EFSim: An Efficient and Flexible Simulation Framework for Digital In-Memory Computing

Digital in-memory computing (IMC) has drawn significant attention in neural network acceleration due to its superior energy efficiency and resistance to non-ideal variations. Simultaneously, the rapid evolution of neural networks and the large design space of IMC systems necessitate an early-stage verification simulator. In this paper, we introduce EFSim, an efficient and flexible simulation framework for digital in-memory computing systems. On the hardware level, a parameterized Chisel-based hardware library is constructed to enable the rapid and flexible generation of customized Verilog IMC circuits. On the software level, the IMC inference engine facilitates the efficient deployment of neural network models. Furthermore, EFSim performs hybrid compilation of designed IMC systems using Verilator, thereby enabling rapid performance evaluation and software-hardware co-optimization. Validation through two designed IMC macros demonstrates that the relative modeling error rate of EFSim is below 10%. Experimental results indicate that EFSim enables efficient evaluation of digital IMC systems and supports extensive design space exploration. The EFSim is open-sourced at https://github.com/cq-work/EFSim.