Streaming processors

Traditionally, performance increases have come from microprocessor frequency scaling. However, due to power and other constraints, scaling looks to only deliver modest performance improvements in the future. In the future large performance improvements demanded by computationally intensive applications must come from exploiting parallelism. Intel and AMD are scaling up the number of cores per chip and processors per node in order to higher degrees of Symetric Multi-Processor (SMP). Existing software has to be modified to take advantage of potentially modest speed improvements that remain limited by a machine's memory bandwidth. The change in software presents an opportunity to move beyond conventional processors to custom accelerators. These accelerators offer the potential of much higher performance by delivering parallelism that is tailored to a particular application. In particular, streaming processors offer a route around the ``memory wall'' by maximising operations performed per data item retrieved from memory. Stream processors can be implemented using Field-Programmable Gate Arrays (FPGAs) and can speed up highly parallel applications by over an order of magnitude. FPGA acceleration has been successfully demonstrated in a variety of application domains including computational finance (Zhang et al. , 2005), fluid dynamics (Sano et al. , 2007), cryptography (Cheung et al. , 2005) and seismic processing (He et al. , 2004).