Intel Kentsfield, Clovertown quite happy on Nvidia

MY OLD YET LARGE bones require a new rule as the age creeps up - no economy-class flights beyond six hours, the very border of physical endurance I set. So, I had to miss the US IDF this time, and please myself only with the local Taipei IDF offshoot a fortnight later.

But, Taiwan is just like Singapore - small yet spicy, so there were interesting observations both from senior Intel hats (more liberal when so far from home HQ?) and local top vendors, who, after all, make a board-level reality out of these Intel and AMD CPUs.

The keynotes are mostly the shortened versions of the previous US ones, with a few new things slotted in over the past two weeks - like updates to its WiMax chippery, for instance. Vector processing potential for the Nehalem and Gesher generations, something brought in from the brutally-murdered Alpha EV9 project, seems more real now, according to the man in charge, John Crawford - more on it in a separate story.

Also, real-time car model ray tracing on Tulsa eight-core system was impressive, as it is still very much a FPU-based, very embarassingly parallel job, and no OpenGL or DirectX can give you that level of realism - but well, a two-socket Clovertown system would probably do it faster, cheaper (same number of cores, but each faster and with faster memory access).

Talking about Clovertown and its desktop twin, Kentsfield - both Intel and the board vendors I spoke to are genuinely happy about the chip. After all, it is not just every day that a senior Intel exec tells you that it is no big deal to reliably overclock a 2.66GHz Kentsfield to 4GHz! And, I specifically asked for accent on 'reliable' (with liquid cooling of course). Since Kentsfields are still in short supply, I'll make sure not to burn mine upfront though.

An interesting thought is that, if really overclocked and running reliably every day at 4GHz, Kentsfield would give you a peak of 64 GFLOPs (full double precision). And, if for some reason the same is valid for Clovertown, we're talking about a 128GFLOPs deskside workstation!

Now, while the Clovertown production version is expected to be the standard 2.66GHz CPU with 1333FSB, Intel seems to be more open now towards the idea of 'special HPC performance editions' of its server CPUs. Essentially a server/workstation equivalent of its Extreme Edition on the desktop, these could also be multiplier-unlocked, and, combined with high-performance chipsets such as the upcoming 2x16 SLI Nvidia dual FSB offering, reach further performance through water cooling, for instance - that dual 4GHz Clovertown (or its 5GHz variety once the 45nm process kicks in) becomes very real then.

Talking about Nvidia, the Kentsfield demo was this time based on SLI - no CrossFire to be seen, anymore. Further to that, I had no luck getting a ATI RD600-based Core2 mainboard from any of the vendors... it is just a no-no, it seems unfortunately. Everyone is waiting for the new Nforce high-end rev for Core 2, the one with three PCI-E X16 slots and upwards of 2GHz FSB capability, sometime in December. Soo, it seems... good bye, ATI on Intel - we'll miss you! µ