AI-Driven Cache Coherence Verification with Graph Neural Networks in SoC-Based Shared Memory Systems

Main Article Content

Santosh Appachu Devanira Poovaiah
Sohil Sri Mani Yeshwanth Grandhi

Abstract

For system-on-chip (SoC) architectures with shared memory to retain data consistency and system stability, cache coherence is essential. Contemporary SoC designs are becoming more complicated and concurrent, making it difficult for traditional verification techniques to keep up. In this study, we suggest utilizing Graph Neural Networks (GNNs) to verify cache coherence in an AI-driven manner. We capture temporal and spatial interdependence between cores, caches, and interconnects by modeling the memory subsystem [5] and coherence interactions of the SoC as a dynamic graph. By learning from simulation traces or runtime logs, the GNN is trained to identify coherence breaches, including stale reads, write propagation delays, and protocol violations. By using this strategy, human debugging effort is greatly decreased, and corner-case scenarios that are hard to find with traditional rule-based techniques are better covered. According to experimental results on common SoC coherence benchmarks, our GNN-based framework is a potential approach for scalable, intelligent verification in next-generation multicore systems since it can detect subtle coherence flaws with high accuracy and few false positives.

Article Details

Section
Articles