Our paper “Reuse Kernels or Activations? A Flexible Dataflow for Low-latency Spectral CNN Acceleration” has been accepted as a full paper at the 28th ACM/SIGDA International Symposium on Field-Programmable Gate Arrays (FPGA ’20). It describes a principle designing an approach for low-latency, low-bandwidth, spectral sparse CNN accelerator on FPGAs to overcome the performance gap between the potential offered acceleration and actual FPGA implementation performance due to high memory burden and off-chip bandwidth requirement for kernel accesses.
