P070 Compute metering and trusted time

Can an on-chip compute and activity meter and a trusted timer count accurately and resist clock, fault and counter-manipulation attacks?

On-chip compute meters, counters and trusted time.

Status

Analyzed Design only: a meter built on existing performance counters with licence-set caps, plus proposed protections (monotonic counters, lockstep redundancy, shielding against fault injection, sanity checks between counters). Nothing fabricated or measured; Ansari 2026 records 'the absence of functional prototypes', and current counters are 'accessible via firmware interfaces and were designed for debugging rather than tamper-resistant governance'. Trusted-time clause: design-level only; CNAS proposes counting clock cycles instead of time because tamper-proof time 'generally is considered very difficult', and RAND says there is 'no airtight defense' against clock attacks. Fault-injection resistance and sanity-check false-alarm rates: no work.

Since these values are already collected in hardware, additional logic could be added to set a maximum value for each counter, determined by a license.
Kulp et al. (2024), Hardware-Enabled Governance Mechanisms (RAND WR-A3056-1) (2024-01), L2179–2180

Posed by

Also raised in: Baker, Kulp, Marks, Brundage & Heim (2025), Verifying International Agreements on AI: Six Layers of Verification for Rules on Large-Scale AI Development and Deployment (RAND WR-A4077-1); Petrie & Aarne (2025), Technical Options for Flexible Hardware-Enabled Guarantees (flexHEG Part 2); Aarne, Fist & Withers (2024), Secure, Governable Chips (CNAS); Kulp et al. (2024), Hardware-Enabled Governance Mechanisms (RAND WR-A3056-1); AI Security Forum, Requests for Discussion.

Sub-questions

IDSub-questionWhat the sources report
P070.1Can an accelerator securely log active GPU-hours such that overclocking or clock manipulation is detected?open
P070.2Has an on-chip FLOP meter been fabricated and shown to count within a stated precision on a production workload?open
P070.3For each candidate meter (arithmetic ops incl. integer, memory bytes, interconnect bytes, joules), can a compiler-level change cut the metered count by a stated margin for the same training result?partly answered
P070.4Can performance-counter FLOP readings be altered or falsified at the source by a firmware-level attacker without detection?implied by a source, not stated as open
P070.5Can laser/EM fault injection reset or freeze a performance or cycle counter without tripping a lockstep redundancy check?partly answered
P070.6What fraction of injected counter faults do counter-consistency sanity checks detect, and at what false-positive rate per GPU-day on production workloads?implied by a source, not stated as open
P070.7Does read-after-write verification of OTP/NVM state defeat high-voltage supply disruption, so meter state persists securely across power-down?implied by a source, not stated as open
P070.8Can a FIB edit disable one distributed security block's gating without tripping neighbouring blocks, at what cost per chip?implied by a source, not stated as open
P070.9Has a tamper-resistant on-chip clock / secure timer resisted clock, power-supply and oscillator injection-locking attacks (minimum detectable slowdown), giving a trusted time source for licence date bounds?open
P070.10Can ageing-sensor readings estimate a chip's cumulative cycles to within a stated precision (an 'odometer')?implied by a source, not stated as open

Sub-questions not yet stated precisely enough to check

IDSub-question
P070.11Should existing counters be hardened or redesigned, and can a formal proof show no hidden reset paths?
P070.12Is counting clock cycles more robust than time for licences?

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 P070.1: Can an accelerator securely log active GPU-hours such that overclocking or clock manipulation is detected?

For P070.2: Has an on-chip FLOP meter been fabricated and shown to count within a stated precision on a production workload?

For P070.5: Can laser/EM fault injection reset or freeze a performance or cycle counter without tripping a lockstep redundancy check?

For P070.7: Does read-after-write verification of OTP/NVM state defeat high-voltage supply disruption, so meter state persists securely across power-down?

For P070.8: Can a FIB edit disable one distributed security block's gating without tripping neighbouring blocks, at what cost per chip?

For P070.9: Has a tamper-resistant on-chip clock / secure timer resisted clock, power-supply and oscillator injection-locking attacks (minimum detectable slowdown), giving a trusted time source for licence date bounds?

For P070.10: Can ageing-sensor readings estimate a chip's cumulative cycles to within a stated precision (an 'odometer')?

Data