IDF server chat focuses on Itanium

SOME OF THE most interesting bits at IDF tend come out of smaller meetings. This time, I attended a round table with Kirk Skaugen, VP of the Digital Enterprise Group, and the other non-bunny suited reporters and I peppered him with questions.

If some of them seem like they are going all over the map, that is because they actually are going all over the map, from reporters themselves coming from all over the map. To his credit, Kirk kept the answers focused on the server related things he is charged with shepherding.

The first one drifted into a discussion of what is the market for Itanium, and there are three - mainframe replacement, high performance RISC replacement, and the highest end of HPC. Some are low volume high spec part markets, the others lower spec with higher volume.

From there it was on to the so called common platform(1), a story in itself, but the questions here were on RAS, and how it tied in. If you have a petaflop computer with thousands or tens of thousands of CPUs, you have about three CPU-years worth of compute time a day per thousand CPUs. 10K procs, 30 CPU-years a day. If you have only one problem per 10 years for a computer, the 1K CPU machine will stay up for about three days before it does something ugly, the 10K CPU machine about eight hours.

To make these machines work for problems that take more than an overnight binge to run, you need either hugely fault tolerant algorithms, insane RAS features, or both. Intel is not counting on the software guys to do the heavy lifting, in this case the software guys are probably more counting on Intel to make a machine that does not crash.

That is why you see Itanium focusing so heavily on RAS and uptime, for much of their target market, it really matters. When the line was launched, x86 was nowhere in the game, some say on purpose, others say it was a cost issue. In any case, that gap is quickly narrowing, but according to Kirk, it is more a case of x86 going up, Itanium going up faster.

On top of this, the common platform does not mean a 1U 2S server needs to implement the same RAS features as a supercomputer, but it will have all the CPU reliability of the big boys no matter what you plug it in to. If you switch to a Xeon, you lose the CPU RAS and may lose some platform RAS The choice is your call, but I think the instruction set will be more of a factor in the purchase decisions for most.

Then I asked about Foxton, Montecito and Montvale. There was no definitive answer about why Foxton was not in Montecito, the most I could get was that it might have delayed the product. Must ... bite ... back ... sarcastic ... comment. When asked directly if it was in Montvale, yes, no or maybe, I got a maybe. There you have it, a definite maybe, count on it.

Most of the other questions were things already covered on The Inq, nothing of earth shattering import, but nothing out of place either. Kirk Skaugen for one sees a bright future for Intel in the server market .µ

(1) Yeah I know, don't write.