Multi-FPGA Evaluation Platform for Disaggregated Computing

Abstract

We present a versatile FPGA-based evaluation platformfor exploring alternative execution strategies on disaggregatedenvironments for applications, considering differentprocessing block types: compute cores, memory, and accelerators. Developers can interconnect different blocks types in orderto create optimal configurations. A user-level software libraryallows quick mapping of applications on real hardware. Wehave implemented a fully working prototype using three ZC706FPGA boards, and evaluated different software / hardwareconfigurations of a matrix multiplication benchmark.