On Designing Efficient and Reliable Nonvolatile Memory-Based Computing-In-Memory Accelerators
Nonvolatile memory (NVM)-based computing-in-memory (CIM) features nonvolatile storage, in-place computing and reduction in data traffic. However, the development of NVM-based CIM is hampered by immature fabrication processes and inevitable operational faults, making cross-layer optimization extremely important. This work focuses on spiking NVM-based CIM that offers both the superior performance of artificial neural networks and the astonishing power efficiency of spiking neural networks. We introduce techniques to enhance inferencing accuracy by increasing programming accuracy, improve area efficiency by minimizing on-chip buffers, boost power efficiency by lowering peak column sensing currents, and bolster system-level reliability by prediction-based error detection and recovery schemes.
