Author: Robert Hyatt
Date: 08:38:15 07/20/05
Go up one level in this thread
On July 20, 2005 at 07:02:52, Anson T J wrote: >On July 19, 2005 at 20:34:32, Robert Hyatt wrote: > >>On July 19, 2005 at 16:34:57, Anson T J wrote: >> >>>As all threads don't calculate the same, the NPS is a fair enough measurement. >>>I'm more interested in the NPS than time to depth. >> >>This metric is simply no good. What if a 1 cpu search hits 1M nodes per second, >>and takes 2 minutes to complete 15 plies. And the 2 cpu search hits 2M nodes >>per second and takes 2.1 minutes to complete 15 plies. Which one is better? > >> >>NPS is like a tachometer in a car. Doesn't say a thing about how fast the car >>is moving, just says how fast the motor is rotating. > >Isn't the definition of "scaling" the number of nps obtained with multiple cpus >divided by the number of nps with a single cpu? You guys are talking about speed >up, which for now I am not interested in. > >The case you mentioned above will be highlighted when I ask for speed up >results. For now, I am interested in how well they scale. You are simply looking at the wrong number. "we" (the authors/developers) are interested in the NPS scaling, because a poor scaling indicates an internal design issue that is causing the hardware to perform below theoretical maximum performance. But that has very little to do with playing chess. The speedup indicates how efficient the parallel search is. And it is a _far_ bigger concern. I spent a lot of "scalability time" last year trying to get the AMD platform to scale properly, and eventually was successful. But think about this. A reasonable parallel speedup for two cpus for any program is 1.7X-1.9X. Reasonable scaling is above 1.9X, all the way to exactly 2.0X. But a program with 1.1 parallel speedup could scale perfectly, and still play lousy chess. The only meaningful number, from a chess skill perspective, is "how much faster does a program reach a specific depth using 2 processors vs using just one?" Any other measure is totally useless from a program-user point of view. Last year Crafty was getting a 1.7X speedup on the AMD box, but only a 1.8X scaling. Later the scaling went to almost 2.0X. Scaling definitely impacts performance, but far less than simple speedup... If you are only interested in scaling, do you only look at the tach when you test-drive a new car??? That doesn't tell the _important_ detail about how fast you are actually going...
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.