《Ray Bittner, Erik Ruf》.pdfVIP

Direct GPU/FPGA Communication Via PCI Express Ray Bittner, Erik Ruf Microsoft Research Redmond, USA {raybit,erikruf}@ Abstract—Parallel processing has hit mainstream computing in traverse through the PCI Express (PCIe) switch twice and the form of CPUs, GPUs and FPGAs. While explorations suffer the latency penalties of both the operating system and proceed with all three platforms individually and with the the CPU memory hardware using the red indirect path. We CPU-GPU pair, little exploration has been performed with the refer to this as a GPU-CPU-FPGA transfer. This additional synergy of GPU-FPGA. This is due in part to the cumbersome indirection adds communication latency and operating nature of communication between the two. This paper system overhead to the computation, as well as consuming presents a mechanism for direct GPU-FPGA communication bandwidth that could otherwise be used by other cluster and characterizes its performance in a full hardware elements sharing the same communication network. implementation. In this paper, we describe and evaluate a mechanism for implementing the green line in Fig. 1 for direct, bidirectional Keywords- GPU; FPGA; PCI Express; PCIe; Xilinx; CUDA; nVidia; Windows; Verilog GP