Fast linking of separately-compiled FPGA blocks without a NoC

Dedicated point-to-point wires (DW) can be used in place of a Packet-Switched Networks-on-a-Chip (PSNoC) for fast linking of separately-compiled FPGA blocks, providing higher bandwidth and performance with less area overhead without increasing compile time. Previous work showed that separate compilation of FPGA modules using a pre-compiled FPGA overlay could reduce the long FPGA compile time by defining and separately mapping small partially reconfigurable blocks (Processing Elements) and using a fixed PSNoC to connect them together. Nonetheless, the lightweight PSNoC cannot meet the high data transmission requirements for some critical links, limiting overall performance. We demonstrate that DWs, where the producer ports and consumer ports are directly connected instead of sharing limited-throughput, packet-switched connections, can provide us with high throughput between Processing Elements (PEs) while preserving the fast compile time; the DWs also reside on partially reconfigurable blocks and can be compiled along with the reconfigurable PEs simultaneously on the cloud. Adjacent pages can be connected by fast links with low latency. Mapping Rosetta Benchmarks, we show that the application-customized direct networks can offer 1.5-10× performance gain and 47-86% interface area overhead savings compared to previous work with PSNoCs.

Download paper here