BM Seer Unofficial thoughts from an anonymous Sun employee

Not comparing E25k & p595

Tuesday Apr 10, 2007

I would compare a 144 core Sun E25K vs. a 64-core IBM p595.

IBM always pitches the more expensive slower p595 against an E25K, and thinks just because it has 64-cores it is better. Why?

  • The Sun Fire E25K 144-core is less expensive than a 64-core IBM p595, on the same application.
  • The Sun Fire E25K 144-core is faster than a 64-core IBM p595, on the same application.
  • Sun Fire E25K even has newer processors than used in the TPC-H result below. Last week we announced 1.95GHz.
  • I'm not sure but the Sun Fire E25K 144-core may even be smaller than the 64-core IBM p595.

    Hey wait they aren't comparable! I guess you buy IBM if you want something older, more costly, and slower. Aren't you glad you can still go to IBM and buy fewer cores. :)

    For some reason IBM thinks 64-cores sounds cheaper and faster than 144 cores, but it isn't true.

    Core count doesn't matter, systems matter! The straight scoop comes from Sun.

      IBM thinks it is about the core count or performance per core. Get real. It is about the whole system. You can do the math based on the info in the TPC-H submissions below...
      Sun: $4,207,126 /144 core = ?
      IBM: $5,358,874 /64 core = ?

      It is clear to see that IBM's cores each cost more than 2.5 times more than Sun's cores. Before you get too confused with 'rotten-to-the-core-math', just remember this. The IBM system costs more and the IBM system is a slower on the TPC-H benchmark. http://blogs.sun.com/bmseer/entry/database_world_record_sun_us.

      • The Sun Fire E25K 1.8GHz outperformed the IBM p5-595 (Power5+) by 14% and also had 31% better price/performance. Also beat the p595 by 26% on the multi-user test (Throughput).
      • The Sun Fire E25K beat the HP Integrity Superdome (Itanium2) by 60% on performance and 34% on price/performance. Sun also beat the Itanium2 Superdome by 72% for the multi-user test (Throughput).
      • Last week Sun announced Sun Fire E25K systems with 1.95GHz processors.

      TPC-H Disclosure Statement:

      Sun Fire E25K 114,713.7 QphH@3000GB, $36.68/QphH@3000GB, avail 04/09/07, HP BladeSystem ProLiant BL25p cluster 64p DC 110,576.5 QohH@3000GB, $37.80/QphH@3000GB avail 06/08/06, Sun Fire E25K 105430.9 QphH@3000GB, $54.87/QphH@3000GB, avail 01/23/06, IBM eServer p5 595 100,512.3 QphH@3000GB, $53.32/QphH@3000GB, avail 03/01/06, HP Integrity Superdome 71,847.8 QphH@3000GB, $55.79/QphH@3000GB, avail 01/18/06, Sun Fire E25K 59,435.7 QphH@3000GB, $100.66/QphH@3000GB, avail 07/27/05, TPC-H, QphH, $/QphH tm of Transaction Processing Performance Council (TPC). More info www.tpc.org.

      Several anon commenters keep bringing up licensing per core. Yes, Sun has more cores, but the pricing listed in this posting accounts for that. Sun is lower cost even with more than double the cores of IBM.

    Like this post? del.icio.us | furl | slashdot | technorati | digg
    Posted at 08:52AM Apr 10, 2007 in Benchmarks
    Tags:
    • Comments:

    I think everyone knows the number of cores matters because of software license fees? It's strange you've been talking about the number of cores for couple blogs now without mentioning that.

    Posted by taos on April 10, 2007 at 10:00 AM PDT #

    IBM likes to compare single-threaded performance because POWER5 is faster in this mode than to have SMT enabled but with one thread running.

    The number of rename registers, issue queues, the LRQ, and the SRQ available to in a thread in SMT mode are less than that in single-threaded mode.

    Now you understand why there is no POWER5 SPECint_rate2006 or SPECfp_rate2006 result!?

    Reference: POWER5 system microarchitecture -- Single-threaded operation

    Posted by Grid on April 10, 2007 at 10:20 AM PDT #

    In a response to the 1st comment. The TPC-H result and the price quoted includes HW, service, and Oracle licensing cost -- Sun is lower price when you wrap everything together.

    2nd comment. Huh? OK great then there is no reason to buy any IBM system larger than an IBM System p5 505 Express (1 processor). Because their POWER5+ design is only good at single-threaded. SPECint_rate2006 measures full system, it sounds like from the IBM design is that just isn't good for big complex applications. The second comment validates that Sun designed so that a whole system runs well, not just so one-core works well. I guess that is why IBM doesn't discuss full systems, just the cores. It would be like an auto company not discussing the car, but focusing on the valve stem.

    For big databases and big applications it is pretty clear Sun solves customer's problems. IBM focused so much on the core that it designed systems that ended up being expensive, old, and slow for whole systems.

    Posted by BM Seer on April 10, 2007 at 10:46 AM PDT #

    Grid, the "because" doesn't make sense in your first statement. and I still don't understand why P5 doesn't have SPEC*rate*: they can run it non-SMT mode if necessary, actually, I expect better results with SMT mode, because SMT helps scale.

    Posted by taos on April 10, 2007 at 10:55 AM PDT #

    BM, you could be a little "fair and balance", you know? I see too much emotion if your posts :) Even if what you said were all true, there're still lots of places for faster-ingle-core design, you know.

    Posted by taos on April 10, 2007 at 10:59 AM PDT #

    Right if you need a faster single core design then you buy small systems. If the faster single core helped full systems then IBM with it's SMT would publish an SPECint_rate2006 to beat the Sun Fire E25K. ...as you can see from the postings in the last week. Sun is getting more an more evidence of Sun E25K clearly beating IBM p595 performance on systems that cost the same (actually IBM is more expensive).

    I really don't understand the 'smoke and mirrors' of IBM saying it has fewer cores so therefore it is necessarily better. Get'em all busy on a workload and show the comparison.

    Posted by BM Seer on April 10, 2007 at 01:19 PM PDT #

    If it is all about the system then why the price per core calculation?

    Posted by rick jones on April 10, 2007 at 01:34 PM PDT #

    I showed the price/core calculation to show everyone that the hardware price per core is all over the place. Vendors implements cores differently and vendors price cores very differently. So just because the word 'core' looks like it refers to the same thing it doesn't. We've all learned that a 3.3GHz processor of one type isn't necessarily faster than 2.1GHz in a different processor.

    Counting cores is as useless as counting number of valves in a automobile engine. Just because my friend has 30 valves doesn't mean his car is faster or better because my engine vendor decided to use 24 valves that are bigger. It really depends on implementation and design.

    Posted by BM Seer on April 10, 2007 at 01:48 PM PDT #

    There seem to be a number of fallacies in your posting: 1. Lots of software is licensed per core - therefore perf/core <--> performance/licensing dollar. And for big systems, software is typically more than hardware...by a lot. 2. SPEC_rate is NOT A SYSTEM BENCHMARK. It doesn't use disks at all, nor does it have any data sharing.

    Posted by Anon on April 16, 2007 at 01:46 AM PDT #

    Post a Comment:
    Comments are closed for this entry.