home *** CD-ROM | disk | FTP | other *** search
- Xref: sparky comp.arch:9415 comp.lang.misc:3068
- Path: sparky!uunet!dtix!darwin.sura.net!spool.mu.edu!sdd.hp.com!uakari.primate.wisc.edu!copper!aspen.craycos.com!sog
- From: sog@craycos.com (Steve Gombosi)
- Newsgroups: comp.arch,comp.lang.misc
- Subject: Re: Goals, Cost, and Flexibility (was Re: "Training" of programmers)
- Message-ID: <1992Sep14.175555.285@craycos.com>
- Date: 14 Sep 92 17:55:55 GMT
- References: <JAN.92Sep12155439@pallas.neuroinformatik.ruhr-uni-bochum.de> <9225808.14389@mulga.cs.mu.OZ.AU> <BuJnA9.JCA@mentor.cc.purdue.edu>
- Organization: Cray Computer Corporation
- Lines: 24
-
- In article <BuJnA9.JCA@mentor.cc.purdue.edu> hrubin@mentor.cc.purdue.edu (Rubin Herman) writes:
- >
- >This doubling is for the minis and micros, which are finding it possible to
- >come closer to the speeds of fast mainframes. The speed improvement from
- >the CRAY 1 to their latest machines is more in using multiple processors,
- >but this will eventually run into physical limits. The switching time has
- >only tripled in a dozen years; there may be breakthroughs, but another 20
- >doublings may very well reach basic physical limits in size, speed, and
- >number of processors.
-
- I believe you meant switching *speed*. A lot of the improvement from the Cray-1
- to the original (2 processor) X-MP was due to speed-ups in the original
- processor design. The faster clock was a minor improvement. Tripling
- the number of vector<->memory paths, adding flexible chaining, reduction
- of vector startup overhead, and allowing
- chaining to memory operations helped a lot more.
-
- Throwing processors at a problem certainly helps, if the problem is
- sufficiently parallel to benefit and the overhead of multitasking it is
- sufficiently low. Everyone won't benefit - if your problem is only 90%
- parallel, you can only speed it up by a factor of 10 whether you have
- 16 processors or 16 million - even *if* the multitasking overhead is zero.
-
- Steve
-