Enhancing system observability: Integrating real-time voltage telemetry and functional monitoring for deeper insights

Modern AI systems-on-chip (SoCs) often include hundreds of compute cores, putting their power delivery networks (PDNs) close to their physical limits during intense HPC or AI workloads. This instability usually leads design teams to use large safety margins, which reduces efficiency. Continuous, non-intrusive monitoring throughout a device’s life allows engineers to recover lost performance, save power and enhance long-term reliability.

Traditional monitoring still leaves a visibility gap: physical sensors can detect critical voltage events but cannot explain the workload conditions behind them, while functional monitors capture system behavior without revealing voltage stability.

Intelligent event triggers provide an opportunity to optimize data collection, thereby reducing the overhead of raw trace data measurements. This paper details a joint solution that pairs physical on-die telemetry from the Movellus Aeonic Insight™ Droop Detector with functional monitoring from Siemens Tessent™ Embedded Analytics (EA) software.

By linking these instruments via hardware cross-triggering, the system offers a single, time-aligned view of what happened and why, directly correlating specific instruction sequences or bus activity with the physical PDN events that caused them.