neural_network_dataflow

Radically Simplified GPU Parallelization: The Alea Dataflow Programming Model

Many programmers still leave the massive GPU parallel power unused – be it because of lacking experience in CUDA or because of limited time and budget. We aim to drastically simplify GPU parallelization by introducing our Alea dataflow programming model based on .NET. Complex computations can be easily and rapidly composed of a set of prefabricated and customizable operations that underlie asynchronous execution. The run-time system automatically translates this abstract model to efficient GPU code and schedules the operations with minimum memory transfers. By way of illustrative application cases of finance and statistics, we explain the model, take a look at the run-time system, and demonstrate its performance that proves to be as good as in manually optimized CUDA implementations.

Presentation slides