innovate > feature

Supercomputers are so named because that’s exactly what they are: superfast and super-powerful -- and until recently, supersized and super-high-priced. But Nvidia, one of the top makers of computer display hardware, is starting to change that image. Just a few months ago, Nvidia unveiled its latest game-changing product: a supercomputer for the desktop.

Built with multiple processors (CPUs), supercomputers offer a serious advantage for simulations, large scale calculations (especially floating-point calculations), and image generation for the professional and scientific fields. While one CPU handles one task at a time, a cluster of CPUs can handle many more tasks at one time, enabling scientists and researchers to use supercomputers to crunch enormous numbers of calculations far faster.

Traditionally built by the big companies -- Cray, IBM, HP, for example -- supercomputers cost hundreds of thousands or even millions of dollars each, but they also tend to be the ones that lead the Top500, the official ranking of the fastest computers in the world. Over the last decade or so, though, enterprising researchers desperate for more supercomputing time have begun to build supercomputers from scratch, so to speak. With cheap, older computers, and free open-source software, the closet hardware geeks connect these together to form a “cluster,” using anywhere from a few computers to hundreds, all interconnected to form one brain. While these aren’t as fast as the commercially built supercomputers, they are at least less costly. But the size problem remains.

Then, in February 2007, Nvidia released its first version of CUDA (Compute Unified Device Architecture), a development language and environment that addresses the GPU (graphics processing unit) on Nvidia’s graphics cards. Soon, company execs noticed that some college and university research departments were using CUDA to repurpose the GPUs on the company’s GeForce graphics cards. By combining several GPUs on a single computer motherboard, these early adopters were able to create what amounted to blazing-fast mini supercomputers in a single desktop box.

Clustering GPUs instead of CPUs makes use of the fact that the basic architecture of the GPU is very different. A CPU, even a very fast one, deals with one computation at a time. A GPU, however, is crafted to handle a graphics screen full of individual pixels, which involves hundreds of simultaneous computations at once, making it well suited for supercomputing. The comparison often used to describe the speed increase in the use of GPUs instead of CPUs is the difference between reading Leo Tolstoy’s War and Peace one page at a time and having the ability to read every page at once. When trying to divine the incredibly complex secrets of cosmology or sub-atomic particle theory, this can be the difference between some progress or none.

“We had university researchers come to us and tell us this has unchained them from their data centers,” says Andrew Humber, Nvidia spokesperson. “Even on these hacked-together systems, they were able to get teraFLOPS speed for their computations, something they’d never seen before.” A FLOPS (FLoating point Operations Per Second) is a measurement of computational speed, and a TeraFLOPS is a trillion FLOPS.

The University of Illinois Urbana-Champaign (UIUC) was one of the first research institutions to leverage the parallel architecture of the GPU to accelerate research. John Stone, Senior Research Programmer at the Beckman Institute for Advanced Science and Technology, UIUC, says that when he first began using GPU technology, he was able to drop the time needed to run one set of computations from four and a half days down to 27 minutes. This also meant he could perform many of his lesser programming jobs right at his desk in far less time, instead of placing them in the queue at the data center.

“We were shocked at how fast we were able to get these programs done using GPUs,” Stone says. “We’ve seen increases in speed more than 10 times, and more recently, even up to 80 times faster than the same work can be done on much bigger machines. It’s just incredible.”

Recognizing it was on the verge of something huge, Nvidia came out with its Tesla line of products last year, repackaging its high-end GPU used in the GeForce line of display cards into the Tesla card, the first dedicated general purpose GPU. Tesla card technology is at the heart of the newly released Tesla Home Supercomputer, which by combining several Tesla cards in a single computer, delivers cluster-level supercomputing performance, enabling it to run computations 250 times faster than standard PCs and workstations. And all this for $6,000 to $10,000, depending on configuration.

It’s not exactly in the price range of the usual home gamer, but it’s not intended for this use, anyway. Programmed to perform as computational tools, the GPUs in the Tesla desktops handle FLOPS rather than determining how the pixels in your avatar for “World of Warcraft” move and look. And for researchers and scientists far too used to seeing price tags in the six-figure range for supercomputers (not to mention the garage-sized space needed to house them), this is a very attractive product.

“Being able to use GPUs for computational work has dramatically changed the landscape of research,” says Stone. “This is perhaps the most exciting thing to come along in a very long time.”

next article archive