Accurate Program/Verify Schemes of Resistive Switching Memory (RRAM) for In-Memory Neural Network Circuits

Resistive switching memory (RRAM) is a promising technology for embedded and its application in computing. In particular, RRAM arrays can provide convenient primitive matrix-vector multiplication (MVM) with strong impact on the acceleration of neural networks artificial intelligence (AI). At same time, affected by intrinsic conductance variations, which might cause degradation accuracy AI inference hardware. This work provides detailed study multilevel-cell (MLC) programming network applications. We compare three MLC schemes discuss their variations terms different slopes characteristics. test two-layer fully connected (FC-NN) as function scheme, number weight levels, mapping configuration. find tradeoff between FC-NN accuracy, size, current consumption. highlights importance holistic approach to accelerators encompassing device properties, overall circuit performance, specifications.