Spatial Computation: Computing without General-Purpose Processors

Mihai Budiu

Wednesday, March 31, 2004
10:00am - SENSQ 5317
Refreshments at 9:30am in SENSQ 5319

Abstract

In this talk I explore the strengths and weaknesses of Spatial Computation (SC), a method of translating programs into hardware by assigning to each program instruction a dedicated hardware implementation. SC synthesizes circuits with ample parallelism, using mostly local communication. SC addresses many of the limitations of contemporary high-performance processor architectures: limited ability to exploit parallelism, reliance on slow global communication signals, design complexity, inflexibility of the instruction-set architecture, low energy efficiency, increasing power consumption.

I will present compilation technologies developed for two spatial architectures: PipeRench (virtualized hardware) and ASH (Application-Specific Hardware). Program transformations developed for these architectures have broad applicability in the realms of traditional compilation, high-level synthesis and asynchronous circuits. This talk will focus on CASH: a compiler for ASH, which translates ANSI C programs into asynchronous hardware circuits. Circuits generated by CASH from dusty-deck C multimedia kernels compare favorably in performance with high-end superscalar microprocessors, while being up to three orders of magnitude more efficient in terms of energy and energy-delay.