A novel mixing-mode architecture that combines the sum-of-resistance and sum-of-current approaches is proposed to address the challenge of handling large-scale sum-of-product operations with number of weights and input values exceeding thousands. The proposed architecture outputs a delay time that is proportional to the sum-of-product result of the perceptron operation, with minimal variation and good linearity. The mixing-mode architecture has high potential in computing-in-memory applications, which offers significant improvements on computing parallelism and energy efficiency for large-scale neural network implementations.