Woodcrest lost in workstation woods

SOME OF OUR fellow hardware geeks, who benchmarked Woodcrest systems in workstation apps, were surprised at the performance advantage it exhibited over the Opteron in a similar configuration. However, there was one weak point: lack of a high-end scalable graphics chipset.

Intel's own Greencreek platform does allow for a single PCI-E X16 slot - sufficient for a typical midrange workstation. But that is far from enough for 3D powerhouse workstation systems, where SLI or CrossFire may be mandatory to get the 3D performance required.

The loads here are usually different than those of gaming PCs. First, the applications - whether engineering, CAD or multimedia creation - are far better multithreaded than even the high end games, justifying quad-core or even eight-core machines easily. Sometimes, on an eight core, you get six to seven times speed-up over a single core in app like Cinema 4D for rendering. So, a multi-core system with truckload of memory is actually helpful here.

Secondly, the engineering models usually may not have much of texture or environment reality trickery, but will have many millions of polygons - navigating or viewing that in real time needs all the vertex processing power a card - or two of them - can provide. In this case, SLI or Crossfire setup is actually necessary, not a fashion statement - too bad that 3DLabs Realizm 800 is not around anymore, it could do almost any job by itself.

So, with Clovertown, you could have up to 8 cores in a 2-socket deskside workstation - but, only one graphics card! Doesn't exactly match, especially when compared to the Opterons, where HyperTransport scalable interconnect provides for a single North Bridge option across all system sizes. For FSB-based systems, each scaling level (single FSB, dual FSB, future quad FSB) needs a very different north bridge: no sharing here.

Then, the FB-DIMM 'problem' to put it politely pops up. Horrendous latency, yet extra heat and power consumption - both performance and power envelope are affected. Greencreek has four channels of FBD DDR2-667, providing a total theoretical 21 GB/s bandwidth. That is sufficient to drive two CPU sockets, each with its own 1333 FSB, however the latency by conversion through AMB chip adds an equivalent of extra CL1.5 to CL2 to every access compared to a non-buffered DIMM of the same chip speed and latency, on top of five to six watts that each AMB chip consumes - per DIMM!

During Computex, just before the ATI-AMD thing happened, I had a chat with ATI's chipset team that architected the RD600 chipset for the Core 2 platform. There was a kind of agreement that it makes good sense to do a dedicated proper dual-FSB Woodcrest/Clovertown chipset with full 2 x 16 PCI-E crossfire and, why now, quad-channel DDR2 standard memory. If combined with some overclocking capabilities, it could even be a good candidate for the ultimate overclocker's system.

Of course, it won't happen right now - the only chance, unless Intel addresses this concern themselves soon, is Nvidia now. Is there a 2 x 16 PCI-E full SLI chipset with standard memory for Intel's Dual Xeon platform on the cards? AMD Opteron platform has both SLI and CrossFire covered well up to the highest end. You could even make a quad-socket ultra-workstation - again, HyperTransport enables you to share a single North Bridge from Athlon64 to Opteron 800 series.

Intel really needs to work this out soon, if it wants to keep a strong presence in the high end workstation market. µ