7.8 A 22nm Delta-Sigma Computing-In-Memory (Δ∑CIM) SRAM Macro with Near-Zero-Mean Outputs and LSB-First ADCs Achieving 21.38TOPS/W for 8b-MAC Edge AI Processing
Peiyu Chen, Meng Wu, Wentao Zhao, Jiajia Cui, Zhixuan Wang, Yadong Zhang, Qijun Wang, Jiayoon Ru, Linxiao Shen, Tianyu Jia, Yufei Ma, Le Ye, Ru Huang
Abstract
In Al-edge devices, the changes of input features are normally progressive or occasional, e.g., abnormal surveillance, hence the reprocessing of unchanged data consumes a tremendously redundant amount of energy. Computing-in-memory (CIM) directly executes matrix-vector multiplications (MVMs) in memory, eliminating costly data movement energy in deep neural networks (DNNs) [2–6]. Prior CIM work only explored the sparsity of DNNs to improve energy efficiency, but the trend of employing non-sparse activation functions, e.g., leaky ReLU, degrade the benefits of leveraging sparsity [1]. Even if sparsity can be exploited, the redundant unchanged input features in analog CIM still consume massive amount of dynamic power (Fig. 7.8.1). From a circuit point-of-view, the energy consumption of analog CIMs is dominated by full-precision ADCs. In different DNN applications, the mean of analog CIM outputs is unpredictable and fluctuating, which requires the ADC to have a high dynamic range to guarantee coverage, introducing a high-power overhead.