TinyML-Enabled Runtime Fault Detection and Self-Healing Framework for Advanced VLSI SoCs

Read full text Summarise with VIRAT
Authors: Ashley Bothello
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

CMOS scaling hasn't just made chips faster and smaller—it has, quietly, made them more fragile. Transistor densities keep climbing while supply voltages shrink, pushing modern SoC platforms ever closer to their physical operating limits. Traditional Design-for-Testability methods handle manufacturing defects well enough. But the moment a chip leaves the fab and meets real-world workloads, things get unpredictable. Aging, voltage instability, localized heating—none of that shows up in a post-fabrication test. Static monitoring strategies don't see it either. This paper proposes a TinyML-enabled runtime fault detection and self-healing framework built for advanced VLSI SoCs. Distributed on-chip sensors feed a lightweight feature processor, which drives a quantized neural classifier, which in turn triggers adaptive recovery logic—all within a single closed monitoring loop. Instead of waiting for a threshold crossing, the system learns degradation patterns directly from fault-injection-generated training data. The resulting TinyML model is compact enough for on-chip deployment, power-conscious by design. When anomalies surface, the framework responds with DVFS adjustments, clock throttling, or selective resets. Results show earlier detection and substantially fewer false alarms than static monitoring. The goal is straightforward: SoCs that don't just compute, but understand when they're wearing down.

Fields of Study & Subjects

VLSI and Analog Circuit Testing Radiation Effects in Electronics Low-power high-performance VLSI design

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
22 March 2026
Language
EN