Ultra-fast and Energy-efficient Write-Computing Operation for Neuromorphic Computing

Abstract

Emerging Non-volatile memory (NVM) has demonstrated superior performance on the computing-in-memory (CIM) architecture. By re-purposing the peripheral circuits, the certain NVM array can perform both storage and computing operations to accelerate the data-intensive convention Neural Networks (CNNs). However, the parallelism of the NVM-based CIM should be considered. In this paper, we present a CIM architecture developed by the Spin-orbit Torque (SOT) MRAM using both read-out and write-in operations. We highlight the memory structure and control method of the write-in operations. With the excellent write performance of SOT-MRAM, the proposed writein operation can obtain ultra-fast and energy-efficient computing data operations. The write-in operation works as a complement of the conventional read-out CIM architecture rather than replace it.