Timing and Memory Telemetry on GPUs for AI Governance

Saleh K. Monfared et al. have posted an arXiv preprint (arXiv:2602.09369, submitted 10 Feb 2026) introducing a compute-based GPU telemetry framework intended to support post-deployment AI governance.

• Main announcement: The paper introduces a measurement framework for GPU telemetry composed of four primitives (PoW-inspired probabilistic workload, VDF-derived latency-sensitive workloads, GEMM-based tensor-core throughput tests, and a VRAM-residency hashing test) and reports evaluation results on contention, architectural alignment, memory pressure, and power overhead; the preprint is available as PDF/HTML/TeX on arXiv (v1, 10 Feb 2026, 781 KB).
• Context and details: The work targets untrusted host/device environments, claims observability without trusted firmware/enclaves/vendor counters, and frames telemetry as actionable signals for post-deployment governance; metadata includes arXiv identifier arXiv:2602.09369, DOI via DataCite pending registration, and a CC BY 4.0 license.