Where’s the Beef? Why FPGAs Are So Fast

Where do all the cycles go when microprocessor applications are implemented spatially as circuits on an FPGA? It is well established that certain sequential applications can be captured spatially and achieve breathtaking speedups when run on an FPGA, but why? Despite running at clock speeds orders of magnitude slower compared to their embedded processor equivalents, FPGA applications can "lose " enough cycles to create exceptionally fast spatially-oriented circuits. We profile and analyze three canonical applications amenable to FPGA speedup to quantify exactly where FPGAs gain that speedup. We compare the FPGA implementations to several idealized software platforms. The idealized software platforms give insight as to how FPGA implementations attain such dramatic speedups. We quantify the effects of parallelizing and pipelining instructions, streaming data, and eliminating the instruction fetch, showing exactly where the cycles are lost in an FPGA implementation. We also show how the memory interface to the FPGA will affect the performance. Our results show that custom memory interfaces are the most effective way at enabling much greater performance on the FPGA, and that memory interfaces traditional software use become a bottleneck when the FPGA uses the same interface. The results, though not surprising, provide a clearer and more intuitive understanding of the performance FPGAs can achieve, offering researchers and engineers alike a new angle to attack the task of parallelizing applications