Temperature Compensation on SRAM-Based Computation in Memory Array

Compute in memory (CIM) is a promising solution for Edge AI applications. However, previous researches on CIM focus more on performance and power efficiency (TOPS/W) improvement, while less considerations on PVT effect. In this paper, an on-chip auto-calibration technique is proposed to compensate the impact of operating temperature on the computation accuracy of the CIM array. Our simulation showed that the proposed technique offer the same BL discharge slope across different temperatures from -40 ° C to 120 °C with little INL/DNL degradation when compared to the conventional design. In normal computation mode, the proposed design achieves 310TOPS/W and after compensation the inference accuracy on MNIST dataset is improved to 96% even under different temperatures.