BM Seer Unofficial thoughts from an anonymous Sun employee

Woodcrest lagging on GHz gain?

Thursday Nov 02, 2006

Woodcrest having issues scaling with GHz? On the SPECfp_rate2000 result website, when Woodcrest goes from 2 GHz to 3 Ghz (50% clock increase), but the SPECfpRate only adds 17.8%. Not good, is it?

System Chip/GHz Config Score
Fujitsu Siemens CELSIUS R540 5160 3.0 GHz 4-core 2-Socket 80.6
Fujitsu Siemens CELSIUS R540 5150 2.66 GHz 4-core 2-Socket 77.4
Fujitsu Siemens CELSIUS R540 5130 2.0 GHz 4-core 2-Socket 68.4
I'm thinking memory latency is an issue? Tomorrow, I'll look at scaling as you add cores, maybe an issue there?

Disclosure

Fujitsu Siemens CELSIUS R540 2.0GHz (2chips,4cores), 68.4 SPECfp_rate2000, Fujitsu Siemens CELSIUS R540 2.66GHz (2chips,4cores), 77.4 SPECfp_rate2000, Fujitsu Siemens CELSIUS R540 3.0GHz (2chips,4cores), 80.6 SPECfp_rate2000. SPEC, SPECint reg tm of Standard Performance Evaluation Corporation. Results from www.spec.org as of 11/2/06.

[5] Comments
Like this post? del.icio.us | furl | slashdot | technorati | digg
Posted at 04:58PM Nov 02, 2006 in Competitor Follies
Tags:
Comments:

It will get even worse with the "quad core" Clovertown. First, Clovertown is not a true quad core, it is just two Woodcrests in multi-chip module. Cache to cache transfers have to go out to the external chipset. Bandwidth per processor is cut in half. The clock rate will be reduced. Early benchmarks for Kentsfield (the desktop "quad-core") show at best a 40% gain in performance. On server apps, the fastest Clovertown will likely provide 50% to 60% more throughput, while requiring 60%-65% more power (130W).

My guess is Clovertown will be about as successful as the very first dual-core Pentiums and Xeons: Not very. In fact, because Clovertown will likely be underwelming, it may give Intel a black eye. Woodcrest is a darn good core, albeit with a feeble interconnect, but the latter is masked by it two-socket implementation. It is a pretty decent two-socket x86 processor, and has recovered Intel's reputation. Clovertown might kill the goodwill Woodcrest has developed.

Posted by Mark on November 02, 2006 at 08:14 PM PST #

Oh my gawd! First Sun managed to beat over than year-old IBM results in OMP2001 and now this! Why everybody seem to be sooo upset that Clovertown/Kentsfield are not 'true quad core' (you probably meant 'on a single die') ? IBM does their MCMs for ages and they perform great! The problem is not the package, but that all those core share single FSB so now you can have 4 cores stuck on memory access rather than one or two. Too bad Intel engineers (or their marketing division) do not learn from the past. I event think it could be more interesting if they made it 2-cores but each core would be 4-threaded like small Niagara. On the other hand, price-wise, the idea of squeezing more threads/cores into one package looks more interesting than rebranding Opterons into 4x4.

Posted by Mike on November 04, 2006 at 02:21 AM PST #

I think it would be really interesting to see how the UltraSPARC T1 memory subsystem scales with cores or frequency. Can you post some memory-intensive floating-point-intensive benchmark scores on those processors?

Posted by Xray on November 04, 2006 at 06:43 AM PST #

...on the OMP result, well IBM doesn't keep posting on newer systems, also they tend to avoid benchmarks once Sun puts out a good one, I'm still waiting for tons of results.

XRAY, The issue really isn't scaling it is about total system performance, but it seems Intel-based vendors love to talk single thread results and get everyone to assume scalability.

Posted by BM Seer on November 06, 2006 at 09:40 AM PST #

And your approach is to cherry-pick from Sun's product lines depending on the argument, to contrast against Woodcrest. For power, you quote T2000 numbers. For floating point and memory intensive workloads, you quote Opteron numbers. Let's see floating point results on T2000 or power measurements on Opteron. Face it, you don't have a product that can compete on all fronts with Woodcrest.

Posted by Xray on November 07, 2006 at 08:22 PM PST #

Post a Comment:
Comments are closed for this entry.