Energy-Efficient Systolic Array for Complex-Valued Convolutional Neural Networks

Complex-valued convolutional neural networks (CVCNNs) are increasingly used in applications involving complex-valued model parameters, offering superior performance over real-valued counterparts. However, the split-ReLU activation function, widely used in CV-CNNs for its implementation simplicity and empirical effectiveness, exacerbates the sparsity in input activations. Our analysis reveals that roughly 70% of activation values contain at least one zero component, resulting in significant redundant computations when mapped to conventional systolic arrays, ultimately leading to energy inefficiency. To address this, we introduce complex systolic array, a hardware accelerator that reduces dynamic energy in CV-CNNs. Leveraging an input stationary dataflow, the architecture employs zero-aware sparsity control and a novel complex processing element to eliminate unnecessary register reads and multiplications. Experimental results demonstrate energy savings of 38.5% in register reads and 33.2% in compute units compared to the conventional systolic array.