A Heterogeneous Microprocessor for Intermittent AI Inference Using Nonvolatile-SRAM-Based Compute-In-Memory

Compute-In-Memory (CIM), characterized by efficient matrix-vector multiplication, has been recognized as a promising candidate technology for edge AI computing. However, applying CIM in extreme edge scenarios, where power delivery is limited and unstable, still faces challenges. The relatively high memory write energy, compared with CIM computing, prevents its further gains on ultra-lower-power devices. The frequent backup/restore in intermittent computing, together with the nonvolatile memory (NVM) with even higher program energy, escalates the gap between compute and memory operations. To overcome these challenges, this work proposes a heterogeneous microprocessor using nonvolatile-SRAM-based CIM (nvSRAM-CIM). The nvSRAM-CIM macro features flexible neural-architecture-aware backup strategies, together with efficient sparsity-aware backup and multiplication. The 8T1R/10T1R nvSRAM/nvSRAM-CIM bit cell reduces 60% and 93% backup and restore energy for weights and activations. On top of the circuit optimization, the nvSRAM-CIM-based AI accelerator shows up to 1.9× energy efficiency with moderate power-failure frequency for neural-network inference.