The changes afoot in computer architecture are shifting our focus toward parallel computing in both SIMD and MIMD style. Despite the skeptics who point to Amdahl's Law and cite the difficulty in making nearly everything run in parallel, the HPC community has had little trouble with Amdahl's Law thanks to ever-growing problem scale. The skeptics are right, though; it's just that the solution they naturally point to, higher single thread performance, is perhaps less effective than another strategy, namely finer parallelism granularity. Accomplishing this will require lower overhead communication among sub-computations, a challenge that has so far been neglected by most of the computer architecture community. This talk is an attempt to kindle interest in the subject.
Wednesday, October 12, 2011
Free and open to the public