A MILLION million

Posted: 02.11.08

ORNL's JAGUAR (Courtesy of ORNL)

On Broadway in New York City, a flop is something you try to avoid, always. In politics a flop preceded by a flip can be the end of a career. In skid-row, a flop house is a place where winos sleep. To sit down without grace is to flop, and flat thong-like shoes are called flip-flops.

But in computers a FLOPS is something to do, it's a Floating Point Operation Per Second—it is an acronym, not a word, and is all capitalized, it is not Flops, or flops, it's FLOPS—that's the law.

How many FLOPS your processor can do can cause a modifier to be added in front of the acronym. If you can do a million FLOPS then you have done a mega FLOPS, or a MFLOPS. If you could do a thousand MFLOPS then you would have achieved gaga FLOPS or GFLOPS. Are you with me so far?

Today the challenge is to achieve a million million FLOPS—or tera FLOPS, or TFLOPS. Doing a TFLOPS for Big Blue was accomplished in 2000. In 2006, Big Blue hit 280 TFLOPS. Then, in 2007 it's Blue Gene/P crossed the magic peta FLOPS line; and in that same year the Cray Jaguar XT$ at Oak Ridge National Laboratories' (ORNL) hit 119 TFLOPS—it was powered by AMD Opteron processors, almost twelve thousand of them. The system was funded at $50 million. Scaling all that and you come up with $4,166 per CPU, or almost 10 GFLOPS per CPU, or 2.4 MFLOPS/dollar.

AMD recently released their R670 GPU which has 320 double precision (DP) floating-point processors (FPs) in it. They then put two of them (and a GB of RAM) on a board (Radeon 3870X2) and will offer it for sale for $500. That $500 AIB does over one TFLOPS.

So if you wanted the equivalent TFLOPS of the ORNL Jaguar you could do it for under $60,000 and actually theoretically hit 125 TFLOPS.

This is the argument that the GPU computing advocates are making about the hidden potential of GPUs. Well maybe it isn't so hidden anymore.

Now ORNLs wanted to catch up with Argonne National laboratory (where Blue Gene is) they could either fork up nine times $50 million for nine more Cray XTS or buy 900 Radeon 3870X2s for a mere half a million dollars

Of course ORNL won't do that and they'll have trillion reasons why it wouldn't be the thing to do, bureaucratic- speak, bureaucraticspeak, bureaucratic-speak. OK, so the empire builders in the super-computer world won't switch to more efficient, less expensive computing technology, it's too late for them. I think we (and they) all know that. And who cares, you can count them on your fingers .

But how about all the super computer wannabe users? How about the folks who are paying too much for server clusters and aren't hide bound by bureaucratic- speak and bureaucraticthink, and have a bottom line to worry about? And, how about all the non-ivy league colleges and universities out there? They could use a megaflops super computer that cost less than a million dollars.

And that's exactly what AMD thinks their Fire Stream will satisfy, and Nvidia thinks their Tesla will satisfy, and Intel thinks their Larrabee will satisfy, and Sony/IBM think their Cell will satisfy, and Mercury computer thinks they will build.

These companies are going to bring huge FLOPS to the masses and none of them are going to flop doing it. s JON PEDDIE'S