Clovertown, Cinema and Woodcrest rock

WHILE AMD'S Dirk Meyer may argue how the Merom/Conroe/Woodcrest hype is about future products, while his team focuses on selling current products, the point is - the future is very near, almost right now, in fact.

Under the shadow of Taipei101, "the stacked lunch box", and the world's tallest skyscraper, at the local IDF this week, Intel repeated its March show-off from the San Francisco IDF. The Clovertown dual-socket, 2x4-core workstation worked well rendering the Cinema4D, and showing smooth accelerated rendering, with a single frame divided into eight equal segments. Inteleers were non-committal on the FSB speed for the dual-die thingie, even though I probed them on the 'benefits of reaching the 1333 FSB on such a bandwidth-hungry device'. But yes, they (at least the couple I spoke to) seemed to be ready to launch Clovertown (as well as its desktop cousin Kentsfield) as early as this Christmas...

Compared to the Frisco event, Woodcrest seemed to be fresher - this time, I assume, with B-stepping, closer to the final production run - those B-step Woodcrests are now everywhere in Taipei, and the vendors I spoke to (on condition of anonimity, of course) seem to be very happy with their performance, reliability and power envelope.

Now, the Intel show had a HP dual-socket Woodcrest at 3 GHz / FSB 1333 vs a Sun dual-socket Opteron 285 2.6 GHz / 1 GHz HT, both with 2 GB RAM etc, running SunGard credit analysis application. The Woodcrest completed the job 35% faster, while the power meter also showed roughly 6% less power consumed - 307 W vs 325 W. Now, this is a July Intel product beating an April AMD product.

Well OK, let's extrapolate this to a, say, September AMD product, a Socket F dual-core 2.8 GHz Opteron with DDR2-800 memory. The extra 7% clock speed, with all else the same, will give us some 5% extra application performance, while the DDR2-800 CL5 should give us another 5% boost vs DDR1-400 CL3 (usual server memory speed grades right now). So, that Opteron should be roughly 10% faster than the Sun unit in the Intel test, while the power consumption may drop a few watts too, due to DDR2 1.8 volt memory instead of DDR1 2.5 volt memory (but the speed increase would largely offset the power savings anyway).

Now, you can also expect that, instead of DDR2-667 CL5 FB-DIMMs, there soon may be a bit faster CL4 version too, giving the Intel plaform an extra, say, 1% in overall performance. Also, keep in mind that FB-DIMMs consume an extra 5W per module for the AMB chips compared to registered DDR2 server DIMMs (20W on four modules!), giving you a perspective how low-power the Woodcrest CPU itself actually is - if it used standard server DIMMs, it would have a 12% system power advantage over the Opteron.

Performance-wise, after the mentioned adjustments, the 3GHz Woodcrest would still be about 25% faster than the 2.8GHz Socket F Opteron in the abovenamed application, and still with somewhat lower power consumption. Well OK, this is an Intel-chosen app. What about more neutral stuff? Here are the expectations based on the 'crystal ball' for, say, SPECint2000 per-core results, as well as SPECint200rate integer 2-core results:

Dempsey 3.73GHz-> 1,800 SPECint2000; 43 CINTrate-peak /May
Woodcrest 3GHz -> 2,400 SPECint2000; 59 CINTrate-peak /July
Opteron 2.8GHz Socket F -> 1,900 SPECint2000 ; 45 CINTrate-peak /September

So, in summary, in these few apps (remember Intel surely singled out the SunGard one for its own benefit too), the July Woodcrest chip is expected to be noticeably ahead of anything AMD will have until K8L comes out early next year. I personally believe that Intel's new baby will be the overall performance leader in the 2-socket space among all general purpose processors for a while - until its own successor, and IBM POWER6, comes along - or unless Alpha EV9 resurrects from the dead.

If you don't believe my 'crystal ball', well too bad - wait till you see the official results! More benchmark input will surely float from the usual 'industry sources' as we come closer to July.

On the other hand, in my opinion, AMD now seems to behave like Intel did five years ago - less friendly, some would say more arrogant too, and way quieter and more secretive about what it is actually doing. I wouldn't be surprised at all to see the K8L core come out faster than expected, and "whatever-the-next-AMD-core-is" also be pushed to 5th gear to wrestle back the leadership, if possible.

Nor would I be surprised to see a 3.33GHz Woodcrest version out soon after as well, or even a HPC-specific unlocked-multiplier Woodcrest with even higher frequency potential if liquid-cooled - while Intel is against that when it comes to commercial servers, the HPC gang will go all the way in performance race (where Opteron leads now), and water coolers are cheaper and more compact than ever, so why not?

But, hey, now that the HyperTransport 3 spec is out in the open (and open to everyone), and addresses a lot of issues that Intel may not have been happy about before, I wouldn't be surprised to "suddenly" see it inside an Intel 45nm follow-on chip, too. If I was at Intel, I would do exactly that - it's not hard to make a better HT3-based CPU than the AMD ones, in particular one providing MORE HT channels!

So, my feel is: Intel wins the 1-socket and the 2-socket soccer match this year - next year, anything can happen... Flames, anyone? µ

See Also
Intel to ship LV Woodcrest server chips in Q3
Intel delivers numbers for Kentsfield, Merom and Woodcrest
Intel attempts dragging quad core Clovertown into 2006