A 2.6 TOPS/W 16-Bit Fixed-Point Convolutional Neural Network Learning Processor in 65-nm CMOS

We present a convolutional neural network (CNN) learning processor, which accelerates the stochastic gradient descent (SGD) with a momentum-based training algorithm in 16-bit fixed-point precision. Using a new cyclic weight storage and access scheme, we use the same off-the-shelf SRAMs for nontranspose and transpose operations during feedforward (FF) and feedbackward (FB) phases, respectively, of the CNN learning process. The 65-nm CNN learning processor achieves peak energy efficiency of 2.6 TOPS/W for 16-bit fixed-point operations, consuming 10.45 mW at 0.55 V.