Date: Sep 02, 2013 Source: (
click here to go to the source)
Reservoir Labs researchers will be presenting results for reducing the amount of energy consumed in computing Fast Fourier Transforms (FFTs) at the High Performance Extreme Computing conference in Waltham, MA on Sep 11. Reservoir's results are based on utilizing an algorithm from MIT's professor Alan Edelman and his collaborators Peter McCorquodale and Sivan Toledo that was published in 1997. Their algorithm utilized the Fast Multipole Method (FMM) to reduce the number of all-to-all communications by 3X. Reservoir's research updates that work with modern kernel-independent FMM and utilizes GPU acceleration for near-field computations to provide speedups with reduction in communications for modern architectures.
With current VLSI scaling trends, the energy required for communication of data is increasingly dominating the energy required for logic switching. This is expected to continue to increase, particular as industry moves into the extreme and exascale computing regimes. Reservoir's work provides a route to computing the critical kernel of FFTs in that new regime.
This work is being shared with collaborators in the DARPA PERFECT (Power Efficiency Revolution for Embedded Computing Technologies) program. Reservoir's SPOTTER (Software Power Optimization to Efficiency Revolution) project is developing advanced algorithms and compiler techniques for PERFECT hardware.
