SACA: System-level Analog CIM Accelerators Simulation Framework: Architecture and Cycle-accurate System-to-device Simulator

Analog Computation-In-Memory (CIM) architectures promise to bring to the edge the required compute and memory demands of TinyML applications while consuming extremely low power. However, the analog CIM paradigm is suitable for accelerating vector-matrix multiplication patterns alone, and the accuracy of the computation itself is stirred by the CIM device and its driving circuit non-idealities. Despite these practical constraints, CIM accelerators are often developed and evaluated in isolation without considering real-world system-level conditions, such as sharing system resources (host CPU, main-memory, and interconnect) for inter-layer pre/post-processing, data alignment, and data movement. These make it challenging to evaluate the energy, performance, area, and accuracy tradeoff for practical, end-to-end applications. To address this, we propose a first SoC level, gem5-based, open-source simulation framework for CIM accelerator design. It supports the modeling of hierarchical CIM accelerators for different device technologies and digital/mixed-signal driving circuit configurations, along with their non-ideal behaviour. The associated full-stack software provides APIs for (re-)configuring the CIM accelerator for offloading computations at the system level. To demonstrate some capabilities of the SACA, we carried out design space exploration on two representative TinyML tasks - Human Activity Recognition and CIFAR10 image classification. The results lead to optimal accelerator profiles and indicate a tradeoff between the energy, area, performance, and accuracy for different configurations.