high performance computing on graphics processing units: hgpu.org

This dissertation project addresses the implementation of Bayesian inference on FPGAs and GPUs, following a top-down approach and using OpenCL. The target application of this Bayesian inference algorithms is artificial perception in robotics. The aim is to improve the power efficiency of Bayesian inference computations. Previous work at our university in the scope of an European project followed a bottom-up approach and developed a toolchain that enabled having custom circuits for Bayesian inference on reconfigurable logic. These had better power efficiency than desktop solutions, but require more design effort. In this work the goal is to use already available vendor tools, namely the OpenCL support from Intel (formerly Altera), to explore the design space in search of low power efficient solutions. To achieve this, the same benchmark problem used in previous works is going to be applied, tested in various dimensions in order to study scaling challenges. The main metrics analysed are nominal power, energy consumed, latency and result’s precision. As expected the results show a great gain in power efficiency in relation to desktop solutions, and comparable performance with previous works developed in the context of the project BAMBI, but with gain in point precision, integration and usability.