Tuesday, November 28, 2006

Cray and IBM in the HPCS Program

For those of you that actually care about the industry that I work in High Performance Computing (ie HPC or supercomputing), somthing big happened this last week. DAPRA awarded their HPCS Phase III contracts. There had been a lot of competition for these contracts. When Phase II had its down select, SGI lost out. This time around, Sun lost out. The results were what you would expect if you were an industry insider: IBM and Cray won. Here's an okay overview.

So what do they have to do? By 2010 both teams need to build and demonstrate a running petaflop computer: note DAPRA is interested in sustained, real world performance, so all you Beowulf ranters can please just go home now, thanx. Cray has $250 million with work with. IBM has $225 million to work with. Please note the numbers there. Please go read about each of the teams approaches: IBM's Power7 and Cray's Cascade. IBM's taking the conservative approach. Cray isn't. Or not as conservative.

In either case, these are not COTS machines! But wait, isn't IBM using a Power processor? Yup, but there's a lot more to this than a processor. For taht matter, there may be some desktop systems that will use the Power7, but I suspect that it won't be many. There are some power workstations out there now, but they're pretty darn pricey for what you get. To me, based on the price tag, based on the current rumor mill in the HPC world, based on the price tag, this doesn't bode well for the future of computing. At least on the small scale end.

During the 1990s we saw COTS clusters pop up everywhere: the so-called attack of the killer micros. Universities could afford them. They provided a very cheap and easily maintained way of bringing very fast assets to researchers on a limited budget. They were scaled up to really big machines: the monsters down stairs are little more than suped up clusters. They're here to stay. They're about to pass into history for the high end machines though.

Consumer parts are simply not keeping up with the demand for HPC platforms. That's because the demand for faster PC processors and bigger memory is but a shadow of what it was during the 1990s. AMD and Intel cannot keep cranking out the improvements in processing power that they did in that time frame (percentage wise). To me, based on what I am seeing, it looks like we've hit a bad case of diminishing returns for the end user and the economics are simply unfavorable for it to grow much more.

Combined with the gobsmacking costs of the highest end HPC platforms that they are now developing - $250 million for a petascale machine? - or recently developed - $800 million +/- for the Earth Simulator? - means that the popular interpretation of Moore's Law is at an end. Or so close that we're almost on the verge. As for comments about quantum computing taking ove rand charging off? Well, they've been talking about 'purely' photonic computers since I was a teenager and working on the bits and pieces even then. The parallels seem very apparent to me.

Let's face it: the Rapture of the Geeks just isn't coming.

No comments: