Running an NVMe host controller in FPGA fabric on the Sidewinder board, the IP, the PCIe path, and the throughput it reaches without a CPU in the loop. Running the host controller in fabric rather than on a CPU core removes the PCIe round-trip through a general-purpose processor entirely, which is most of where the latency in a software NVMe stack comes from. The Sidewinder board carries the Zynq UltraScale+ device the IP targets, and the demo shows sustained throughput with no CPU cycles spent servicing the storage path at all.

Learn More at https://fidus.com/sidewinder/