Author: Robert Hyatt
Date: 10:40:46 10/14/02
Go up one level in this thread
On October 14, 2002 at 05:12:59, Vincent Diepeveen wrote: >On October 14, 2002 at 02:43:37, Slater Wold wrote: > >>On October 13, 2002 at 23:10:38, Keith Evans wrote: >> >>>On October 13, 2002 at 14:25:00, Vincent Diepeveen wrote: >>> >>> >>>>You cannot use killermoves in hardware and you cannot use hashtables >> >>You sure the hell can. > >in theory you can, but the slowdown is too big. Vincent, why don't you simply shut up when you don't know what you are talking about? It is _not_ about "slowing down" a thing. It is simply about circuit complexity. It wouldn't cost a bit of speed to speak of. It _would_ make the circuit more complex. Ken Thompson can tell you why if you want, he made the original decision to not do that and published why. > >IBM asks for the maximum nodes a second. CPUs are clocked 20 or 24 Mhz. >the 20Mhz is easier to calculate with. > >It gets about 2 million nodes at 1 cpu. > >That's 10 clocks. 20Mhz / 10 clocks = 2 million nodes a second. > >If you add killermoves, then in move ordering phase it will require 2 >extra clocks to use them, i assume storing them goes in parallel again. Where on earth do you pull these numbers from? On second thought, they stink so badly I _know_ where you pull them from. > >that's 20Mhz / 12 clocks = 1.7 MLN nodes a second. > >If you add hashtables that'll eat like 4 clocks in the sequential search. > How so? Belle _hashed_ in the hardware. Deep Blue hashed in the hardware, but it didn't have memory to probe/store in, but the hardware already had the hash table circuitry designed and built in. He just didn't have time to design and build the memory because other things took his time. >So then you drop to 20 / 16 = 1.25MLN nodes a second. > >Now i would still go for it, if it just eated 4 clocks and 2 clocks for >killermoves. Killermoves give more like 20% speedup. > >However things get more complicated than this when the searches were >very small in the hardware, and according to hsu's statements so far it >seem they were. > >>>I don't see any technical reason why this would be true. It shouldn't be a >>>problem to hack killermoves into a Belle style move generator - see what Marc >>>Boule did in his master's thesis. >>> >>>Even Hsu mentions that hashtables are possible. (Not in the latest chat - just >>>see his previous publications.) >>> >>>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.