IBM RS/6000 MIPS ratings
The UCS Stat/Math Center
I'm working with the usage reports for three clusters of RS/6000's and
I'm trying to standardize the measure of CPU usage to MIPS for several
different platforms but I'm told by the system manager that IBM only
reports the SPECfp92 ratings and not MIPS any ideas?
These are the ratings I need to convert:
Machines SPECfp92 rating * ? = MIPS
---------- --------------- -----
cluster #1 121.1 ?
cluster #2 97.6 ?
cluster #3 134.6 ?
Thanks ahead of time.
--
UCS Stat/Math stat...@ucs.indiana.edu
UCS Center for Statistical and Mathematical Computing
618 E. Third Street Phone 855-4724
Michael Kraemer
Organization: GSI, Darmstadt, Germany
Subject: Re: IBM RS/6000 MIPS ratings
Keywords:
I'm wondering why you want to have MIPS. Specmarks are a much better measure
of system performance because they test a mixture of "real life" programs.
MIPS only measures sort of a maximum "naked" processor speed which is totally
useless in real life. See the hype about the Alpha: lots of MIPS, but at the
bottom line Alpha _systems_ aren't faster than IBMs or HPs.
Jerry Leslie
: Message-ID: <1994Sep1...@rzri6f.gsi.de>
I thought "MIPS" stood for "Meaningless Instructions Per Second" :-)
Since we will use RS/6000s, if our clients prefer, what RS/6000 compares
to the DEC ALPHA 3000/900 ?:
Key characteristics of the 3000/900:
CPU DECchip 21064
Clock Rate 275 MHz
SPECint92 189
SPECfp92 264
Memory 32 Mbytes to 1 Gbyte
Cache 2 Mbytes
Storage 2 Fast SCSI-2 controllers at 10 Mbytes per second
I/O Slots 6 TURBOchannel slots at 100 Mbytes per second
Graphics (T/C option) 8/24 plane; 3D 24 plane
Monitor 21 inch color
Thanks in advance,
--Gerald (Jerry) R. Leslie
Staff Engineer
Dynamic Matrix Control Corporation (my opinions are my own)
P.O. Box 721648 9896 Bissonnet
Houston, Texas 77272 Houston, Texas, 77036
713/272-5065 713/272-5200 (fax)
gle...@isvsrv.enet.dec.com
jle...@dmccorp.com
==============================================================================
BTW, DEC has announced faster versions of the AXP chips:
Alpha AXP 21164-300MHz : 330 SPECint92, 500 SPECfp92 (estimated)
Alpha AXP 21164-266MHz : 290 SPECint92, 440 SPECfp92 (estimated)
Michael Kraemer
Organization: GSI, Darmstadt, Germany
Subject: Re: IBM RS/6000 MIPS ratings
Keywords:
Should be something like the 590, at 120 SpecInt, 260 SpecFP, appr. 70 Mhz.
Floating point like the alpha, integer some 30% less. However, it is nearly
one year "old" now.
BTW, I don't look at _chips_, let's see their _systems_.
Robin D. Wilson
:I'm trying to standardize the measure of CPU usage to MIPS for several
:different platforms but I'm told by the system manager that IBM only
:reports the SPECfp92 ratings and not MIPS any ideas?
Since I'm sure you've been told about 1000 times by now, MIPS is no longer a
useful measure of system (or even CPU) performance. A CPU instruction can
mean so many different things from chip-to-chip that to measure how fast one
completes a given 'instruction' is a worthless measure of how fast the system
will operate on real world stuff.
But, in case you were wondering...
The CPU on an RS/6000 can complete multiple instructions per clock cycle.
The Power2 architecture can complete something like (up to) 4 instructions
per cycle. The PowerPC is similar, but may be less. The older RIOS (Power?)
architecture could complete (up to) 2 instructions per cycle. This all
depends on which instructions, and what they are attempting to do... So, to
calculate MIPS for a given CPU architecture, you would multiply the clock
speed by the (theoretical) average number of instructions completed per cycle
-- on some machines the clock runs at 80Mhz, on some it runs at 60Mhz (and on
others...) .
Please note the difference between "can complete" and "does complete"...
Real world use may show that only .5 instructions are being completed per
clock cycle -- depending on the application and supporting hardware.
--
-----------------------------------------------------------------------------
*** These are my opinions... Mine! All Mine! Minemineminemineminemine! ***
-----------------------------------------------------------------------------
Robin D. Wilson ro...@pencom.com Pencom Software
701 Canyon Bend Dr. 9050 Capital of Texas Hwy
Pflugerville, TX 78660 Austin, TX 78759
Michael Kraemer
Organization: GSI, Darmstadt, Germany
Subject: Re: IBM RS/6000 MIPS ratings
Keywords:
In article <356lu7...@usenet.pa.dec.com>, nei...@nestvx.enet.dec.com (Burkhard Neidecker-Lutz) writes:
|> In article <354oup$g...@rs18.hrz.th-darmstadt.de> kra...@rzri6f.gsi.de (Michael Kraemer) writes:
|> >|> Key characteristics of the 3000/900:
|> >|>
|> >|> SPECint92 189
|> >|> SPECfp92 264
|> >
|> >Should be something like the 590, at 120 SpecInt, 260 SpecFP, appr. 70 Mhz.
|> >Floating point like the alpha, integer some 30% less.
|>
|> Must be some new math:
|>
|> 189/120 = 1.575 or 57% better integer performance
|>
OK, if you like to argue with me on some silly numbers, that's how I did it:
1 - 121.6 / 189 = 1 - 0.64338624 = 0.35661376 = 35.661376%
I apologize for dropping 5.661376 % in favor of the IBM.
|> >BTW, I don't look at _chips_, let's see their _systems_.
|>
|> Which is fine for the 21164 numbers posted, but the 3000/900 decidedly is
|> a shipping system.
|>
Sure, but it doesn't look much greater than the IBM.
regards,
Michael
Burkhard Neidecker-Lutz
>|>
>|> SPECfp92 264
>
>Should be something like the 590, at 120 SpecInt, 260 SpecFP, appr. 70 Mhz.
>Floating point like the alpha, integer some 30% less.
Must be some new math:
189/120 = 1.575 or 57% better integer performance
>BTW, I don't look at _chips_, let's see their _systems_.
Which is fine for the 21164 numbers posted, but the 3000/900 decidedly is
a shipping system.
Burkhard Neidecker-Lutz
GLASS Project, CEC Karlsruhe
Advanced Technology Group, Digital Equipment Corporation
nei...@nestvx.enet.dec.com
Burkhard Neidecker-Lutz
> <DEC 3000/900 and IBM 590 numbers deleted, gist being that they
> are equal on FP and the DEC is about 36% faster on integer code>
>Sure, but it doesn't look much greater than the IBM.
Mmh. The 590 costs $57,500, the 3000/900 $38,000. Faster and quite
a bit cheaper. But then it always has been somewhat more expensive
to buy IBM.
Burkhard Neidecker-Lutz
GLASS Project, CEC Karlsruhe
Advanced Technology Group, Digital Equipment Corporation
nei...@nestvx.enet.dec.com
"Alpha - The Next Generation: 21164-300, SPECint92 330, SPECfp92 500"
Ronald S. Woan
There are good reasons for not reporting the infamous Misleading
Indicator of Processor Speed, mainly because the numbers are virtually
worthless in trying to determine relative performance across differing
architectures. The SPEC suites are far better for the technical
and development computing environment.
I don't think there is any way to fairly translate to any type of MIPS
measurement from just SPECfp, and people compute MIPS in a number of
differing fashions.
comp.arch and comp.benchmarks could give you a better idea of fairer
measures of system or processor performance.
--
+------All Views Expressed Are My Own And Not Necessarily Shared By IBM-----+
+ Ronald S. Woan (IBM VNET)WOAN AT AUSTIN, wo...@exeter.austin.ibm.com +
+ outside of IBM wo...@austin.ibm.com or wo...@cactus.org or r.w...@ieee.org +
+ other wo...@csua.berkeley.edu Prodigy: XTCR74A Compuserve: 73530,2537 +
Tony Gast
|>
|> There are good reasons for not reporting the infamous Misleading
|> Indicator of Processor Speed, mainly because the numbers are virtually
|> worthless in trying to determine relative performance across differing
|> architectures. The SPEC suites are far better for the technical
|> and development computing environment.
|>
All the same, wouldn't just be fun to have the MIPS numbers? Even though they
don't mean anything, the mainframers around here don't speak SEPC, and I doubt
the mainframe CPUs are rated on SPEC. If they are, I would be interested in
the SPEC measurements for the ES/9000 CPUs.
--
Reply-to: meaddata.com!cisdfl01!cis115
Tony Gast (606)344-4528x4667
Square D
UNIX Systems Administrator (Guy, dude, etc)
Michael Kraemer
Organization: GSI, Darmstadt, Germany
Subject: Re: IBM RS/6000 MIPS ratings
Keywords:
Wrong. Back in '90 DEC had the chuzpe to charge DM 37000,- (including
Xmas discount) for a VAX 3176 with 8MB memory + gfx. At that time a comparably
equipped (gfx+memory) IBM 6000/320 was around DM 40000,-, maybe less, but 2 - 2.5 times
more powerful.
In '91 our group purchased 320Hs and DECstations 5000/xxx (ULTRIX)
at very similar prices between 30000,- and 40000,- (don't remember exactly).
For most of our applications the IBMs still were about a factor 2 faster,
the DECs had lousy graphics and lousy compilers (the FORTRAN and C DEC sold us).
Compared to AIX, ULTRIX is junk.
So who is expensive then ?
DECs price/performance may have improved since then (market pressure),
but still it is not better than their competitors.
In fact, the tons of papers I get from
IBM, HP, DEC, ... show very similar ratios.
Assuming the DEC 3000/900 (and DEC 2100) are rather new systems (months old,
the IBM 590 is about a year old now, I guess),
lets await what IBM and HP have to offer
end of September. Another round in the price war ?
Michael
Burkhard Neidecker-Lutz
> <lots of stuff about the old Digital deleted>
> Compared to AIX, ULTRIX is junk.
With more recent versions of AIX, agreed. With older versions of both,
they are both junk.
>So who is expensive then ?
IBM today.
>DECs price/performance may have improved since then (market pressure),
>but still it is not better than their competitors.
It is quite a bit better.
>In fact, the tons of papers I get from
>IBM, HP, DEC, ... show very similar ratios.
>Assuming the DEC 3000/900 (and DEC 2100) are rather new systems (months old,
>the IBM 590 is about a year old now, I guess),
August 94 and April 94, respectively.
>lets await what IBM and HP have to offer
>end of September. Another round in the price war ?
We'll see.
Burkhard Neidecker-Lutz
GLASS Project, CEC Karlsruhe
Advanced Technology Group, Digital Equipment Corporation
nei...@nestvx.enet.dec.com
"August 94: DEC 7000/700, SPECint 193.8, SPECfp 292.6, 275 Mhz 21064A
The Next Generation: 21164-300, 330 SPECint92, 500 SPECfp92"