P018 Telemetry read-path integrity and accuracy

Does any shipping GPU expose an authenticated, tamper-resistant power/clock telemetry path a root-privileged host cannot forge or replay?

Authenticity and accuracy of GPU power/clock telemetry read paths.

Terms: NVML (Nvidia Management Library): the standard software interface for reading a GPU's power, clock and temperature.

Status

Analyzed A requirements design for the read path (attested enclave or secure-boot-chained firmware, hardware-rooted signing, nonce-binding against replay, attested monitoring software), pointing to the flexHEG guarantee-processor designs; nothing built, and no shipping GPU is named as having it - every source says such paths 'do not yet exist'. Telemetry accuracy (sampling fraction, counter error) is in the statement but not the headline: related, with measured work (Yang et al. 2023: only 25% of runtime sampled for power on A100/H100).

The monitoring agent must read counter values from a source the adversary cannot tamper with—e.g., an attested enclave on the GPU firmware boundary (Aarne et al., 2025) or a verified read path through secure-boot-chained firmware (Petrie, 2025; O’Gara et al., 2025). Without this, an adversary with software privilege simply returns forged values.
Rahman & Tajdari (2026), Detecting Hidden ML Training With Zero-Overhead Telemetry (MIRI; arXiv 2606.19262) (2026-06-17), L776–783

Posed by

Also raised in: Abdelghafar & Kulp (2026), Privacy-Preserving AI Verification via Minimal Information Disclosure; Rahman & Tajdari (2026), Detecting Hidden ML Training With Zero-Overhead Telemetry (MIRI); Kimpson, Baker & Graham (2026), Can Power Draw Constrain Covert Compute? Limits of Analogue Verification for AI Governance, arXiv 2610.07476.

Sub-questions

IDSub-questionWhat the sources report
P018.1Does any shipping GPU expose an authenticated, tamper-resistant read path for power/clock telemetry (signed in-GPU, nonce-bound) such that a root-privileged host cannot forge or replay NVML readings?open
P018.2Can monitoring software, classifier weights and feature code be attested so they cannot be swapped for a permissive version?open
P018.3Can an operator inject a counterfeit current waveform into a clamped external probe (shunt / bypass conductor) without detection?open
P018.4What fraction of runtime does NVIDIA's on-board power sensor sample (≈25% on A100/H100), and has Blackwell changed it?a source suggests an answer (not checked)
P018.5What is the GA100 NVML total-energy counter's error against an external reference meter?open
P018.6Has per-instruction energy attribution been verified against external measurement on Ampere/Hopper, and has a PowerSensor3-class external meter been used for verification on a data-centre GPU?partly answered

Sub-questions not yet stated precisely enough to check

IDSub-question
P018.7How can evidence be checked for origin, integrity, coverage, freshness and execution binding when the integrity layer cannot make a compromised source honest?

Possibly relevant work found by automated search

Found by AI agents searching for this problem's sub-questions, mostly by keyword and citation in the OpenAlex scholarly database and on the web; a second AI pass reviewed each one, and we spot-checked some by hand. They seemed reasonably related, as far as one can tell from outside each paper's field. Some come from neighbouring fields, such as arms control and financial audit, and bear on the problem by analogy. Not reviewed against the status above; a paper listed here does not change that status. 928 of 1052 quotes checked word for word against the paper's text.

For P018.1: Does any shipping GPU expose an authenticated, tamper-resistant read path for power/clock telemetry (signed in-GPU, nonce-bound) such that a root-privileged host cannot forge or replay NVML readings?

For P018.2: Can monitoring software, classifier weights and feature code be attested so they cannot be swapped for a permissive version?

For P018.3: Can an operator inject a counterfeit current waveform into a clamped external probe (shunt / bypass conductor) without detection?

For P018.4: What fraction of runtime does NVIDIA's on-board power sensor sample (≈25% on A100/H100), and has Blackwell changed it?

For P018.5: What is the GA100 NVML total-energy counter's error against an external reference meter?

For P018.6: Has per-instruction energy attribution been verified against external measurement on Ampere/Hopper, and has a PowerSensor3-class external meter been used for verification on a data-centre GPU?

Data