Abstract

Field-programmable logic devices, such as Field-Programmable Gate Arrays (FPGAs) are ‘programmable hardware’, and may be configured at startup or on-the-fly to implement a wide variety of digital electronic designs. These devices are commonly used for rapid prototyping of complex computational systems. However, I will discuss a different paradigm – the use of FPG As as general purpose computers, or as part of a general purpose computer system. In this approach, conceived by multiple researchers over the last 10-15 years, but only now coming of age, the general purpose FPGA is configured to act as a special purpose computational machine for whatever computation is required at the present time; the architecture is designed around the algorithm.

I will briefly discuss the evolution of FPGA architectures and draw comparisons with the evolution of general purpose processors, before discussing two main advantages of modern FPGA architectures, the ability to design numerical representation around the algorithm, and the potential memory bandwidth available.

Although companies such as Cray and SRC now incorporate FPG As in their new supercomputers, one of the main challenges remains how to program such devices. We will touch on this question via an example providing food for thought. Finally, I will present some numerical comparisons of performance achieved using FPG As compared to traditional general purpose processors, showing 1-2 orders of magnitude improvements in speed for particular applications.