Author: Robert Hyatt
Date: 21:13:05 02/28/03
Go up one level in this thread
On February 28, 2003 at 19:30:21, Torstein Hall wrote: >On February 27, 2003 at 16:43:45, Robert Hyatt wrote: > >>On February 27, 2003 at 15:57:04, Brian Richardson wrote: >> >>>As I recall, 5 years ago folks were saying only another 10-12 years for Moore's >>>Law speedups...now they are still saying another 10 years or so. I agree that >>>at some point physics will dictate limitations, but then there is more >>>parallelism. Sun just outlined plans for running 4 threads on each of 4 cores >>>on a single chip in the 3-5 year time frame. That would be roughly 16x. Both >>>Intel and IBM have similar plans to extend on-chip parallelism. >>>Bottom Line: Just as coding for 64 bits will become the norm soon, so will >>>coding for parallel searching with multiple threads. >>> >>>Brian >> >> >>If you look back over the past 5 years, I've said that a hundred times. >>"Moore's law" >>is definitely fading fast. > >That statement do not sum correctly if I remember right. Is it not 5, 6 or 7 >years ago a 200MHZ Pentium Pro was really hot? At the moment Intel is fast >moving beyond 3Ghz. Thats a hefty 15 times clock speedup in the period, not >counting paralelism etc. etc. The pentium pro 200 was released in late 1995 if I recall correctly. I received mine in 1996 and used it in Jakarta at the WMCCC event that year. 200mhz to 3000ghz today. 15x faster. 8 years. log2 (15) is less than 4, which turns into roughly a doubling every two years, depending on where we are at the end of 2003. It was originally 18 months. But look back maybe 3 years and it seems to be a bit less. And based on current "roadmaps" it seems to be somewhat less for the next two years as well. The engineers claim they are really _struggling_ at .08, and that going beyond that is a _real_ challenge. They might go to x-rays for a shorter wavelength, perhaps, but then they still have to deal with making oxide layers insanely narrow where you can count the atoms across a single path with your fingers and toes... > >So if this is right Moores law is breaking, the speedup seen from a consumers >point of view is much higher! > >Torstein > > > > But a "pseudo-moore's law" dealing only with >>performance, >>has a hope for quite a bit longer, but in the realm of parallel programming. N >>cpus on >>a chip has been done already. SMT is a different take on the same thing. A >>single chip >>with N cpus and M threads makes complete sense, although that will only extend >>the >>limit a few years, because you can't keep adding cpus without making the cpu die >>size >>requirement smaller. And that is what is coming to an end...
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.