A couple of weeks ago I requested information on benchmarks. I didn't
recieve a huge amount of information, in fact I got more requests
for the information than actual replies, so it appears a summary is
called for. Our purchasing desision is leaning back towards a Sparc 10,
simply because we need a solid, compatible environment, with no suprises.
The comments below have left me with the feeling that the HP and IBM
machines, whilst doubtless excellent machines, have too many rough edges
for some of our novice users. So barring some unassailable fiscal reasoning
on the part of HP or IBM, it looks like a Sparc 10. Anyone with strong
feelings, who wishes to convince me one way or the other, is welcome to
drop me email.
Thanks to:
Dana Jacobsen
Peter Baer Galvin
Ron Nash
Steve Blair
Here are the messages:
>From jacobsd@xanth.CS.ORST.EDU Thu Aug 27 00:05:18 1992
-header deleted-
> We won't be getting our hands on any Sparc 10's until mid-September (if then!),
>but you said "Any information", so I'll give you the publicly available stuff.
>
>Machine Spec92Int Spec92FP LINPACK MFLOPS
>
>
>Sun Sparc 2 21.8 22.8 4.2
>Sun Sparc 10/30 44.2 52.9 10.6
>Sun Sparc 10/41 52.6 64.7 17.2
>
>HP 9000/710 31.6 47.6 12.2
>HP 9000/720 36.4 58.2 17.9
>HP 9000/730 47.8 75.4 23.7
>
>IBM RS/6000 220 15.9 22.9 6.5
>IBM RS/6000 320H 20.9 39.4 11.7
>IBM RS/6000 340 27.2 51.9 14.8
>IBM RS/6000 350 34.6 65.0 18.6
>IBM RS/6000 520H 20.9 39.6 11.5
>IBM RS/6000 530H 28.2 57.7 20.2
>IBM RS/6000 550 35.4 71.7 25.6
>IBM RS/6000 560 42.1 85.5 30.5
>
>SGI Crimson 61.7 63.4
>
>Intel 486/50 30.1 14.0 2.2 (est)
>Sun Sparc IPC 13.8 11.1 1.8
>HP 9000/705 21.9 33.0 8.0
>HP 9000/750 48.1 75.0 23.7
>IBM RS/6000 970 47.1 93.6
>
>
> From experience with a Sparc 2, an HP 730, and an IBM 340, these are
>my impressions:
>
> Gaussian Elimination: The IBM machines, hands down. Those things are
>FAST on this (comes from having a mult+add instruction).
> General floating point: The HP 730 wins on this, with speed being about
>2-3 times the Sparc 2 for most of our code. Some code ran only 1.5 times
>faster, some up to 3 times.
> Integer Speed: HP 730 wins this also. The IBM RS6000's just aren't that
>fast on integer. The 340 just barely edges out the Sparc 2 here.
> Overall speed: The HP 730 feels the fastest. The graphics are very fast,
>the CPU is about twice the Sparc 2, and compiles are lightning fast. The
>IBM feels sluggish, and it's graphics are very poor. In my opinion, the
>RS6000 is not a good general purpose machine -- it is good for floating
>point only, and I don't think it's a good buy against the HP 730.
> Porting: Unless your code is already written for an HP or an IBM, the
>Sparc is by far the easiest to port and compile on. We haven't heard
>back from our IBM reps about the "Internal Compiler Errors" we got (after
>they spent so much time BSing about how 'easy' and 'standard' the IBMs
>were). HP/UX 8.0 is definetely an improvement over 7.0 -- they have all
>the POSIX libraries (though their include files do some non-POSIX things
>which are annoying to work around). We had quite a few programs that
>gave incorrect output on the HPs (and some fortran code which core-dumped
>their preprocessor), but nothing like the god-awful hell of the IBMs.
>
> Note that these comparisons are between two $30k machines (the HP 9000/730
>and the IBM RS6000/340) and a $15k machine (the Sun Sparc 2). From Sun's
>estimates of performance (and the Spec92 numbers), one should expect between
>two and three times the Sparc 2 performance from the 10/41, which will be
>the same price or less than the IBM and HP boxes. Since our code ran about
>2-3 times faster on the HP and IBM boxes, I believe we will get Sparc 10's
>as it solves the porting problems, leaves us with more flexibility on the
>upgrade path, and means we don't have to buy all-new licenses for our
>commercial software.
> I'll see if I can mail you an update when we get our Sparc 10's in.
>Meanwhile, summarize your results! I'm eager for more info. Thanks,
>--
>Dana Jacobsen Disclaimer: This is not an official announcement
>jacobsd@cs.orst.edu of any kind, it may contain blatent falsehoods and
>jacobsd@solar.cor2.epa.gov strong opinions that are solely the authors. This
>(God I hate politics) article may be reposted only with this disclaimer.
>From pbg@cs.brown.edu Thu Aug 27 10:00:33 1992
-header deleted-
>According to sun:
>
> SS10 Model 30 SS10 Model 41 HP9000/730 IBM 6000/350
>specint92 44.2 52.6 48.1 34.6
>specfp92 52.9 64.7 75.0 65.0
>
> --Peter
>
>------------------------------------------ --------------------------------
>Peter Baer Galvin (401) 863-7623
>Systems Manager, Brown Univ. Comp. Sci. pbg@cs.brown.edu
>Box 1910 (115 Waterman Street) uunet!brunix!pbg
>Providence, RI 02912 (02906) pbg@browncs.bitnet
>From nash@ucselx.sdsu.edu Thu Aug 27 10:17:20 1992
-header deleted-
>In article <1992Aug27.015145.18244@cs.wm.edu> you write:
>>I'm sure this was discussed ad nauseam a few months ago, but now I find myself
>>in need of the information. Specifically, I need benchmarks for the SPARCstation
>>10 models 30 and 41. Floating point benchmarks are most important to us, and
>>relative performance comparisons to the HP snakes and IBM Powerstations would
>>be useful too. Any information greatly appreciated.
>>
>>---
>>+-----------------------------------------------------------------------------+
>>|..Alastair Neil................................| |
>>|..(804)-221-3533..[ajn@physics.wm.edu].........| None Shall Sleep |
>>+-----------------------------------------------------------------------------+
>
>>From the May 1992 SunFlash Vol 41 #34
>
> Model 30 Model 41 Model 52 Model 54
> -------- -------- -------- --------
> Uni Uni 2-way MP 4-way MP
>SPECint92 44.2 52.6 58.1 per CPU
>SPECfp92 52.9 64.7 71.4 per CPU
>SPECthruput89 N/A N/A 109 (est.) 218 (est.)
>MIPS 86.1 96.2 200+ 400+
>Mflops (DP) 10.6 17.2 38 76
>TPS (est.) 100 120 180 220
>
>Hope this helps.
>
>
>--
> ,--, | Ron Nash San Diego State University
> _ ___/ /\| | nash@sdsu.edu
> ,;`( )__, ) ~ |
> // // '--; | Gin-N-Tonic Learning to be an endurance horse
> ' \ | | Luv on Fire trusty trail horse
>From sblair@upurbmw.dell.com Wed Sep 2 14:17:26 1992
-header deleted-
>Here's a chart of several current systems. I used it recently for
>a study of new machines:
>
>MACHINE SPECfp-92(*) SPECint-92(*) SPEC89 MFLOPS MIPS
>-----------------------------------------------------------------------------
>
>Solb DX4000 20.2 19.3 21.3 3.9 23.2
>(cost = $ 20K)
>Sun SS2 22.8 21.8 24.7 4.2 28.5
>(cost = $ 20K)
>Sun ss10-30 52.9 44.2 57.3 10.6 86.1
>(cost = $ 25K)
>Sun ss10-41 untested untested 71.2 17.2 96.2
>(cost = $ 28K)
>SGI 4DIndigo 61.0 57.0 70.0 17.9 85.0
>(cost = $20K)
>H/P 720 47.7 31.6 59.5 17.9 57.9
>(cost = $ 45K)
>RS6000-320H 39.4 20.9 41.2 11.7 37.1
>(cost = $ 35K)
>
>NOTES:
>------
>
>1)The Sun SS1-41 is not currently finished with the SPEC International
>evaluation testing. All numbers on the ss10-41 are *PROJECTED* values
>(from Sun) and are unknown as final numbers until the results are posted.
>The machine will not start shipping until late-October/early-November 1992.
>
>2) The Sun ss10-30 is just entering shipping volumes, with a projected
>ARO of 30 -> 45 days we are informed.
>
>3) The Sun ss10-41 is yet to enter shipping, with a current projected
>ARO of 60 - > 90 days we are informed.
>
>
>
>--
>Steve Blair DELL UNIX DIVISION sblair@dell.com
>====================================================================
>"Multiple Processors Do Not A Symetric Multiprocessing Machine Make"
--- +-----------------------------------------------------------------------------+ |..Alastair Neil................................| | |..(804)-221-3533..[ajn@physics.wm.edu].........| None Shall Sleep | +-----------------------------------------------------------------------------+
This archive was generated by hypermail 2.1.2 : Fri Sep 28 2001 - 23:06:48 CDT