LEGOSim: A Unified Parallel Simulation Framework for Multi-chiplet Heterogeneous Integration

The rise of multi-chiplet integration challenges existing simulators like gem5 [ 45 ] and GPGPU-Sim [37 ] for efficiently simulating heterogeneous multiple-chiplet systems due to incapability to modularly integrate heterogenous chiplets, high synchronization overheads in parallel simulation, and high inter-chiplet communication modeling overhead. To address these limitations, this paper introduces LEGOSim, a unified parallel simulation framework capable of flexibly integrating various open-source and in-house designed chiplet simulators as processes in parallel simulation, referred to as "simlets" with minimal modifications needed. It introduces a three-stage simulation process that decouples chiplet simulation from inter-chiplet communication modeling to mitigate the communication modeling overhead. The framework also integrates Network-on-Interposer (NoI) simulator for modeling inter-chiplet communication, enabling accurate assessment of various interconnection architectures’ performance. Furthermore, it employs an on-demand synchronization protocol, ensuring synchronization only occurs when necessary, thus reducing overhead while maintaining correctness. Evaluated with diverse benchmarks, LEGOSim shows high accuracy in simulating multi-chiplet architectures like SIMBA [55 ] and a CiM-based accelerator [13 ], with average errors of 3.79% and 3.94%, respectively. It significantly reduces synchronization overhead by up to 99.9% compared to per-cycle synchronization and by 66.1% compared to time quantum synchronization, without synchronization errors. Five case studies show that LEGOSim also provides precise system performance metrics and stall cause reporting, simplifying tasks such as performance analysis and optimization, and can be used for design space exploration of various multi-chiplet systems.