Seminar: Control-Flow versus Data-Flow Supercomputers

Veljko Milutinovic, University of Belgrade, presented Seminar: Control-Flow versus Data-Flow SupercomputersVeljko - Siena, Italy, November 12th, 2012
ABSTRACT: This presentation starts with a comparison of various supercomputer types as far as the following issues: (a) Speed, (b) Power, (c) Size, (d) Programming effort, (e) Debugging effort, and (f) Compilation time. It continues with details of the Maxeler approach to data-flow supercomputing, using a number of examples. It concludes with a projection of future trends. If finishes with an elaboration of a PHD research methodology inspired by the scientific success of Maxeler (a spin-off of Stanford and Imperial College London). Data-flow supercomputers compile application code down to the gate level, which helps obtain a number of advantages over Control-flow supercomputers of the same purchasing price. Speedups, for various applications in physics/chemistry/biology, are about 20 times or more, and up to about 200 times for specific business applications, as published by JP-Morgan (a 20% owner of Maxeler). Monthly electricity bills are down for the factor of about 20, which is an important issue, since the two-year electricity bills may overpass the initial investment in the case of Control-flow supercomputers. The size reductions go down also for the factor of about 20. Speedup related data are shown for selected applications in physics, geo-physics, banking, and econometry. A group of PhD-student researchers in Belgrade now develops code for a number of applications not covered so far. They all follow the same methodological path, the details of which will be elaborated in this talk.