QUALCOMM Incorporated
Ternary computation memory systems and circuits employing binary bit cell-XNOR circuits particularly suited to deep neural network (DNN) computing

Abstract:

A multiply-accumulate (MAC) operation in a deep neural network (DNN) consists of multiplying each input signal to a node by a respective numerical weight data and summing the products. Using ternary values for the input signals and weight data reduces memory and processing resources significantly. By representing ternary values in two-bit binary form, MAC operations can be replaced with logic operations (e.g., XNOR, popcount) implemented in logic circuits integrated into individual memory array elements in which the numerical weight data are stored. In this regard, a ternary computation circuit (TCC) includes a memory circuit integrated with a logic circuit. A memory array including TCCs performs a plurality of parallel operations (e.g., column or row elements) and determines a popcount. A TCC array in which logic circuits in columns or rows employ a single read-enable signal can reduce routing complexity and congestion of a metal layer in a semiconductor device.