Massively Parallel SIMD -- the CSX Add-on Architecture
Benedict Gaster
ClearSpeed Technology, Bristol
Massively parallel multi-core systems, particularly SIMD arrays, are no
longer the province of the large super computer, but instead they are making
their way into a single PC and even embedded processors with
applications ranging from video encode/decode to physical modeling.
The CSX architecture provides a novel approach to a new generation of low
power/high performance SIMD array processors. Used as either a co-processor
to existing systems or integrated as part of a system-on-chip for embedded
applications the CSX allows critical sections of an application to be off
loaded and computed in parallel.
In the first part of the talk we outline the current and future CSX
architectures focusing on the instruction set and system architecture.
Moving on to describe an alternative approach to programming these machines
with a simple extension to C we are able to model naturally the programming
of arbitrary sized multi-core arrays while retaining the semantic
completeness of the parent language.