Skip to Main Content
Field-programmable gate arrays (FPGAs) commonly implement massively parallel circuits that require significant memory bandwidth. Due to I/O and memory limitations, parallel tasks often share bandwidth via arbitration, whose efficiency is critical to ensure parallelism is not wasted. In this letter, we introduce a bandwidth-sensitivity-aware heuristic for arbitration that analyzes the effect of memory bandwidth on performance for each application task, and then accordingly allocates bandwidth to minimize execution time. When compared to round robin (RR) arbitration, application speedups as high as 6.5×are achieved.