Device Variation-Aware Adaptive Quantization for MRAM-based Accurate In-Memory Computing Without On-chip Training

Hardware-accelerated artificial intelligence with emerging nonvolatile memory such as spin-transfer torque-magneto-resistive random-access memory (STT-MRAM) is pushing both the algorithm and hardware to their design limits. The restrictions for analog-based in-memory computing (IMC) include the device variation, IR drop effect due to low resistance of STT-MRAM and read disturbance in the memory array at the advanced technology node. On-chip hybrid training can recover the inference accuracy but at the cost of many training epochs, reducing the available lifetime for updating cycles needed for on-chip inference. In this work, we show the unique feature of device variations in the foundry STT-MRAM array and propose a software-hardware cross-layer co-design scheme for STT-MRAM IMC. By sensing device level variations, we can leverage them for more conductance levels to adaptively quantize the deep neural networks (DNNs). This device variation-aware adaptive quantization (DVAQ) scheme enables a DNN inference accuracy comparable to on-chip hybrid training without on-chip training. Besides, this DVAQ scheme greatly reduces IR drop effects. Overall, the DVAQ allows one to achieve less than a 1% accuracy drop compared with in-situ training under 40 % device variation/noise without on-chip training in several DNN applications.