Author: Robert Hyatt
Date: 12:28:50 08/16/05
Go up one level in this thread
On August 16, 2005 at 03:06:43, Majd Al-Ansari wrote: >Hi Bob, > >I have been following your implementation of new hardware with great interest. >I think it is great to push software to match the leading hardware. I was >wondering what you think of Hydra's way of using programmable cards as a >scalable way to increase hardware power. Is it more cost effective? I'm not a fan. The problem is, a program like Crafty can already reach 60M nodes per second with today's hardware. There are 16 processor opteron boxes around (not cheap) and dual-cores turn that into 32 processors, 4x faster than what I am using now (and hitting 16M typically). There are "issues" in a cluster like Hydra uses that are not issues in NUMA architectures, or at least nowhere near the same in terms of impediments to performance. Nothing wrong with the hardware approach. Witness Belle in the late 70's/early 80's, then HiTech and Deep Thought/Deep Blue. But there is a high cost, and you bear it all yourself since you get to buy special-purpose stuff that won't do anything else useful except for what you design it to do... > Or do you >think that as more and more cores get added to CPU's that is a better >alternative. While I am not a programmer, I was especially interested about >Hydra's claim that adding "chess knowledge" does not reduce the programs speed. That's a known advantage of special-purpose hardware. If you add an eval term, you can do it in hardware in parallel with the other terms. There is a very tiny speed loss when you use some sort of adder tree to combine all the individual scoring components, but that is a real benefit. In a software program, anything you do in the eval takes time, period, because it is all done serially in a general-purpose CPU. >I think with every engine you can get reach certain positionss where even the >best engine can play moves that even a 1800 player can do better. I have one >position in mind where Shredder 9 UCI will resign as black in an absolutely >drawn endgame (gives evaluation +9.8) and I am sure you have similar experiences >with certain positions. That's a problem. But I should add the number of such positions 20 years ago was _far_ larger than it is today. The programs are getting better. They are just a long way from perfect. > Ofcourse the programmers are aware of these things but >will not add this chess knowledge because it will reduce the strength of the >program in other areas. If Hydra's claim of being able to add chess knowledge >without reducing speed is true, then isn't that the way to go? Thanks.. The problem is that their special-purpose hardware will be slow in 2-3 years, compared to off-the-shelf cpus. Each new generation of processor gives us the ability to either go faster, or add more "stuff" and stay at the same speed. Or strike a happy medium between fast and smart... > > > >On August 15, 2005 at 11:03:54, Keith Hyams wrote: > >>Does anyone know if/where a list of the hardware that the programs are using has >>been published? Details of version numbers would be interesting too. >> Regards >> Keith
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.