hoberoi@eagle.wesleyan.edu (06/26/91)
Hi, here goes: any comparisons of RS6000 530/540 etc with the HP Apollo 9000 series 700 machines ? HP claims better SPECmarks for all the comparable models SPEC IBM 320 HP 720 IBM 530 HP 730 mark 24.6 55.5 32 72.2 int 16.3 39.0 20.4 51.0 fp 32.4 70.2 43.4 91.0 I would be interested in the performance AIX vs HP-UX. How better/worse is the OS. graphics- IBM offers the SGI Personal Iris board for the 500 series machines. HP has the T1/T2 based boards. How do the two compare ? thanks Himanshu hoberoi@beaver.wesleyan.edu
dave@visual1.jhuapl.edu (Dave Weintraub) (06/26/91)
In article <1991Jun25.160925.53455@eagle.wesleyan.edu>, hoberoi@eagle.wesleyan.edu writes: |> Hi, |> here goes: |> any comparisons of RS6000 530/540 etc with the HP Apollo 9000 series |> 700 machines ? |> |> HP claims better SPECmarks for all the comparable models |> |> SPEC IBM 320 HP 720 IBM 530 HP 730 |> mark 24.6 55.5 32 72.2 |> int 16.3 39.0 20.4 51.0 |> fp 32.4 70.2 43.4 91.0 |> |> |> I would be interested in the performance AIX vs HP-UX. How better/worse |> is the OS. |> |> graphics- IBM offers the SGI Personal Iris board for the 500 series |> machines. HP has the T1/T2 based boards. How do the two compare ? |> |> thanks |> Himanshu hoberoi@beaver.wesleyan.edu |> No answer, but a comment: Beware of HP's claims. Their machine is *hot*, but they tend to be into hyperboil (?sp). See Dvorak's column in PC Magazine, where he reports HP's claims of a 720 vs a Cray, and interprets these with a wise ton of salt. The HP videotape I saw cited a SAS developer compairing SAS on the HP versus SAS on a 3090-600E. Only problem is, SAS/C (in which SAS is now written) does not, to the best of my knowledge, take advantage of the IBM vector facility, multitasking, or QSAM chaining; the comparison was a little skewed. I would also use the caveats suggested by Dvorak, in terms of scalability of the comparisons. Mind you, I repeat: the 720 is a *hot* machine! And the PC simulator beats pcsim by miles.
jackv@turnkey.tcc.com (Jack F. Vogel) (06/26/91)
In article <1991Jun25.214124.29573@aplcen.apl.jhu.edu> dave@visual1.jhuapl.edu (Dave Weintraub) writes: [ various numbers and statistics deleted...] >Beware of HP's claims. Their machine is *hot*, but they tend >to be into hyperboil (?sp). ^^^^^^^^^ Sounds like a salesweenie's social disease :-}. Suppose they have to go in eventually and get it lanced :-}!! >The HP videotape I saw cited a SAS developer compairing SAS on the >HP versus SAS on a 3090-600E... >Mind you, I repeat: the 720 is a *hot* machine! But, but, but.... it couldn't be as "*hot*" as a 3090, after all it doesn't have to be water-cooled :-}! Sorry folks, its been a long day and I needed a giggle! Disclaimer: The usual applies. -- Jack F. Vogel jackv@locus.com AIX370 Technical Support - or - Locus Computing Corp. jackv@turnkey.TCC.COM
de5@ornl.gov (Dave Sill) (06/27/91)
In article <1991Jun25.214124.29573@aplcen.apl.jhu.edu>, dave@visual1.jhuapl.edu (Dave Weintraub) writes: > >Beware of HP's claims. Their machine is *hot*, but they tend >to be into hyperboil (?sp). See Dvorak's column in PC Magazine, >where he reports HP's claims of a 720 vs a Cray, and interprets these >with a wise ton of salt. It's clear from this column# that Dvorak doesn't know his asymptote from a hole-in-the-graph when it comes to benchmarks. He presents *no* first hand data to back up his claims, sheds no light on what the referenced benchmark, AN-SYS something-or-other, measures, and then proceeds to claim it has no basis or relevance. He uses arguments like "this benchmark indicates that the HP is half as fast as a Cray, and *I* know that isn't the case, so this result is bogus." What a crock. He says how he "was told" that the Cray's time on the test was nearly all system overhead, and that if the test were lengthened 100 times, its time might not increase whereas the HP's time would likely be 100 times greater. What garbage. He concludes with a comment from an unnamed CISC weenie that RISC machines are only marginally faster than equivalent CISC machines, as if neither of these guys (Dvorak nor his source) had a stake in CISC. Sheesh. I suggest Dvorak's readers interpret his column with a carload of salt. Or better yet, skip it favor of a column with some meat in it. Can anyone shed any more light on this AN-SYS benchmark and what it measures? # I received a copy of a Dvorak article on this topic electronically. I'm assuming it's substantially the same as the one referenced above. -- Dave Sill (de5@ornl.gov) Tug on anything in nature and you will find Martin Marietta Energy Systems it connected to everything else. Workstation Support --John Muir
wdh@hrshcx.csd.harris.com (W. David Higgins) (06/27/91)
In article <1991Jun26.191020.26093@cs.utk.edu> Dave Sill <de5@ornl.gov> writes: > >Can anyone shed any more light on this AN-SYS benchmark and what it >measures? > I'm not a user of ANSYS, but I've been involved in porting the program to several platforms. Take all of this with a small chip of salt... ANSYS is a finite element analysis program, used to model physical structures and the stresses placed upon them. SP-3 is a "moderate sized 3-D solid statis analysis of a pressure vessel containing 1020 eight node solid elements". ANSYS is a product of Swanson Analysis Systems Inc. (SASI), Houston PA. From what I remember ANSYS in general, and solving SP-3 in particular, spends much of its time doing linear algebra (dot product, etc.). I suspect a machine that does well on the Linpack benchmark would do equally well with ANSYS. I don't believe Dvorak's column gave SP-3 times for Intel x86 boxes; let me remedy that by giving some times from SASI's 03/20/91 benchmark report: SP-3: Machine CPU Elapsed Comments -------- ---- ---- ------------------------------ CRAY-2 27 29 One processor used for timings CRAY Y-MP 8/128 17 17 One processor used for timings IBM 6000/540 68 70 ALR 486/33mHz 360 360 HP 486/25mHz 550 550 Gateway 386/33mHz 804 804 No wonder Dvorak conveniently forgot to include x86 times. There is some small truth to Dvorak's claim that SP-3 doesn't give the Cray a chance to strut it's stuff, although it is probably vector length and not system overhead that is involved. SASI's LS4 benchmark problem is considerably larger in size than SP-3. Some times follow: LS4: Machine CPU Elapsed -------- ---- ---- CRAY-2 647 699 IBM 6000/540 4139 4656 Notice that the IBM-6000 took 2.5x the Cray running SP-3, but the ratio grew to 6.4x when solving LS4. LS4 is a _large_ problem; SASI says 600mb of disk is required to complete the run. The benchmark reports I am quoting from did not include HP snake times, so I cannot give the HP times for LS4. Dvorak gave SP-3 CPU times of 49 seconds for the HP 9000/730, and 68 seconds for the HP 9000/720. I hope this information is useful. -- -- W. David Higgins wdhiggins@hrshcx.mkt.csd.harris.com -- Harris Computer Systems, Ft. Lauderdale, FL 33309 305-973-5351
fangchin@leland.Stanford.EDU (Chin Fang) (06/28/91)
In article <1145@hrshcx.csd.harris.com>, wdh@hrshcx.csd.harris.com (W. David Higgins) writes: |> In article <1991Jun26.191020.26093@cs.utk.edu> Dave Sill <de5@ornl.gov> writes: |> > |> >Can anyone shed any more light on this AN-SYS benchmark and what it |> >measures? |> > [some stuff deleted] |> ANSYS is a finite element analysis program, used to model |> physical structures and the stresses placed upon them. SP-3 is |> a "moderate sized 3-D solid statis analysis of a pressure vessel |> containing 1020 eight node solid elements". ANSYS is a product |> of Swanson Analysis Systems Inc. (SASI), Houston PA. |> [some benchmarks deleted] As an experienced aerospace engineer (five years at two USA's primier avaition companies) and a long time user of many Finite Elment Analysis can programs, I would like to add more to above. ANSYS is capable of using the Finite Element Method to solve engineering programs in (1) static structural analyses, like landing/crach loads on the wing of an fighter both in linear and non-linear formulations, (2) structural dynamic analyses, like the the manovering loads on the wing of a F15, (3) heat transfer problems, like the temperature effect on the nozzle of a rocket. and many physical problems that can be mapped into (1)-(3) above. It's probably the secondly popular FEM package used in Aerospace/Automotive/ Civil/Neuclear/.. The primier pkg is the enhanced commerical version of the formally NASA sponsored NASTRAN, called MSC/NASTRAN, (MSC, MacNeal and Schwandler, both were heavily involved in the original NASA NASTRAN project). The core of such pkgs is a set of numerical linear algebra routines and usually both Swanson and MSC program them to suite the specific platform they are intended to run on. So other than the smartness of programmers, the differences among the hardwares' capabilities of handling numerical linear algebra problems play a dominant role as well. eg. MSC has a vector version of MSC/NASTRAN for IBM 3090. Such can programs are usually front-ended by a graphics preprocessor for building a math model interactively, and back-ended by a graphics postprocessor for viewing TONs of results generated by such pkgs :-( It's my experience that in solving a problem, typically the human time is the longest portion of the entire project time (modeling the physical structure, checking out the math model, interpreting results, meetings, and yes, reports) so such benchmarks are seldom of *VERY* much importance even though all of us probably would like to run the darned problem on a fast machine. As a example, a rocket nozzle may take several weeks to model, maybe hours to solve on a typical powerful technical workstation like RISC 6000, and another few weeks to prepare everything else I mentioned above. As to comparing 486 boxes to technical workstations like RISC 6000s, I think it is better for joke than for reality, as a CISC processor with some hybrid pipe- line features built in, even at 33Mhz, it offers less than 2 MFLOPs. Even a wimpy SPARC 1+ can handily beat it in this regard. This leads me to... a 40Mhz i860 offers 80 MFLOPs. Can IBM Austin people design a addon board to utilize i860 for us? After all, IBM owns part of Intel and if RISC 6000 lines are designed for technical users, I would say the more MFLOPS the bigger your market share will be :-) Now this is THE MOTIVATION for IBM :-) 80 MFLOPS! Beat Cray if you want (you can already, the hardware is there :-). Sincerely, Chin Fang Student Unix System Administrator Academic Information Resources Stanford University grad student Mechanical Engineering Department Stanford University fangchin@leland.stanford.edu
mccalpin@perelandra.cms.udel.edu (John D. McCalpin) (06/28/91)
>>>>> On 27 Jun 91 19:42:24 GMT, fangchin@leland.Stanford.EDU (Chin Fang) said:
Chin> This leads me to... a 40Mhz i860 offers 80 MFLOPs. Can IBM
Chin> Austin people design a addon board to utilize i860 for us?
Chin> After all, IBM owns part of Intel and if RISC 6000 lines are
Chin> designed for technical users, I would say the more MFLOPS the
Chin> bigger your market share will be :-) Now this is THE MOTIVATION
Chin> for IBM :-)
A 40 MHz RISC-System/6000 can also do 80 MFLOPS, and IBM seems to have
done a much better job at writing good compilers for their CPU than
Intel has.
I have important pieces of my fluid dynamics codes that run at over 25
MFLOPS on the Model 320 and over 60 MFLOPS on a Model 550.
Unfortunately not all of the code runs that fast yet --- there are
still memory bandwidth bottlenecks to deal with....
If someone wants to build a "Real (tm)" number-crunching machine, then
it is going to have to support 2 64-bit loads, 1 64-bit store, 1
64-bit FP add, 1 64-bit FP multiply, and 1 logical operation/branch
per cycle. Right now if you want one of those, you will have to call
Cray Research....
--
John D. McCalpin mccalpin@perelandra.cms.udel.edu
Assistant Professor mccalpin@brahms.udel.edu
College of Marine Studies, U. Del. J.MCCALPIN/OMNET
mas@everest.corp.sgi.com (Michael A. Schulman) (06/28/91)
In article <1145@hrshcx.csd.harris.com>, wdh@hrshcx.csd.harris.com (W. David Higgins) writes: |> |> SP-3: |> Machine CPU Elapsed Comments |> -------- ---- ---- ------------------------------ |> CRAY-2 27 29 One processor used for timings |> CRAY Y-MP 8/128 17 17 One processor used for timings |> IBM 6000/540 68 70 |> ALR 486/33mHz 360 360 |> HP 486/25mHz 550 550 |> Gateway 386/33mHz 804 804 |> |> |> give the HP times for LS4. Dvorak gave SP-3 CPU times of 49 seconds for |> the HP 9000/730, and 68 seconds for the HP 9000/720. |> It is important to look at the Elapsed times, not just the CPU times. If a system is not tuned correctly, although the CPU times may be low, the total time from start to completion may be high which is what someone doing similiar work probably cares about. From the 5/7/91 ANSYS Report, the SP-3 times for the HP are CPU ELAPSED 720 66 121 730 48 109 Michael Schulman Silicon Graphics Computer Systems 2011 N. Shoreline Blvd. Mountain View, CA 94039 415-962-3308 mas@sgi.com
jfh@rpp386.cactus.org (John F Haugh II) (06/28/91)
In article <1991Jun25.160925.53455@eagle.wesleyan.edu> hoberoi@eagle.wesleyan.edu writes: >Hi, > here goes: > any comparisons of RS6000 530/540 etc with the HP Apollo 9000 series > 700 machines ? > > HP claims better SPECmarks for all the comparable models You need to compare some HP model against the 550 as well. I won't go into the dangers of comparing today's hot box against last years hot box, but try to keep in mind that with technology changing as rapidly as it does, you really have to compare the latest technology against the latest technology. You also have to compare bang/$$$ as well. Merely stacking the lowest performance HP against the lowest performance IBM machine is neither accurate nor fair (but commonly does pass for a marketing strategy ...) -- John F. Haugh II | Distribution to | UUCP: ...!cs.utexas.edu!rpp386!jfh Ma Bell: (512) 255-8251 | GEnie PROHIBITED :-) | Domain: jfh@rpp386.cactus.org "UNIX signals are not interrupts. Worse, SIGCHLD/SIGCLD is not even a UNIX signal, it's an abomination." -- Doug Gwyn
jlitvin@morticia.intel.com (John Litvin ~) (06/29/91)
In article <1991Jun27.194224.29879@leland.Stanford.EDU> fangchin@leland.Stanford.EDU (Chin Fang) writes: > This leads me to... a 40Mhz i860 offers 80 MFLOPs. Can IBM Austin > people design a addon board to utilize i860 for us? After all, IBM owns > part of Intel and if RISC 6000 lines are designed for technical users, I > would say the more MFLOPS the bigger your market share will be :-) Now this > is THE MOTIVATION for IBM :-) IBM owns no Intel stock. During the nasty years in the mid 80's, IBM did buy a chunk of Intel's stock (as a sort of loan to make sure that the 80386 would come out), but as soon as Intel started making money again, IBM sold all of the shares. John Litvin Intel Corp. jlitvin@morticia.intel.com