James W. Demmel
University of California, Berkeley
Mathematical and Physical Sciences
Demmel is a world-renowned expert in developing efficient high performance algorithms and creating widely used software for solving linear algebra problems used in science and engineering. Some examples: In parallel processing systems, communicating results among processors is a key performance bottleneck. Demmel pioneered a communications avoiding approach for matrix manipulation algorithms that minimizes communication. Different execution orders cause different floating-point roundoff-errors. Demmel's scalable algorithms achieve bit-wise reproducible floating-point summation for any summation order. Analysis to improve performance by using the least floating-point precision necessary to get a sufficiently accurate answer is difficult. Demmel's method does this analysis automatically. Additionally Demmel is a leading developer of ScaLAPACK, the standard parallel linear algebra software system.