Systolic Array Based Convolutional Neural Network Inference on FPGA

Convolutional Neural Networks (CNNs) possess a particular edge over its predecessor, the Multi-Layer Perceptron (MLP). This is due to its weight sharing features that allows the CNN to use less parameters for the same number of outputs as compared to the MLP. Systolic arrays capitalize on the weight sharing property of CNNs to do data reuse while performing convolutional operations, in order to reduce the power consumption from the memory accesses. A kernel fitting systolic processing element array was designed with only positive multiplication to increase the throughput and power efficiency of the CNN accelerator, while using weight stationary dataflow to achieve data reuse in the systolic array. A cost-optimized lightweight solution is implemented through low-cost FPGA hardware so as to allow for greater accessibility. The CNN accelerator consumes 0.363 W power at 100 MHz operating frequency. A peak throughput of 10.98 GOps/s was achieved with peak performance density of 0.200 GOps/s/DSP and peak power efficiency of 30.26 GOps/s/W. Even with the added support for additional functions, proposed design achieved up to 1.59x better power efficiency compared to other systolic implementations and up to 6.17x better power efficiency compared to non-systolic implementations.