Article ID: 23.20260460
FPGA accelerators for convolutional neural networks need high array utilization despite tight on-chip storage, routing, and clock-frequency budgets. Operand delivery limits systolic-array CNN accelerators when storage and window generation do not follow the array wavefront. This letter proposes a feed path that encodes alignment, padding validity, and window overlap in local control. Skew-aware bank addresses replace wide-data realignment; validity masks and copy markers suppress padding-only steps and repeated window transfers. Four physical 16 × 16 cores form a logical 32 × 32 array. Results show that the data path meets the 250-MHz timing constraint after placement and routing on XC7Z100. Cycle-accurate RTL simulation at this frequency yields convolution-backbone throughputs exceeding 413, 363, 471, and 204 GOP/s for AlexNet, ResNet-18, VGG-16, and MobileNetV2, respectively.