Why GPGPUS for evolutionary computation?

Abstract

In 2006, for the first time since they were invented, processors stopped running faster and faster, due to heat dissipation limits. In order to provide more powerful chips, manufacturers then started developingmulti-core processors, a path that had already been taken by graphics cards manufacturers earlier on. In 2012, NVIDIAcame outwith GK110 processors boasting 2,880 single precision cores and 960 double precision cores, for a computing power of 6 TFlops in single precision and 1.7 TFlops in double precision. Supercomputers are currently made of millions of general purpose graphics processing unit cores which poses another problem: what kind of algorithms can exploit such a massive parallelism? This chapter explains why and how artificial evolution can exploit future massively parallel exaflop machines in a very efficient way to bring solutions to generic complex inverse problems.