Highest SPECfp on the Planet!

Please note the added required disclosure statement at the end
With the introduction of SunFire X4600 servers and new compiler patches, Sun Studio Compilers, Solaris and SunFire Opterons (4P)combine for the highest SPEC CPU2000 FP(peak value =3538) on the planet. There is some secret sauce in the newly introduced patch, plus the autopar scaling of some benchmarks made possible by this new configuration that helps overwhelm Intel's Woodcrest and even the respected IBM Power5+ number!
The interesting nugget for Intel's Woodcrest here is that its own FP number(2783), announced with the much-respected Intel Compilers, is trumped by PathScale Compilers(3056) on what appears to be a very similar box. Begone, Intel FUD!

Heres the tally of all the leading architectures and their SPEC CPU2000 FP (peak values only):
  1. SunFire X4600:     3538
  2. IBM Power5+:     3513
  3. HP/Woodcrest/PathScale:     3056
  4. Intel/Woodcrest/Intel Compilers:     2783 (using Fujitsu Siemens PRIMERGY RX300)
  5. Bull Itanium2/1600MHz:    3017
  6. Intel Pentium 4 EE 3.73GHz:    2236
  7. Fujitsu 2.16GHz SPARC64:    2236
  8. Intel Pentium 4 Xeon 3.73GHz::    2150
  9. HP 1.3GHz Alpha:    1684

My favorite site for looking up SPEC numbers is this SPEC dataminer, from Aces Hardware site.

Required Disclosure Statements:

Sun Fire X4600 3538 SPECfp2000 (4 cores, 4 chips, Solaris 10),
IBM p5 575 3513 SPECfp2000 (2.2GHz POWER5+) (1 core, 1 chip, AIX),
HP BL480c 3.0GHz Woodcrest 5160 3049 SPECfp2000 (1 core, 1 chip, RHEL4u3).
Bull Novascale 3045 3017 SPECfp2000 (1600MHz  Itanium2) (8 cores, 4 chips, Bull Advanced Server 4 (linux kernel 2.6.12, 64K pages))
Intel(R) D975XBX 2236 SPECfp2000 ( 3.73 GHz Intel Pentium(R) processor Extreme Edition 965 , 1066 MHz bus)  ( 2 cores, 1 chip, Windows XP Professional SP2)
Fujitsu Limited PRIMEPOWER650 2236 SPECfp2000 (2164MHz SPARC64 V)  (1 core, 1 chip, Solaris 10)
HP DL380 G5 2150 SPECfp2000 (3.73GHz, Intel Xeon Processor 5080) (1 core, 1 chip, Hyperthreading disabled, RHEL4)
HP AlphaServer GS1280 7/1300 1684 SPECfp2000 (1300MHz  Alpha 21364) (1 core, 1 chip, Tru64 UNIX V5.1B-1 + PK4)

SPEC, SPECfp are Registered Trademarks of Standard Performance Evaluation Corporation. Results from www.spec.org as of Jul 11, 2006.


Vijay, you need to put in the proper disclosure statement on this posting. Here is a start, but you need this for each of the competitors you list -- can't just post results and links, sorry SPEC rules (even in blogs) not mine.

Sun Fire X4600 3538 SPECfp2000 (4 cores, 4 chips, Solaris 10), IBM p5 575 3513 SPECfp2000 (2.2GHz POWER5+) (1 core, 1 chip, AIX), HP BL480c (3.0GHz Woodcrest 5160) 3049 SPECfp2000 (1 core, 1 chip, RHEL4u3). SPEC, SPECfp reg tm of Standard Performance Evaluation Corporation. Results from www.spec.org as of Jul 11, 2006.

Posted by BM Seer on July 19, 2006 at 07:18 AM PDT #

Thank you for pointing this out. I read the rules and my interpretation was that the links were sufficient. I've updated the blog to be more explicit, hopefully this will help.

Posted by Vijay Tatkar on July 20, 2006 at 04:17 AM PDT #

Is the benchmark single threaded? Why did we run on a 4 processor machine if it's single threaded? (Just for more RAM?) If it was actually using all 4 chips, doesn't that mean the P5 is 4 times faster at FP than the Opteron chip we're using? Those are pretty strange numbers you posted, Vijay. As a caveat, I know almost nothing about this subject, as you can tell from all the questions I'm asking. ;-)

Posted by Chris Quenelle on July 21, 2006 at 02:06 PM PDT #

To clarify: this is pretty standard. The SPECfp benchmarks can be run in auto-parallel (-xautopar) mode and Sun has been submitting results in autoparallel, for peak (not for base, because base has limited flags) for years now. Almost all good Fortran compilers out there implement an automatic parallelization option. Since these are not rate benchmarks, but CPU measures, the scale-up is rarely scales close to #CPUS. Typically, it improves 10-12% on SpecFP like benchmarks (and several real-life programs, BTW for 2p and decreasing over additional processors. Still, there is nothing that stops anyone from using all four processors even if the gain is only about 20% or so over single processor count. The importance of this, besides being allowed by SPEC (so, why not?) is that - it gives application writers an idea if they can exploit the extra processors on a machine. The answer is "yes" but its diminishing returns. Still, since Sun's primary market is 4p to 16p, theres nothing wrong, and everything right, with using all 4 processors - secondly, as processors transition to dual-core now and quad-core (or 8-core or more as in the case of Niagara), exploiting other cores is paramount and this is the right thing to show for these machines. BTW, our single Processor, non-autopar numbers are also better than Woodcrest, but 4p gives us a chance to trump IBM P5+. 4P, BTW, is a standard config for the new SunFire X4600 (the other is 8P, but 8P is generally seen as being in a different league, tho again, not disallowed by SPEC). Another way of saying this is that the SPECfp for this machine is different than the SPECfp for this chip. As a systems provider, its our obligation to show this difference, so this is also morally correct. All IMO, of course

Posted by Vijay Tatkar on July 21, 2006 at 04:02 PM PDT #

Post a Comment:
  • HTML Syntax: NOT allowed

I have worked with Sun and Oracle for 25 years now; in compilers and tools organization for most of these years followed by a couple of years in Cloud Computing. I am now in ISV Engineering, where our primary task is to improve synergy between Oracle Sun Systems and our rich ISV ecosystem


« July 2016
Interesting Links