high performance computing on graphics processing units: hgpu.org

A novel language system has given rise to promising alternatives to standard formal and processor network models of computation. An interstring linked with a abstract machine environment, shares sub-expressions, transfers data, and spatially allocates resources for the parallel evaluation of dataflow. Formal models called the a-Ram family are introduced, designed to support interstring programming languages (interlanguages). Distinct from dataflow, graph rewriting, and FPGA models, a-Ram instructions are bit level and execute in situ. They support sequential and parallel languages without the space/time overheads associated with the Turing Machine and l-calculus, enabling massive programs to be simulated. The devices of one a-Ram model, called the Synchronic A-Ram, are fully connected and simpler than FPGA LUT’s. A compiler for an interlanguage called Space, has been developed for the Synchronic A-Ram. Space is MIMD. strictly typed, and deterministic. Barring memory allocation and compilation, modules are referentially transparent. At a high level of abstraction, modules exhibit a state transition system, aiding verification. Data structures and parallel iteration are straightforward to implement, and allocations of sub-processes and data transfers to resources are implicit. Space points towards highly connected architectures called Synchronic Engines, that scale in a GALS manner. Synchronic Engines are more general purpose than systolic arrays and GPUs, and bypass programmability and conflict issues associated with multicores.