Computing Without Processors
The hardware software divide must be eliminated to allow us to tackle ever demanding requirements for reductions in latency and energy consumption. There is a limit to the problems we can solve using blades of multicore processors in the cloud. The most challenging problems will require us to brandish not only multicore processors but also GPUs (graphics cards hijacked for parallel computing) and FPGAs (soft hardware that can be moduled into a customized architecture for implementing a specific algorithm along with its computational engines and optimized data-movement). We have the technology today to build such heterogenous computing devices. What we lack are the programming models to allow agile development for a collection of components that must work together but are not all regular processors. This presentation describes what such heterogeneous computing architectures might look like and proposes a specific technique for mapping data-parallel computations onto these devices.
Satnam Singh works in the technical infrastructure division of Google that is responsible for producing the systems that manage the world's largest distributed computer. He has previously worked at Microsoft, Xilinx, British Telecom, VLSI Technology/Compass, European Silicon Structures as well as an academic at the University of Glashow and the University of Birmingham along with other roles at the University of Washington, Imperial College London, and Chalmers Technical University.