Konferenzen und Rahmenprogramm
Direct Data Exchange Between FPGAs and GPUs Using GPUDirect
Today€™s heterogeneous computer systems combine CPUs, GPUs, and FPGAs with different archi- tectures. GPUs with immense parallelization are best fitted for real-time video and signal processing. On the other hand, FPGAs are ideal for capturing and pre-processing multiple video streams. Therefore, partitioning computational tasks between these computational units and designing the data-paths between them are very important. Close and efficient interaction of the individual components is, therefore, indispensable. With PCIe, modern computers provide a high-speed internal data network between CPU, GPU, and FPGA. However, since communication usually takes place via the CPU, this communication becomes the bottleneck in applications. This paper compares two different implementations for an efficient FPGA-GPU co-design without CPU bottleneck. The discussed implementation reaches a transfer jitter of only 4 Î¼s and utilizes 72% of available PCIe bandwidth without any CPU interaction.
--- Datum: 27.02.2020 Uhrzeit: 14:30 - 15:00 Uhr Ort: Conference Counter NCC Ost