andyrose@batcomputer.tn.cornell.edu (Andy Rose) (11/20/90)
At SuperComputing '90 last week in NY City, IBM had a RS-6000 550 mowing down linpacks at ~65 mflops. $140,000 to $500,000. The theory center library is working on a magazine subscription list. Please e-mail the names of magazines and scientific journals you have found useful in visualization. I will post an edited list in a week. <thanks> There seems to be a lively debate about GL vs. Phigs as emerging standards for display list languages. I see more applications written in GL, and faster hardware running GL but this is changing. Phigs is 'died in blue' meaning it has IBM support (too what extent I couldn't guess). There are phigs drivers for Stardent and Suns also. I have heard that GL can go faster because phigs has not support for an interactive display list. I may have this very wrong and I would like to know the truth. Throw enough hardware at it though and phigs should run fast.(?) questions... Is conversio from GL to phigs (and vice versa) trivial? Can the two coexist? Are there other libraries in a position to be industry leaders? (Dore) What hardware runs what (see below)? Does X sit on top of these, i.e. do I care which if I write X code? How about Motif? Widget libraries? Phigs on RS-6000 (and IBMs with AIX flavor Unix), Suns (how about Sparc 2?). GL on RS-6000 and Silicon graphics. Personally I like GL (especially lookat, and stack management) and I haven't tried phigs. V for vis. -- Andrew Newkirk Rose '91 Department of Visualization CNSF/Theory Center 632 E & T Building, Hoy Road Ithaca, NY 14583 607 254 8686 andy@cornellf.tn.cornell.edu
mccalpin@perelandra.cms.udel.edu (John D. McCalpin) (11/20/90)
>On 19 Nov 90 20:41:35 GMT,andyrose@batcomputer.tn.cornell.edu(Andy Rose)said:
Andy> At SuperComputing '90 last week in NY City, IBM had a
Andy> RS-6000 550 mowing down linpacks at ~65 mflops. $140,000 to
Andy> $500,000.
Please note that this performance level is for the 1000x1000 "anything
goes" LINPACK test, not the more commonly quoted 100x100 "you can't
even modify the comments" version. I am anticipating that the latter
(which IBM uses for their published FP performance numbers) will be
about 19 MFLOPS, based on the speedup of the clock from the 540 to the
550 (41.? MHz vs 30 MHz).
It is possible to get 1000x1000 LINPACK performance of about 36 MFLOPS
on a Model 530 by careful coding (i.e. following the advice in Ron
Bell's technical report on the RS/6000). I expect that IBM will
release the fully optimized LAPACK package within a few months which
should provide similar performance levels for most dense linear
algebra stuff. Right now, only a few of the level-3 BLAS routines are
optimized --- DGEMM (in libblas.a) runs at about 45 MFLOPS on the 530
or about 30-32 MFLOPS on the Model 320.
--
John D. McCalpin mccalpin@perelandra.cms.udel.edu
Assistant Professor mccalpin@vax1.udel.edu
College of Marine Studies, U. Del. J.MCCALPIN/OMNETgrunwald@foobar.colorado.edu (Dirk Grunwald) (11/22/90)
>>>>> On 20 Nov 90 02:02:49 GMT, mccalpin@perelandra.cms.udel.edu (John D. McCalpin) said: >On 19 Nov 90 20:41:35 GMT,andyrose@batcomputer.tn.cornell.edu(Andy Rose)said: Andy> At SuperComputing '90 last week in NY City, IBM had a Andy> RS-6000 550 mowing down linpacks at ~65 mflops. $140,000 to Andy> $500,000. JDM> Please note that this performance level is for the 1000x1000 "anything JDM> goes" LINPACK test, not the more commonly quoted 100x100 "you can't JDM> even modify the comments" version. I am anticipating that the latter JDM> (which IBM uses for their published FP performance numbers) will be JDM> about 19 MFLOPS, based on the speedup of the clock from the 540 to the JDM> 550 (41.? MHz vs 30 MHz). --- One should also note that the ``anything goes'' doesn't mean that they coded it in assembler or anything, which is what you need to do to get speed out of the Intel i860. In fact, they just used block algorithms to increase cache locality, if I recall.