Friday lecture: Brian Vinter, Professor, The Niels Bohr Institute, eScience, University of Copenhagen

Abstract:

The realization that processor performance is severely limited by the three walls; power-wall, frequency-wall, and memory-wall has led to the advent of a generation of processors commonly known as accelerators and more advanced conventional processors. A very common misunderstanding is that programming these next-generation processors is an exercise in parallel programming, but while parallelization is part of the solution correct memory layout and access patterns are still the primary key to performance.

In this talk we will take a look at different processors, their similarities and differences, how they may be programmed and how we may help automate the programming of them. To help understand the practical problems we will look at some examples of HPC tasks from industrial production, financial modeling and classic applications of supercomputing in science.