I just returned from the 14th SIAM conference on parallel computing in Seattle. I was presenting a SIMD matrix compression algorithm Ive been working on for some applications in quantum chemistry. There were a lot of interesting technical talks but my favorite was a non-technical plenary talk by Horst Simon on energy-efficient computing, which you might think would be a tough sell to the supercomputing community. However Simon obviously knew his audience and made a very convincing argument to the effect that greater energy efficiency is an absolute must in order to even build an exaflop machine.
The reason I liked the talk so much though was that he just used that fact as a hook to pull you into a more general discussion of energy in America and the future of computing. I will relate one interesting point he made which is that you cant just build a competitive supercomputer anywhere these days. There are two big reasons for this: one is that you have to have stable access to a huge amount of (preferably renewable) energy; the other related reason is you want to maximize your Power Usage Effectiveness which essentially means building in a cool climate so as to not waste resources on cooling. One result of this is that you are going to see more and more remote ‘data factories’ in the Pacific Northwest such as Google’s Oregon Data Center. These are essentially supercomputers dedicated to fabricating data for consumers in the form of cloud apps such as Gmail or Amazon’s EC2.
