One Joule per GFlop for BLAS2 Now!

Energy consumption is today one of the major topics that the HPC community tries to tackle. In this paper, the authors present a thought experiment aiming at building a node of a supercomputer based on a GPU (Nvidia GTX280). The paper concentrate on BLAS2 operations that dominate many scientific applications. They show that it is already possible today to build a node that is able to perform one, GFlop while consuming only one Joule. This corresponds to an energy reduction by one to two orders of magnitude. Technical limitations imposed by computer manufacturers hinder the realisation of such ecological resources. In a second part of the paper a new FLOPS-driven governor is presented aiming at boosting the CPU performance for applications that are not main memory access bound, and showing results on an Intel mobile processor.