Acceleration of Compute-Intensive Applications on Field Programmable Gate Arrays

In recent years, the field of high-performance computing has been facing a new challenge: achieving high throughput at the lowest energy cost. Recent interest in field-programmable gate arrays (FPGA) has been spurred by their significant growth in density and speed. While they were, until recently, considered an alternative to application-specific integrated circuits (ASIC) for low volume designs, they have become an alternative compute platform that achieves much higher floating-point operations (FLOPS) per unit of energy.To partially offset the massive cost of the energy consumption in CPUs and GPUs, this dissertation explores the design and implementation of high-throughput energy-efficient compute-intensive applications on FPGAs. I show how these demanding applications can be built. To this end, I have chosen three applications from diverse domains: (a) Human Action Recognition from the field of computer vision and image processing, (b) Quantum Dynamics Simulations from the field of computational physics, and (c) the QR decomposition of Tall-and-Skinny Matrices from the field of high-performance linear algebra. Regarding (a), I show that FPGAs combined with GPUs outperform homogeneous platforms by a factor of 1.3 while consuming 50% less energy. In regards to (b), for systems having over a thousand atoms, I show that FPGAs using wide pipelines oriented towards the processing of sparse matrices surpasses competing platforms by a factor of 1.5 while consuming 4.0x less energy. In terms of (c), for tall-and-skinny matrices having over 50K rows, I show that FPGAs using wide and deep pipelines can exceed the performance of competing platforms by a factor of 1.5 while executing as much as twice more FLOPS per unit of energy