Research Summary: Workload Identification with Physical Side Channels for AI Governance

Research demonstrates a method for external observers to identify the class of Artificial Intelligence (AI) workloads running on NVIDIA H200 GPUs by analyzing physical side channels, specifically power draw. This approach offers a potentially verifiable and independent means of monitoring AI compute usage, addressing a critical challenge in international AI governance and compliance.

Why it matters

This development is strategically important as it introduces a novel, independently verifiable mechanism for monitoring AI compute activities, which is crucial for international AI governance frameworks. It provides a technical pathway to address compliance and transparency challenges associated with powerful AI systems, enabling more effective oversight and trust in adherence to regulations.

Key insights

• AI compute verification is identified as a primary tangible and tractable point for international policy on AI governance.
• Regulating authorities require the ability to discern how compute resources are being utilized by operators, including frontier labs, to ensure compliance with agreements.
• The study shows that an external observer can identify the class of a workload running on an NVIDIA H200 GPU by monitoring its power draw.
• This physical channel method offers an advantage over on-chip telemetry (e.g., NVML), which can be manipulated, by allowing independent observation without operator cooperation.
• The research collected extensive data, comprising 930 five-second traces at approximately 10 MHz, to cover seventeen open Large Language Model (LLM) families for analysis.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2609.00309