Characterizing latency overheads in the deployment of FPGA accelerators

FPGA hardware accelerators have recently enjoyed significant attention as platforms for further accelerating computation in the datacenter but they potentially add additional layers of hardware and software interfacing that can further increase communication latency. In this paper, we characterize these overheads for streaming applications where latency can be an important consideration. We examine the latency and throughput characteristics of traditional server-based PCIe connected accelerators, and the more recent approach of network attached FPGA accelerators. We additionally quantify the additional overhead introduced by virtualising accelerators on FPGAs