Last week, the Folding@home team reported that they achieved five petaflops of processing power for their popular protein folding research project, which relies on processor cycles contributed by hundreds of thousands of people. That's more processing power than can be found at any single US DOE lab or supercomputing center. What's more, the five teraflops corresponds to real application performance for the project's protein simulation software, so we're not just talking peak hardware performance.

Perhaps even more impressive is that the project crossed the one petaflop mark only 18 months ago, on Sept. 16, 2007. At this rate, they'll hit an exaflop in about five years. But it's doubtful whether the Folders will really be so fortunate. Most of the performance increase over the last year and a half was the result of the GPGPU revolution. In September 2007, the project had a mere 42 teraflops of GPUs working for them. Today that number stands at 3,295 teraflops (3 petaflops). Two thirds of those are NVIDIA GPUs; one third are ATI (AMD) GPUs.

The remainder of the performance increase over the last year and a half was gleaned from Cell BE-based PlayStation3 consoles and CPU-based PCs and workstations. While more GPU-based systems will surely be added to the Folding@home infrastructure in the future, the increased performance will likely follow more of a Moore's Law type curve (albeit an accelerated one that corresponds to the faster evolution of GPUs).

In sheer computing power the five petaflop Folding@home infrastructure represents more than four times the Linpack performance of the 1.1 petaflop IBM Roadrunner supercomputer at Los Alamos. Of course, that's an apples-to-oranges comparison since supercomputers are monolithic machines built for tightly-coupled applications. Folding@home works more like a typical distributed computing system, where an application is divvied up over a large number of machines and then the results are aggregated. Some purists wouldn't call Folding@home a supercomputer at all since it doesn't exist as a stand-alone system. Nevertheless, the Folders are doing real HPC work and are pushing the envelope in both high-end computing and protein modeling.

Whether Folding@home can stay ahead of the supercomputing performance curve remains to be seen. Depending upon the kindness of strangers may turn out to be a precarious model for ultra-scale computing. The advent of cloud computing means over-provisioned PCs may end up morphing into thin clients with much less processing power to share. On the other hand, if our client systems become the visual computing platform of choice -- as Intel, AMD and NVIDIA seem to be angling for -- our PCs and even televisions will be chock-full of GPUs, multicore CPUs or some hybrid of the two. In that case, the Folders will continue to have a large reservoir of machines to tap into.

Getting to an exaflop on a distributed computing platform shouldn't be all that difficult. When you consider that a high-end gaming GPU today offers more than a teraflop of peak performance, you would need only a million or so client machines to get an aggregated peak exaflop. Only some fraction of that will translate into application performance, but considering that GPUs continue to get more powerful and the software is getting better at extracting performance, exaflop protein folding is certainly within reach.