: >I'm interested in coordinating an effort to build the next >world-champion chess program. FPGA technology has advanced since Deep >Blue's construction such that replicating its architecture is now >inexpensive. All you need is a cluster guy (me), a pile of FPGAs >(corporate sponsor), someone who knows how to route a board evaluation >chip, and someone interested in building a message-passing chess >program which uses the board evaluation chips as an accellerator. > You will need more than a board evaluation chip, the chips in DB did a 4 ply deep search before the eval. This will be much more complex to make, but due to the bandwidth between the chip and the computer it is in, it is vital. >In the tradition of Deep Blue, I'm fairly clueless about chess. Actually, while IBM is fairly clueless about chess, the DB team itself had a fair amount of chess savvy, and *hired* a top level player to give them advice on its play. >Unlike Deep Blue, I plan on making this engine available for more than >just a few games, so that its behavior can be studied. Cool. >If you are interested in this project, please drop me email. I'm >especially looking for a person/group seriously interested in >constructing the overall program. I can provide the cluster & support, >find the FPGA corporate sponsor, and find someone to route the chip. > >-- greg I've seen an idea like this one a number of times before, but this is the first time I have seen an offer to provide the hardware and money part of the problem. This boosts your odds of success from 0.000% up to 0.1% or maybe even 1% if you are really stuborn. Here are a few things to consider I think. 1. How long are you prepared to support the development of this before you get your "world champion" player? DB took more than a decade, and its developers were working on it full time for a fair amount of that period, volunteers can not be expected to put as much time in. While the hardware available now is much nicer than when DB was started, the software part will have to be done nearly from scratch. 2. As you no doubt know, message passing parallel systems are prone to subtle race condition and deadlock bugs that are very hard to find except on the target system. This means that at least a small part of the hardware cluster will need to be built and be available for testing of the code for the duration of the project. 3. Will the code produced be GPL? If so, will this disturb your corporate sponsor? If not, it will be even harder to get volunteers. 4. Since nobody want to donate work to a project that is doomed, expect people to want a lot of information about every little detail of this project before they offer to help. 5. The program you are asking for will have very little in common with the chess players that many of the readers of this board have written. In the usual chess program, a great deal of time and effort has gone into designing very fast move generation and board evaluation code, and this is mostly irrelevant to a hardware assisted program. A few people here have written SMP based programs, but they assume a huge block of shared memory for the hash table which is not available in a message passing system. Good luck, and hope this doesn't depress you too much. Angrim(email? whats that? ;)
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.