Author: Eugene Nalimov
Date: 14:34:50 05/14/99
Go up one level in this thread
You really have to read Hsu article in IEEE Micro, and maybe several previous... Here are my 2 cents: RS/6000 CPUs did first plies of search, plus extensions. Chess chips searched last 4 plies, plus quiescence search, plus did all the evaluation. Evaluation function was hardcoded in the sense that all the evaluation factors were fixed and you cannot easily add the new ones (there were a way to do that, but it required implementation of the new hardware and was never used). You can modify weights of those factors (and can effectively turn off the factor by assigning a zero weight to it). During the second match with Kasparov DB team greatly increased king safety score after the game one (I think that really happen before the game 3). RS/6000 is not the best integer CPU - if I remember it correctly, its' SpecInt resembles SpecInt of the Pentium/200. It has much better floating point unit, but it chess programs do not use FP. On the other hand, it has much better memory bandwith - that's essential for 30-CPU system, but not so important for simple- or dual-CPU system, "Efficiency" of the parallel search for such huge # of CPUs was 25-30%. By replacing 480 chess CPUs by 4 chess CPUs that are 16 times faster than the original you'll greatly decrease overhead - I beleive Bob showed that for small # of CPUs (less than 16) speedup can be nearly linear. So, 'simplified' modern 4 chess CPUs DB can easily achieve the performance of the 1997 monster. 1 CPU system will be 3 times slower - not bad for $200. Special-vs-general hardware: yes, special hardware technology is usually worse that technology used in mainstream CPUs. Nevertheless, specialized products are usually faster, and you routinely use them. I'd bet that your video board with 2D and 3D acceleration is built using 0.35 to 0.5 micron technology - and it can do what it supposed to do much faster than your 0.25 micron processor. And processor will not catch it in ~5 years. You cannot directly compare Fritz (or Nimzo, or Junior, or Rebel, or Crafty) nps with DB nps. Hsu estimated that their evaluation functions will use ~40K instructions of general CPU. Based on Fritz nps, its' evaluation function is 1-3K instructions. So here is additional factor of 20. Eugene On May 14, 1999 at 15:45:12, Gregor Overney wrote: >On May 14, 1999 at 09:52:14, Robert Hyatt wrote: > >>>1) You just simple can't put DB on a single chip. >> >>Back to the drawing board on _that_ comment. DB _was_ based on a single >>VLSI chip. They then used 480 of them to do a very fast parallel search. >>Hsu, in the current issue of IEEE Micro has said that he is re-designing the >>chip using new fab facilities, and can make a single chip run 15 times faster >>than the chip used in "DB". IE a single chip will search about 36M nodes per >>second... >> > >Does DB stand for Deep Blue? Or is this discussion about another "DB", such as >Deutsche Bundesbahn, maybe? :-) . To cite IBM's statement: > >"Deep Blue is at heart a massively parallel, RS/6000 SP-based computer system >that was designed to play chess at the grandmaster level." > >Where are the RS/6000 chips on this DB chip? And how would you put 480 chips in >one? Lets make a some small calculation: DB (1997 version) runs at 200 >Mnodes/sec (right?). It uses 480 chips and 32 (?) RS/6000 CPUs. Lets just assume >that those 32 (?) RS/6000 CPUs are just sitting there to diplay an IBM logo. >Well, 200,000,000 divided by 480 times 15 gives 6.25 Mnodes/sec (why 36?). > >So, what do we need to make this single DB-chip a success? > >1) The expected factor 15 is real (has to be seen) >2) The search is all that's needed to play chess (I doupt it!) >3) This chip will hit the market before 2001 and sells for less than $100 (in >quantity of 1000). Otherwise it will be too expensive once it hits the consumer >market. > >> >> >> >>> >>>2) You still need to write the correct algorithms to make this chip work. And >>>those algorithms are pretty complex (see evaluation functions etc.) >>> >> >>But it has _already_ been done. All that is left is to use the "new" fab >>process to increase density and clock speed.. DB's chess chips only ran at >>20-24 megahertz. running that up to 16x faster seems quite easy with todays >>silicon capabilities as that would still be a modest < 400mhz processor. >> >> > >What are you telling me is that the evaluation functions are _hardcoded_ in the >chip that are supposed to make the search? What were those RS/6000 chips doing? > >> >>>3) It is not the first time, people try to design super-fast search engines on a >>>chip. Most of those efforts were gradually falling behind "real" CPUs. It's a >>>nightmare and not very profitable. Just look at those countless chips that have >>>been designed for image and speech recognition. A standard DSP with the right >>>software does the trick much cheaper. For chess, a solid SMP 64-bit architecture >>>and the right algorithms should always succeed. >>> >>> >>>Gregor >> >> >>But a special-purpose chip will _always_ be at least an order of magnitude >>faster than a general purpose solution. Always has, always will be... > >Unfortunately, special-purpose chips are always released behind the market's >release of new general purpose chips. Let's see in 2002 if a 6.25 Mnodes/sec >general-purpose chip is a factor 10 faster than DEC's new Alpha that will be >released in 2002. - If I am not completely mistaken, an Intel pIII/500 runs more >than 300 knodes/sec using Junior 5/32. (Even an old p6/200 cranks 100 knodes/sec >out of Crafty). And that's just today's entry level CPU for new systems. Intel >has already announced that they can produce 1000 GHz versions of it (in large >quantities). In the close future, it will be really difficult to be a factor 10 >faster. - At the end it is supposed to play chess and not just make a simple, >but super-fast brute force search. > >Gregor
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.