The Shrine

I'm not kidding. One of the guys in my group does research on the Cell processor - it really is an interesting architecture. It's a hot topic in high performance computing right now.

At the moment, there are two ways of getting your hands on a Cell processor: a PS3, or a Cell Blade server. The only company that makes Cell Blade servers doesn't put their cost on their webpage. That's a hint that, minimum, it will cost you $10,000, maybe as much as $30,000.

A PS3 costs $600. That's an expensive toy, but a cheap research machine. So, with a university credit card, we went to Game Stop and picked up one of the 60 GB models. I was surprised they had any; we went there almost on a whim. Since my roommate isn't home yet, I assume he's still trying to get Linux on it. If all goes well with this one, we'll probably buy more and make a cluster.

That company is IBM btw.

Actually, no. It's <a href="http://www.mc.com/cell/products/index.cfm">Mercury Computer Systems</a>. The Cell was a joint project with IBM, Sony and Toshiba, but they're not building systems with it yet. IBM probably will eventually.

Other Kupek research projects include the following:

1) Research at gentlemen's clubs
2) drinks at bars
3) trips to Rio



The Cell is a pretty sweet processor, but perhaps too complicated ( a recurring complaint). Also it may have been too expensive - even at $600/pop it seems Sony is taking a bath with these and it's very hard to manufacture with signifacant yeilds which is probably why you see virtually all of them going to the console.

Anyway, keep us up to date on your "research".

The reason we (well, my group, I don't work with it, although I might try to port stuff to it) are interested in it is that this is what processors in supercomputers might look like in the future. Someone in my group ported a biology application to Cell, and it performed equally as it did with some 2-way CMP, 2-way SMT Power5s we have. This might not sound impressive, but those Power5s cost about $2,000 a pop. So while the Cell might be expensive for a toy, it's cheap for scientific computing.

The published version of that work: <a href="http://www.cs.wm.edu/~cda/ppopp07.pdf">Dynamic Multigrain Parallelization on the Cell Broadband Engine</a>. And no, I'm not one of those people; it's not my research. I'm basically pimping a colleague's work.

(has a brief mental image of Kup in a dropped Cadillac, driving through the hood, collecting hundred dollar bills from research papers in short miniskirts and bitch boots)

I would throw in a Beowulf joke in there, but I won't.