home *** CD-ROM | disk | FTP | other *** search
- Xref: sparky comp.benchmarks:1660 comp.arch:10637
- Newsgroups: comp.benchmarks,comp.arch
- Path: sparky!uunet!mcsun!news.funet.fi!cs.joensuu.fi!jahonen
- From: jahonen@cs.joensuu.fi (Jarmo Ahonen)
- Subject: Re: DEC ALPHA Performance Claims
- Message-ID: <1992Nov12.074705.19215@cs.joensuu.fi>
- Organization: University of Joensuu
- References: <1992Nov11.043149.17257@engage.pko.dec.com> <BxKI38.DM7.2@cs.cmu.edu>
- Date: Thu, 12 Nov 1992 07:47:05 GMT
- Lines: 53
-
- lindsay+@cs.cmu.edu (Donald Lindsay) writes:
-
-
- >sharpe@adodem.enet.dec.com writes:
- > [ A list of the MFLOPS the Alphas get, inverting dense matrices. ]
- > [ I have extended the list to give people a feel for the context. ]
-
- >>
- >>System 100x100 dp 1000x1000 dp
- >>
- >Compaq Deskpro 486/33 1.4
- >DEC 5000/200 (25 MHz R3000) 3.7
- >Cray-1 11 31
- >>DEC 3000/400S AXP 26.4 70.8
- >>DEC 10000/610 AXP 42.5 111.6
- >Convex C3810 (1 proc) 44 113
- >Cray C90 (1 proc) 387 874
-
- Numbers are always interesting, but I think that I might tell you
- some of my experiences which always make me to calm down when I hear
- excellent performance numbers. I have tried to run some in-house written
- code on Suns, HPs (730, 750, and 710), and Convex C3840 and C3420
- machines. With very small datasets those "real life" programs are
- *very* fast on big HPs (actually more than 1.3 times the C3420 if
- you run the program with only one processor on the C3420), but
- with large datasets the C3420 provides more than twice the
- performance of HPs (C3420 with options that the program
- is not parallerized and all processes fit into RAM, i.e.
- no paging).
-
- The C3840 is, according to my experience and my codes, more than
- twice and sometimes even more than thrice the performance of C3420
- (remember, no parallerization).
-
- The suitability of the program for the architecture has, of course,
- an enormous effect on the real-life performance. For example the
- much used BP -package for neural-network simulations is dog slow
- on Convex machines (at least relatively slow), and some FORTRAN
- codes originally written for VAX are fairly slow on HPs and very
- fast on Convex.
-
- For those reasons I have decided to run my programs on different machines
- and then decide how fast they are. I'm eagerly waiting for the possibility
- to rum my programs on Alpha-based machines.
- I'm especially interested in the relation between the size of the dataset
- (and the overall process size) and the relative performance.
- Interesting times ahead, I believe :-).
-
- ---------------------------------------------------------------------
- Jarmo J. Ahonen
- Computing Centre, Lappeenranta University of Technology, P.O.Box 20,
- SF-53851 Lappeenranta, Finland. email: Jarmo.Ahonen@lut.fi
-
-