Empowering Parallel Computing with Field Programmable Gate Arrays

After more than 30 years, reconfigurable computing has grown from a concept to a mature field of science and technology. The cornerstone of this evolution is the field programmable gate array, a building block enabling the configuration of a custom hardware architecture. The departure from static von Neumannlike architectures opens the way to eliminate the instruction overhead and to optimize the execution speed and power consumption. FPGAs now live in a growing ecosystem of development tools, enabling software programmers to map algorithms directly onto hardware. Applications abound in many directions, including data centers, IoT, AI, image processing and space exploration. The increasing success of FPGAs is largely due to an improved toolchain with solid high-level synthesis support as well as a better integration with processor and memory systems. On the other hand, long compile times and complex design exploration remain areas for improvement. In this paper we address the evolution of FPGAs towards advanced multi-functional accelerators, discuss different programming models and their HLS language implementations, as well as high-performance tuning of FPGAs integrated into a heterogeneous platform. We pinpoint fallacies and pitfalls, and identify opportunities for language enhancements and architectural refinements.