Performance of Old Intel Processors
Zhiwei Xu
1. The peak MIPS rate and/or Mflop/s rate of the following processors:
8088, 4.77 MHz, as used in IBM PC/XT
80286, 6 MHz, as used in PC/AT
80386, 12 MHz (or any other clock)
80486, 20 MHz (or any other clock)
Pentium, 60 MHz, (or any other clock)
2. The SPEC-92 baseline numbers for
8088, 4.77 MHz, as used in IBM PC/XT
80286, 6 MHz, as used in PC/AT
One more related question: What are the number of flops in the SPEC92fp or
SPEC95fp benchmarks? Did anyone from SPEC ever run the codes on, say, a
Cray C-90 and use the Hardware Performance Monitor to measure the flops?
I found it difficult to use SPEC numbers to help optimize user programs.
Suppose my application runs at 30 Mflop/s on a machine with SPEC92fp=80
and 150 peak Mflop/s. The SPEC number is almost useless to me. The peak
number, on the other hand, at least tells me my code runs at 20%
utilization (or efficiency, if you rather), and there may be room for
improvement.
In other words, SPEC might be useful for vendor advertising and customer
purchasing, it does not seem to be helpful for users.
Did I miss something?
Zhiwei Xu z...@diana.usc.edu
Mike Schmit
>
>I would appreciate any information or pointer to
>
>1. The peak MIPS rate and/or Mflop/s rate of the following processors:
Here is what I have: (MIPS)
8086, 4.77 MHz .75
80286, 6 MHz, 1.5
80386, 16 5
80486, 25 27
Pentium, 60 MHz 112
>2. The SPEC-92 baseline numbers for
>
> 8088, 4.77 MHz, as used in IBM PC/XT
> 80286, 6 MHz, as used in PC/AT
SPEC would not run on those, I suspect.
Mike Schmit
-------------------------------------------------------------------
msc...@ix.netcom.com author:
408-244-6826 Pentium Processor Programming Tools
800-765-8086 ISBN: 0-12-627230-1
-------------------------------------------------------------------
Jeffrey Reilly
>
>1. The peak MIPS rate and/or Mflop/s rate of the following processors:
>
Note that it is the opinion of some that "MIPS" and "MFLOPS" have
limited value because there is no "standard" definition for what these
mean (i.e. - how are they generated, etc.)
Intel does not include this information in their latest performance briefs
(probably for the above reason in favor of "better" benchmarks, SPEC, BAPCo,
etc.) You may be able to contact Intel Literature and see if they have any
older Performance Briefs (800-548-4725).
>
>2. The SPEC-92 baseline numbers for
>
> 8088, 4.77 MHz, as used in IBM PC/XT
> 80286, 6 MHz, as used in PC/AT
Intel has never reported measured SPEC92 numbers for these processors. In my
opinion, this is probably because at the time SPEC92 was introduced a) these
processors were not performance competitive in the market (if I recall correctly,
Intel provided 80486DX-50 results at the SPEC92 introduction) and b) setting up
an environment for SPEC (UNIX and compilers) was probably a hassle, particlarly
if the same type of environment as the 80486DX-50 results was desired.
>
>One more related question: What are the number of flops in the SPEC92fp or
>SPEC95fp benchmarks? Did anyone from SPEC ever run the codes on, say, a
>Cray C-90 and use the Hardware Performance Monitor to measure the flops?
I believe that for SPEC92 this information was published in the SPEC Newsletter
Volume 3 Issue 4, December 1991. SPEC can be reached at: 703-331-0180 or
spec...@cup.portal.com for info on reprint information.
>
>I found it difficult to use SPEC numbers to help optimize user programs.
>Suppose my application runs at 30 Mflop/s on a machine with SPEC92fp=80
>and 150 peak Mflop/s. The SPEC number is almost useless to me. The peak
>number, on the other hand, at least tells me my code runs at 20%
>utilization (or efficiency, if you rather), and there may be room for
>improvement.
>
>In other words, SPEC might be useful for vendor advertising and customer
>purchasing, it does not seem to be helpful for users.
As part of SPEC's charter, it states "The founders [of SPEC] believe that the
user community will benefit greatly from an objective series of application-
oriented tests, which can serve as a common reference point and be considered
during the evaluation process." The SPEC benchmarks define a set workload which
generates a comparitive metric. THis may be useful for advertising, purchasing
and some users.
It sounds like you are trying to use SPEC as a means of predicting your
application's performance. This is not the primary goal for SPEC. In this case,
you are correct in that the SPEC numbers are probably not going to provide a
lot of value to you without more analysis (i.e. compare and contrast
the characteristics of your program to the benchmarks, run your program on
machines with a known SPEC value and try to perform some correlations, etc.).
Also, it may not be very useful to try to predict application performance
using "MIPS" and "MFLOPS" values (instruction mix, memory requirements, etc.
often have an influence that is not recognized in "peak MIPS"; it
depends on the user program but its rare to find one that approaches "peak MIPS").
The SPEC Newsletter also included an article once showing the difficulties in
doing this.
Jeff
Associate Editor, SPEC Newsletter
Jeff Reilly |"A lot of close plays at first (base) could
Intel Corporation | be eliminated by moving the base back a
jwre...@mipos2.intel.com | foot." - L. Anderson
(408) 765 - 5909 |
Disclaimer: Speaking only for myself, not my employer... use at your own risk
Hans du Buf
: 1. The peak MIPS rate and/or Mflop/s rate of the following processors:
deleted
: 2. The SPEC-92 baseline numbers for
: 8088, 4.77 MHz, as used in IBM PC/XT
: 80286, 6 MHz, as used in PC/AT
: Zhiwei Xu z...@diana.usc.edu
What's wrong with linpack :-)
I remember that a 486DX2@66MHz has 2.5 Mflops
if running linpack 100x100
Hans
Willy Geerts
>
> I would appreciate any information or pointer to
Try URL: http://hpwww.epfl.ch/bench/bench.html
I have to write a paper on benchmarks and as a complete ignorant of the
subject I found valuable information there (with very complete results
for all sorts of machines and CPU).
So, maybe that a more experienced person like yourself could find what
he is looking for ?
--
Geerts Computer Systems
Frans Baetenstraat 38/2, 2100 Antwerpen, BELGIUM
tel ,info(VRU) : +32 (0)3 3222162 fax : +32 (0)3 3663354
g...@glo.be -- http://www.glo.be/globe/users/gcs/index.html
Jacob Parnas
> 8086, 4.77 MHz .75
I find it hard to believe that a 8086 is .75 mips when a Vax 11/780 is about
1 mip. I think I saw a figure a long time ago about an 8086 being about
is about .1 MIPS.
Jacob
------------------------------------------------------------------------------
| Jacob M. Parnas |
| IBM Thomas J. Watson Research Ctr. |
| Internet: jpa...@jparnas.cybercom.net |
------------------------------------------------------------------------------
--
------------------------------------------------------------------------------
| Jacob M. Parnas |
| IBM Thomas J. Watson Research Ctr. |
| Internet: jpa...@jparnas.cybercom.net |
------------------------------------------------------------------------------
Ron Nicholson
I think early 8088 (eight not six) systems ran at 4.77 MHz.
The early 8086 based PC/AT systems ran at 8 MHz.
> I find it hard to believe that a 8086 is .75 mips when a
> Vax 11/780 is about 1 mip. I think I saw a figure a long
> time ago about an 8086 being about is about .1 MIPS.
I would believe that systems based on a 1 MHz 6502 or a 4 MHz
Z80 would be around 0.1 times the performance of an 11/780 when
running assembly language coded very small integer programs.
An 8 MHz 8086 based system was much faster than these 8-bit
micros.
My (vague) recollection was that xlisp 1.6 ran almost as fast
on an 8086 based HP Vectra as on an DEC VAX 11/780.
Someone must have a list of old dhrystone numbers for these
machines lying about. These old numbers might still be useful
for comparisons, since they predate the era of the -Odhrystone
option in optimizing compilers.
---
Ronald H. Nicholson, Jr. r...@engr.sgi.com, r...@netcom.com
#include <canonical.disclaimer> // I speak only for myself, etc. YMMV.
David A Willmore
>Jacob Parnas wrote:
>> In article <4bab22$6...@cloner2.ix.netcom.com>,
>> > 8086, 4.77 MHz .75
> ^^^^
>I think early 8088 (eight not six) systems ran at 4.77 MHz.
>The early 8086 based PC/AT systems ran at 8 MHz.
The IBM PC and XT ran a 8088 at 4.77MHz. The AT ran a 80286 at 6 MHz and
later at 8MHz. The only 8086 based machine that I'm aware of IBM producing
was the PS/2 Model 30. Later it became the Model 30/286 which sub'd a 80286
for the 8086. I believe that both of those machines ran at 10MHz.
>> I find it hard to believe that a 8086 is .75 mips when a
>> Vax 11/780 is about 1 mip. I think I saw a figure a long
>> time ago about an 8086 being about is about .1 MIPS.
>I would believe that systems based on a 1 MHz 6502 or a 4 MHz
>Z80 would be around 0.1 times the performance of an 11/780 when
>running assembly language coded very small integer programs.
>An 8 MHz 8086 based system was much faster than these 8-bit
>micros.
I have to disagree. A 1 MHz 6502 could run from .5 to .14 native MIPS. I
think that the Z80 used 9 clock cycles for its fastest instruction which
would give .44 MIPS on a good day. These are native 8 bit MIPS. There
weren't many 8 MHz 8086 systems. These older processors competed against
a 4.77 MHz 8088 which was a dog of a processor. The Z80 could beat it
as could the 6502. Now, the 80286 had a glint of a promise of being a
real processor. It could beat a VAX 11/780. The 8086 just wasn't used
in enough (popular) machines to make it worth discussion.
>My (vague) recollection was that xlisp 1.6 ran almost as fast
>on an 8086 based HP Vectra as on an DEC VAX 11/780.
I can believe this. The 11/780 was not a fast machine. The instructions
were very slow. What it had going for it was more memory and I/O than almost
any PC of the time. That goes a long way towards making a better system.
Throw in a real operating system and good compilers and it's a heck of a lot
better place to write software than a PC.
>Someone must have a list of old dhrystone numbers for these
>machines lying about. These old numbers might still be useful
>for comparisons, since they predate the era of the -Odhrystone
>option in optimizing compilers.
The problem you might have is there were never reasonable C compilers made
for these machines. Native compilers suffered from a tiny memory space
for run time and compile time. Optimization would have required more memory
and more CPU--things that they had too little of. Berkly, they guys who did
GEOS, used a cross compiled environment, but I'm not sure if it was C.
Anyone with a good cross-compiled C version of dhrystone for the 6502?
Cheers,
David
BTW, I don't speak for Motorola.
Graeme Gill
>Jacob Parnas wrote:
>I would believe that systems based on a 1 MHz 6502 or a 4 MHz
>Z80 would be around 0.1 times the performance of an 11/780 when
>running assembly language coded very small integer programs.
>An 8 MHz 8086 based system was much faster than these 8-bit
>micros.
I'd dispute that. My 6Mhz Z80 ran rings around a 4.77 Mhz 8088
when running applications. Even if this disparity was due to
system issues (disk, memory, OS and display performance), the
8088 could not be described as "much faster".
Graeme Gill.
Timothy Cullip
>Ron Nicholson <r...@sgi.com> writes:
>>Jacob Parnas wrote:
>>> Vax 11/780 is about 1 mip. I think I saw a figure a long
>>> time ago about an 8086 being about is about .1 MIPS.
>
Well MIPS mean different things to different people. If you take
the somewhat accepted view of DRYSTONE MIPS where a VAX 11/780
being about 1560 drystones/sec = 1 MIP, then the IBM PC/XT 8088 at
4.77 MHz and ~300 drystones/sec would give you 0.2 MIPS.
An 8 MHz 8086 is about 700 drystones/sec = 0.5 MIPS.
The above is taken from the comment table of the drystone source
code I got from somewhere.
--
Tim Cullip
cul...@radonc.unc.edu
Timothy Cullip
>4.77 MHz and ~300 drystones/sec would give you 0.2 MIPS.
>An 8 MHz 8086 is about 700 drystones/sec = 0.5 MIPS.
>The above is taken from the comment table of the drystone source
>code I got from somewhere.
Oh, and by the way, I just ran that code on my Pentium P100 and got
110 MIPS, seems we've come a ways with the x86. I wonder what a
200 MHz Pentium Pro would give?
--
Tim Cullip
cul...@radonc.unc.edu
bill davidsen
Jacob Parnas <jpa...@kalypso.cybercom.net> wrote:
| In article <4bab22$6...@cloner2.ix.netcom.com>,
| > 8086, 4.77 MHz .75
|
| 1 mip. I think I saw a figure a long time ago about an 8086 being about
| is about .1 MIPS.
Unfortunately CPUs aren't linearly scaled. There's a ratio for FP,
for 8, 16, 32 and 64 bit integer, etc. For a small program which fit
in 64k code + 64k data, I can believe this. For fft it's obviously
false.
I wrote a benchmark which measured all the common ops, like 16/32
bit add, mpy, etc, text and branch, transcendental functions, and
lots of other things. It gave me a profile of operations, and if I
could tell what % of a program was doing that op, I could predict
which CPU was faster. It was great fun, but unless you have a
program which does only 8 bit adds and 16 bit divides, I don't think
it is useful. For most load SPEC is a good rule of thumb.
Norbert Juffa
performance (dhry) of various Intel CPUs with 1990 C compilers. Note
that a VAX 11/785 has 1857 Dhrystones (with a 1985 C compiler I believe).
This allows us to compute so called Dhrystone MIPS (dmips), based on the
claim that a VAX 11/785 was a 1 MIPS machine. If I remember correctly,
the CPI for typical large mode real mode programs was about 19 for a 8088,
17 for a V20, 14 for a 8086, 11 for a V30 and 6 for a 286. These numbers
are the basis for the native MIPS (nmips) computations below.
machine cpu clk os compiler switches drhy dmips nmips
IBM PC 8088 4.77 PC-DOS 3.3 MS-C 6.0 /AL/Ox/Gs/G0 427 0.23 0.25
IBM PC V20 4.77 PC-DOS 3.3 MS-C 6.0 /AL/Ox/Gs/G1 500 0.27 0.28
Amstrad 1512 8086 9.54 PC-DOS 3.3 MS-C 6.0 /AL/Ox/Gs/G0 1141 0.61 0.68
Amstrad 1512 V30 9.54 PC-DOS 3.3 MS-C 6.0 /AL/Ox/Gs/G1 1505 0.81 0.87
IBM AT-02 80286,1W 6.0 PC-DOS 3.3 MS-C 6.0 /AL/Ox/Gs/G2 1404 0.75 1.00
IBM XT-286 80286,0W 6.0 PC-DOS 3.3 MS-C 6.0 /AL/Ox/Gs/G2 1717 0.92 1.00
I used the large model to compile Dhrystone since I though that would be the
fairest comparison to machines that have a address space of > 64 KB, like the
Motorola 68K CPUs or the VAX. Use of the small model would have inflated the
numbers by 30% or so.
I also remember that a 6 MHz Z80 on a Microsoft Softcard (for the Apple 2e)
gave roughly the same performance as the original 4.77 MHz IBM PC.
-- Norbert Juffa (nor...@nexgen.com)