Design productivity is a major concern preventing the mainstream adoption of FPGAs. Overlay architectures have emerged as one possible solution to this challenge, offering fast compilation and software-like programmability. However, overlays typically suffer from area and performance overheads due to limited consideration for the underlying FPGA architecture. These overlays have often been of limited size, supporting only relatively small compute kernels. This paper examines the possibility of developing larger, more efficient, overlays using multiple DSP blocks and then maximising utilisation by mapping multiple instances of kernels simultaneously onto the overlay to exploit kernel level parallelism. We show a significant improvement in achievable overlay size and overlay utilisation, with a reduction of almost 70% in the overlay tile requirement compared to existing overlay architectures, an operating frequency in excess of 300 MHz, and kernel throughputs of almost 60 GOPS.
展开▼
