Project Description

Multi FPGA compute acceleration PCIe card

KR developed a full-length, full-height PCIe card with sixteen XCZU4CG MPSoC chips, a XCKU95 FPGA, and Zynq XC7Z020 SoC for use in heterogeneous compute acceleration applications. The sixteen MPSoCs served as distributed compute nodes with algorithm acceleration in logic fabric and provided logic assisted access to 32TB of NVMe storage. The KU095 (a VU13P upgrade is also possible for additional compute capacity) served as the PCIe bridge and data router as well as limited capacity compute node. The Zynq7000 allowed for Ethernet enabled remote debugging with a JTAG multiplexer implemented in logic fabric, which allowed the dynamic reconfiguration of the JTAG chain of all 17 target FPGA / SoCs on the card.

The system was successfully deployed by our client in several PoC. The proof of concept deployments ranged from accelerated risk analysis of an investment portfolio to the analysis of satellite data for detecting and tracking atmospheric rivers off the coast of California. Each PoC demonstrated a speed advantage of at least 100:1 compared to the legacy implementation. The design resulted in multiple issued international patents.

Knowledge Resources contributed the complete hardware design, including structural and cooling metal work, implemented logic elements, and assisted the customer’s software team in implementing and testing the acceleration software environment and drivers. All units were built utilizing our on-premise manufacturing capabilities. KR also handled the certification process and the module passed FCC and UL certification in preparation for the initial market introduction.