GEMM Accelerator on Cyclone V SoC
Custom RTL matrix-multiply accelerator taken from SystemVerilog through Qsys integration to a full Linux bring-up (custom kernel, device-tree overlay, UIO binding, and user-space driver). An 8x8 heterogeneous systolic array runs INT8/INT16/INT32/FP16/FP32 in one multi-precision MAC. Fitter shows 112/112 DSP at 48% ALMs; a wide-port DMA fast path, operand-locality tuning (resident A, double-buffered B), and a hardware tile-row batch sequencer walk memory stall down 93% to 76% to 19% and take INT8 512^3 to 3.15 GOP/s at 50 MHz, about 27x an ARM A9 software reference, at 60% compute occupancy, bit-exact across all precisions (60/60 on-board tests).
SystemVerilog · FPGA · Cyclone V SoC · Linux driver · DSP
read the case study →