A Hybrid Analog–Digital Dynamic Compute-in-Memory Architecture with Workload-Aware Dynamic Mode Selection for Edge AI

Abstract

By cutting down on the amount of data that needs to be sent back and forth between memory and processor units, Compute-in-Memory (CIM) architectures make deep neural network workloads more efficient and save energy. On the other hand, current mixed-signal CIM systems have problems with phase-change-memory-based crossbars, analog-to-digital-converter-less hybrid designs. This research introduces a Hybrid Analog-Digital Dynamic Compute-in-Memory (HAD-DCIM) architecture that integrates analogue accumulation with effective digital refinement using a workload-aware dynamic mode-selection mechanism. The proposed architecture is evaluated using a custom Python-based behaviour model. The model calculates latency, leakage power and overall power consumption. Experiments are performed on convolutional, fully connected and single head attention layers at an operating frequency of 600MHz. Unlike traditional designs like PCM CIM, HCiM, and AIMC Simulated testing shows that HAD-DCIM can cut latency by up to 20% and gives 8.7ns, lower leakage energy of 35.5pJ which is reduced by 35%, dynamic power consumption of 0.020mW which also reduced by 25%. It also provides energy efficiency of 6.7 TOPS/W which is improved by 20% than the existing designs. These results suggest that HAD-DCIM is a very effective, accurate, and flexible solution for edge-AI applications that don't have a lot of power.