Author: Vincent Diepeveen
Date: 13:11:00 12/27/05
Go up one level in this thread
On December 27, 2005 at 14:20:50, Dann Corbit wrote: >On December 27, 2005 at 12:21:42, Gian-Carlo Pascutto wrote: > >>On December 27, 2005 at 11:29:22, Vincent Diepeveen wrote: >> >>>You know, without 64 processors, having 2 hardware cards which are development >>>cards at just 60Mhz or so, that has simply zero chance against software. >>> >>>those fpga boards have been made to *develop* chips, not to use them as release >>>processors. released fpga hardware runs at near 1Ghz, not 60Mhz!! >>> >>>So hydra at 1Ghz would surely be a good match for the software, but at 60Mhz, >>>no chance, really. fpga development boards weren't designed to be used as >>>production hardware :) >> >>Well, regardless of the speed issues, which could be "solved" by using faster >>boards (but which might require redesigning the Hydra hardware entirely due to >>routing/timing issues), I think the largest problem of Hydra is the long >>development and testing time of an FPGA based solution compared to a general >>purpose CPU. >> >>Given that the problem of computer chess is not finding nice algorithms, but >>testing what works, I don't think Hydra has the faintest chance in the long run, >>and it's probably already eclipsed by Zappa, Fruit and Rybka now (silly claims >>of the Hydra authors notwithstanding). > >I think the biggest problem with the hardware approach is one of cost. >For a few million dollars, they could remake all the boards faster. Chrilly is a true genius in tuning. By now in 2005 there is 3 programs which for sure have a better tuning. Fruit, Rybka and Zappa. In 2006 we can expect a new Sjeng version, a new Diep version and a new Shredder version. Hopefully joining the world top is also Spike in 2006. Perhaps other programs will follow too. The issue Gian-Carlo raises is therefore very valid. The whole machines software and hardware programming gets all done by 1 person and 1 person only. Just a single recompile of his own fpga board takes like 1 day. Just testing something simplistic which takes me litterary a few minutes to modify source code, then i recompile; that recompile in software eats just 10 seconds at most. The same procedure in hardware takes like 2 days. First of all some hours to bugfix the verilog code. So even the tiniest bugfix of a pattern you can wait another day. Then if he plans to test it and ships it to testers, is he going to ship a fpga card with each shipment? Imagine the time it takes for that camel to reach Abu-Dhabi with that fpga card on its back! Each year we see clearly chessprograms must progress over 50 rating points. Extra search depth helps a tiny bit, but most are simply software improvements. New algorithms, small tweaks and enhancements. More chessknowledge especially. Fruit was the first this year to improve bigtime. Zappa we all saw coming of course, now there is Rybka. The progress keeps going and keeps going. Of course 2005 is also having the free downloadable source code of fruit 2.1, so many will of course make a jump now, but obviously: In 2006 one needs to improve again. I just don't see a single person with such complex hardware problems make that jump. The real genius of Chrilly is tuning and getting the job done at the right time. I also know a commercial programmer, whose name may not be revealed, who says: "Chrilly is always very good in creating a lot of noise around him." (i hope i translated the dutch way of expressing things to english correctly) Especially the last 3 programs we saw coming, namely fruit, zappa and rybka, it is clear these programs have been very well tuned. So based upon that i conclude Hydra is an outdated program now. I am very very sure we will never see it again participate in a computerchess event against other programs. >Or you can take a PC program wait for 10 years, and get the same thing for $50. > >Now, the multi-million dollar solution will always be a bit faster, but every >new generation will have stupendous retooling costs for any upgrade. > >Now, suppose that the present Hydra were +200 Elo better than any existing >program (I am pretty sure that it's not, but as an absolute and utter edge, I >think it is very safe that it is not more than that). I can take a $50 copy of >Rybka on fast hardware and let it think for 32 hours on a problem instead of one >hour and get 5 doublings of compute power. This would give better analysis (at >a cost of some time) than the multi-million dollar solution if it really were >that much better. Now, if I wait 5 years, I can put Rybka on a 32x faster >machine which will cost the same as today's machine, and it will be actually >faster than Hydra. Now, they could upgrade Hydra again -- but at stupendous >cost. I will never have to make the same kind of absurd expense. I just have >to be patient. I will always have equivalent analysis even with the latest >iteration of the hardware giant if I just let the program analyze a lot longer. >And in a few years, the software solution will outstrip the hardware solution. > >The Chess-MIPS per dollar will always be far superior with the software solution >than with the hardware. > >But I guess that custom hardware will always be a bit faster than the fastest >software system (albeit at enormous expense) if they choose to retool and >rebuild to get the utmost in speed. > >Furthermore, I guess that Hydra is perhaps a little better than Rybka on very >fast 64 bit hardware, but maybe not. So we may already get Hydra performance >for $50 instead of 5 million. I know which alternative I would choose.
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.