X
Innovation

Rupert Goodwins' Diary

Wednesday 20/7/2005It’s getting increasingly difficult to divine the fate of Itanium. There’s a steady trickle of news about it — Intel launching a new version with a faster bus, SGI producing a new server based on the chip, and HP will give you a $3250 Itanium server if you attend a $2000 developer workshop — but also a fair dollop of anti-news.
Written by Rupert Goodwins, Contributor

Wednesday 20/7/2005

It’s getting increasingly difficult to divine the fate of Itanium. There’s a steady trickle of news about it — Intel launching a new version with a faster bus, SGI producing a new server based on the chip, and HP will give you a $3250 Itanium server if you attend a $2000 developer workshop — but also a fair dollop of anti-news. That selfsame HP (Intel’s closest partner in Itania) is keen for us to think of Itanium as the chip of choice for mammoth servers with thirty two processors or more, yet has just sold a whopping great 1024 processor supercomputer to the United States Department of Defence — built out of AMD Opterons. Curious stuff.

The new version with a faster bus is also mildly confusing. Although Intel will happily sell you one, it won’t sell you the supporting chipset needed to make it do anything other than paperweight duty, nor a board to plug it into. The only people who do make such a chipset — and thus the only customers for the chip — are those crazy guys at Hitachi. No idea how many servers they ship but this time last year their quarterly sales — when combined with those from four other Asian companies — were pegged at seventy. Let nobody say that Intel doesn’t care for its smaller customers.

The faster bus is, of course, nicked from Montecito — the forthcoming and widely-trailed dual-core Itanium that is due to appear in a blaze of glory (and a whirr of fans) later this year. It’s being bigged up something rotten, with Intel claiming that it will blow the socks off any other chip for floating point performance, with a single four-way, two core device clocking in at 45 gigaflops. Twenty of these should be able to reach a teraflop. For comparison, the first Cray-1 supercomputer in 1976 ran at 160 megaflops — that’s about three hundred times slower than a single Montecito — and required around a thousand times more power from its own 150 kilowatt generator. And no, it didn’t run Windows.

I know we’re supposed to worry about the commercial side of things: just look at Cray if you want to know where the pursuit of power for power’s sake gets you. And I don’t honestly know what the global market is for teraflop-sized commercial systems, nor I suspect does anyone else. It’s even impossible to imagine what on earth would happen if the Silicon Fairy turned up, waved its magic wand and gave us all teraflop computers sitting on our desks — what sort of software would appear, and would it really look that different to what we have now?

But would I like one? Hell, yeah. Even if it was just a paperweight.

Editorial standards