HP's launch of its Arches chipset, powering the latest generation of Integrity and Superdome servers, is a strong reminder that processors aren't everything. Arches supports 3TB more memory than its predecessor, has a new crossbar switch with twice the memory bandwidth and 25 percent lower latency, and many new reliability features. In all, says HP, it'll do 30 percent more work.
All this was due to be an expensive birthday present for Montecito, Intel's long-delayed dual-core Itanium. As it is, it'll be a retirement present for current Itanium stalwart Madison — and HP's last-of-the-line PA-RISC chips. When Montecito finally arrives later this year, Arches will further build on the twofold jump in performance Intel promises.
That's good, because we'll be waiting a while for the next leap forward. After the 90nm Montecito, it'll shrink to the 65nm Montvale — not this year, as initially hoped, but 2007. Then it's a two-year pause until Tukwila in 2009, and things really get steaming with a major architectural revamp in Poulson at "a later time". This will have Common System Interconnect, promising board-level compatibility with the rest of Intel's server range.
We'll also have to wait and see with IBM's Power6 architecture in 2007 and Sun's Rock in 2008, but the real competition is closer to home. Itanium is now a generation behind its Core Architecture brothers, which will be in 45nm next year and — Moore willing — 32nm by 2009. Core has shown that low-power ideas intended for portable computing are essential for efficient servers: performance above all else isn't the way to get the best.
Intel is proud of the $10bn investment promised by the Itanium industry through to 2010 — $5bn of which is coming from HP. Yet put in the context of the $50bn/year server market, let alone the $200bn PC sector, it's not enough to close the gap. What used to be low-end chips are now being driven by the same design imperatives that characterise Itanium, only with more ferocity.
In a world getting used to the idea that key server technologies can migrate up from the laptop, be under no illusions where the real revolution will come from. HP isn't, and it isn't throwing away that $5 billion. By 2010, the Common System Interconnect means that the grandson of Arches will work not just with Itanium but with Core. That $10bn looks a lot more like a bet on CSI than a pure Itanium investment. After all, processors aren't everything.