Computing has become much more than just an augmenter of science; it advances all science and engineering areas. The demand for more powerful computers is ever-growing. As Dennard scaling ceased to be valid and Moore's law is slowing down, innovations in high-performance computing (HPC) architectures are necessary. The main driving avenue of innovation for new architectures of HPC systems will be energy efficiency.
Besides improving energy efficiency, there is a challenge to keep programmability high and avoid overspecialization as the problem domains in HPC keep changing frequently.
We propose a system consisting of a general-purpose, application-class core that is tuned for programmability and runs a full-fledged, state-of-the-art operating system such as Linux and a specialized cluster architecture delivering the energy-efficient compute power.
Compared to a wide range of other systems, we achieve a double-precision efficiency gain of up to 15x.weiterlesen