Supercomputing

CNN has an article posted about supercomputers. It touches briefly on the originals (CRAY1, 1976) running at 100MFLOPS (10^6), and the current frontrunners - the Japanese RIKEN-K running at 10.5PFLOP (10^15).  Of course, it's worth noting that the RIKEN-K runs on 705,000 cores and 1.4Petabytes of RAM.

For the sake of comparison, a quick google shows most modern quad-core desktop computers running in the 4-7GFLOP (10^9) range, or about fifty times faster than the CRAY1 was running forty years ago.

Next step? ExaFLOP, or 10^18, about 100 times more powerful than the RIKEN-K. Of course, this is starting to get into serious money, space, and power. According to the CNN article, we're talking rooms the size of football fields, with monstrous cooling arrangements, and power consumption around 100 MW. (A modern nuclear reactor will produce power in the 600-1000MW range.)

In other words: there have to be some leaps before this is going to happen. But we're getting there.

And, that brings us to what will likely be the next arrangement: distributed computing. It's already used extensively on a volunteer basis for things like SETI@Home and Folding@Home, where people can devote extra processing cycles to the search for ET life, protein folding for cancer research, or even the search for the next Mersenne Prime.

Applied to the above goals, instead of a client app that uses spare cycles, why not contract out X number of towers that will sit in colocated data centers, distributing the power/cooling load as well as the vulnerability? With the ready availability of dark fiber in the US (and nationally), the data bandwidth to connect widely-spaced cores is there... it makes sense, at least from my perspective.

And what to do with these PFLOP and EFLOP arrays? Particle modelling, crypto, weather research, gene and cancer research... you name it.

Top500.org lists the most powerful systems in the world. Worth a click and skim through.