Author: Robert Hyatt
Date: 20:45:00 04/27/01
Go up one level in this thread
On April 27, 2001 at 16:37:14, Vincent Diepeveen wrote: >On April 26, 2001 at 22:28:21, Robert Hyatt wrote: > >>On April 26, 2001 at 22:10:14, Robert Raese wrote: >> >>>On April 26, 2001 at 19:58:43, Robert Hyatt wrote: >>> >>>>On April 26, 2001 at 18:11:20, Rajen Gupta wrote: >>>> >>>>>hi bob: you mentioned in you recent post regarding hsu's chip being capable of >>>>>doing 1 billion nodes per sec-is this an advanced version of the "chess chips" >>>>>which were present oin the deep blue?-as far as i remember the deep blue had a >>>>>number of general purpose powerpc processors and a number of what they termed >>>>>"chess chips" if this is so >>>> >>>>NO... here is the math. DB2 used 480 chess processors. About 1/2 of the >>>>processors ran at 20mhz, the other half ran at 24mhz. >>> >>>ok this is probably a stupid question, but isn't 20mhz kind of slow for such >>>advanced hardware? what am i missing? >> >>Not particularly. These are ASICs and they do a _bunch_ per cycle. In fact, >>it takes just 10 clock cycles to handle one node... That translates to >>several thousand machine language instructions for the typical chess problem. >> >>20mhz was slow. Hsu had planned on a 15X faster chip without really pushing >>the fabrication limits of today... He even posted once that 30M nodes per >>second per chip was easily doable... >> >>Don't confuse 20mhz in a special purpose finite state machine vs 20mhz in >>a microprocessor... they are really not comparable... > >such a chip with nullmove and with all kind of enhancements like >SRAM on chip to use a small hashtable, now that would be interesting. > >Of course both features slow down NPS huge, but that's no big deal if >you do so many millions a second! > >Also can use my evaluation then on chip. Going to be quite a bit of >clocks though as tough code needs other tough code to in the >end again combine things. So 10 clocks is, unless more sequential things >can be done in a clock nowadays as in 1997, not possible anyway, >which will be no problem as using a hashtable is going to slow down >things most. > >Would be real interesting to see how fast something like that gets! > >What i really do not understand is why Hsu never put hashtables at >his chips. Didn't he have enough budget to do that or was this in >1997 technical quite hard? > >I'm pretty sure that also in 1997 there was SRAM! > SRAM is made of nothing but transistors... not capacitors. Which means SRAM takes a lot more space. But then it is _not_ slow. That is what register files are made of. The chess chips have a memory addressing scheme built in already. But Hsu didn't have time to design and build the memory system (16-ported for 16 processors on a single board) in time for the 1997 match. It was doable... and at a 20mhz clock speed it wouldn't have slowed things down one iota... >Best regards, >Vincent
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.