kahn@batcomputer.tn.cornell.edu (Shahin Kahn) (02/22/90)
In article <157@illini.osc.edu> djh@osc.edu (David Heisterberg) writes:
<
<Does it really matter whether the 780 was 1 MIPS or not? As long as
<you compare on the same scale the numbers should be somewhat meaning-
<ful. A few of DR's CPU2 pseudo MVUPS follow
<
<SUN-4/260 SunOS 4.0 5.91
<
<DECstation 3100 Ultrix 3.0 12.88
<
<Apollo DN10000 Domain/OS SR 10.0.0.p 16.60
<
<Multiflow Trace 7/300 Trace/Unix V5.0.4 31.45
<
<Convex C210 (1 CPU) Unix 7.1 fc 5.1 44.68
WRONG--------^^^^^
That's for a 4-processor machine.
From DR lab's article October 2, 1989:
Convex C240 1 CPU: 36.42 +/- 8 MVUPS
2 CPU: 39.62 +/- 10 MVUPS
3 CPU: 42.24 +/- 13 MVUPS
4 CPU: 44.68 +/- 16 MVUPS
(** I am estimating the "+/-" values based on the published graph,
but the MVUPS values are those that are printed."
The article on VAX-9000 also mentions
Multiflow Trace 28/300 at 46.89 MVUPS and
FPS 500-EA (1 scalar, 1 vector) at 47.06 MVUPS.
It only mentions these three and the VAX.
<
<VAX 9000-210 unspecified 52.78
<
<These don't seem out of line to me. But it seems that the 95% confidence
<interval gets fairly large for the faster machines, and especially for
<vector and multi-processor machines.
<
<When real people have real VAX 9000s then we can check out the SPEC
<numbers. But it looks like DEC's done a respectable job of engineering,
<though at a cost: half the circuitry is diagnostic, and prices start
<around the $1M mark.
<
<David J. Heisterberg djh@osc.edu
<The Ohio Supercomputer Center djh@ohstpy.bitnet
<Columbus, Ohio ohstpy::djh
Well-said. We need to see the whole report for the VAX-9000.
Starting prices seem too high, since they usually mean a bear-bone
system. , The question is whether the starting configuration yields
the reported performance.!
gil@banyan.UUCP (Gil Pilz@Eng@Banyan) (02/22/90)
In article <9768@batcomputer.tn.cornell.edu> kahn@tcgould.tn.cornell.edu writes: >Well-said. We need to see the whole report for the VAX-9000. >Starting prices seem too high, since they usually mean a bear-bone >system. ^^^^ ^^^^ My God, is that what they're making them out of !? I wonder why we haven't heard from the animal-rights people on this one ? Gilbert W. Pilz Jr. gil@banyan.com
mmm@cup.portal.com (Mark Robert Thorson) (02/26/90)
> Depends on what you call a MIP. I just did the following on a VS2000 > (basically a uVAX II) > > MOVL #100000,R2 ; do 10 million instructions > 10$: NOP ; \ > <97 more NOPs> ; \ 100 instructions > Command: > DECL R2 ; / > BNEQ 10$ ; / > > and got a result of 2.25 seconds, that works out to 4.444 MIPs This reminds me of someone I knew who bought a nice 386 machine from Five Star. He wrote a little benchmark and set it for a miilion iterations. POW! It came right back. "Gee, that's pretty fast," he thought. Then he set it for 10 million iterations. POW! It came right back. Now he knew something was wrong. He popped the object code into the debugger, and he discovered his compiler had realized no products of the loop were being examined, and it removed the loop from his program.
hascall@cs.iastate.edu (John Hascall) (02/27/90)
mmm@cup.portal.com (Mark Robert Thorson) writes: }I write: }> Depends on what you call a MIP. I just did the following on a VS2000 }> (basically a uVAX II) }> MOVL #100000,R2 ; do 10 million instructions }> 10$: NOP ; \ }> <97 more NOPs> ; \ 100 instructions }> DECL R2 ; / }> BNEQ 10$ ; / }> and got a result of 2.25 seconds, that works out to 4.444 MIPs [386 machine returns instantly on 1 & 10 million iter. loops] }Now he knew something was wrong. ... compiler ... removed the loop Are there optimizing assemblers!?!? Anyhow, MACRO-32 doesn't, the loop is actually there. If memory serves, the uVAX instruction prefetch is big enough for eight NOPs (and some of the more `hideous' instructions don't even fit) which probably explains this `unusual' performance. (The VS3200 was > 10 MIPS :-). Assuming uVAXII == 1MIP, it looks like the average VAX instruction takes 4 1/2 cycles on a uVAXII. (Compare with the VAX9000 which appears to be just over 1 cycle/instr). John Hascall / hascall@atanasoff.cs.iastate.edu