P019 Bounding compute from power and analog measurement
How tightly can power and analog observables bound a chip's or facility's compute, and how much does the bound loosen from on-die to feed meter under adversarial choice?
Power-side compute bounds (beta*, energy per op) and the chip/node analog observables and meter limits that feed them, from on-die to feed meter.
Terms: β (beta): the largest hidden computation a power trace cannot rule out, as a fraction of the declared machine capacity. It depends on declared utilisation (how busy the declared work keeps the chip); β (beta) is its worst case over all utilisations. The paper's headline "β = 1.16" is that worst case; 0.41 is what a demonstrated attack actually hid. Energy per operation: joules consumed per arithmetic operation, used to turn power readings into a compute estimate. ΔP0 (delta-P0): the spread of the chip's non-compute overhead power (idle draw, cooling, I/O), which the verifier cannot see separately from compute; 41.2 W on the A100 measured. w_C: how much energy per operation varies with where the data sits (the 'locality' factor; 1.04× on the A100 measured). N: the number of devices.
Status
Started One NVIDIA A100 model, one driver, one cluster; energy read from on-die telemetry with a perfect meter assumed; the authors' own adversary model; the bound tightens to about 0.06 only with re-execution at an observed operating point; a 0.41-1.16 gap left open. The loosening from on-die to feed meter and the multi-device facility bound are explicitly not addressed ('assume a perfect meter'; 'do not offer a facility-level, multi-device version'), so that clause has no work.
Measurements on NVIDIA A100 GPUs constrain β = 1.16 in the worst case, while adversarial matched-energy strategies are shown to hide at least β = 0.41 of compute.
Posed by
- Scher & Thiergart (2024), Mechanisms to Verify International Agreements About AI Development (MIRI):
It is unclear how closely power draw correlates with HFU (especially in an adversarial environment).
(Building Blocks › Power draw, p. 91, L4382–4383)
Also raised in: Harack et al. (2025), Verification for International AI Governance (Oxford AIGI); 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); Canary Institute research; Wasil, Reed, Miller & Barnett (2024), Verification Methods for International AI Agreements; Kimpson, Baker & Graham (2026), Can Power Draw Constrain Covert Compute? Limits of Analogue Verification for AI Governance, arXiv 2610.07476; Proofworks, Living Verification Documents.
Sub-questions
| ID | Sub-question | What the sources report |
|---|---|---|
| P019.1 | Can analog measurement attest the chip itself (integrity, identity) — has a physics-based attestation (PAtt-style) scheme been built for a GPU? | implied by a source, not stated as open |
| P019.2 | Can a prover controlling the GPU make active-probe telemetry under-report true utilisation when host and device are both untrusted? | partly answered |
| P019.3 | Can a verifier learn a GPU's operating clock (and hence voltage) without trusting the operator - e.g. an antenna beside a populated rack recovering SM clock to a stated precision? | open |
| P019.4 | Can an external sensor identify GPU idle state (cold / context up / post-load) at a stated accuracy, or a trusted channel report static/overhead power separately? | implied by a source, not stated as open |
| P019.5 | How much do energy-per-op bands vary across DVFS range, vendors (A100 vs MI300) and operand patterns (does any pattern exceed the checkerboard's energy per op by a stated margin)? | open |
| P019.6 | At what unit cost and volume can side-channel suppression components be produced, and does look-through jamming suppress leakage by a stated amount while leaving the inspector's measurement usable? | open |
| P019.7 | Using external sensors only, what error (%) is achieved in inferring a chip's operations over an interval, and how well does power track HFU under adversarial padding? | open |
| P019.8 | For a fixed workload type, can OP/W be predicted from power, temperature and clock to a stated precision? | open |
| P019.9 | How does the error in inferred AI-chip energy/compute scale with metering granularity (per chip interposer vs server vs PDU/rack)? | open |
| P019.10 | For each auxiliary channel (power, EM, thermal, acoustic), has a study shown it can confirm a declared workload's compute volume to a stated precision? | open |
| P019.11 | What is the error (+-%) of a cumulative-FLOP estimate from facility energy at the grid connection, across accelerator generations and utilisation regimes? | open |
| P019.12 | How much do real meter noise and the feed-side conversion/cooling draw widen a single-device bound (beta*, delta-P0) relative to a perfect on-die meter? | open |
| P019.13 | Does the bound for N devices behind shared power and cooling exceed the single-device bound when the prover chooses allocation? | open |
| P019.14 | Is there a retrofit per-server or per-chip power sensor with tamper evidence and stated accuracy in production? | open |
| P019.15 | Under the A100 protocol's adversary, what is beta* (the largest hidden computation a power trace cannot rule out, as a fraction of declared capacity) on H100/B200, and does re-execution without an observed operating point stay at beta*~1? | a source suggests an answer (not checked) |
| P019.16 | What is the per-op energy gap for the same declared kernel across accelerator generations, firmware and cooling? | open |
| P019.17 | What is the largest covert load a meter-accepted schedule can hide (closing the 0.41-1.16 gap)? | open |
| P019.18 | By what factor does joint energy-time inversion lower beta* relative to energy alone? | open |
| P019.19 | Does modelling power-ramp and DVFS-governor limits from the resolved trace lower beta* by a stated margin (a target to be agreed)? | open |
| P019.20 | What is the minimum marginal energy (pJ/op) of a loss-improving training step on a model above a stated size (a target to be agreed)? | open |
| P019.21 | What is the locality width w_C for memory-bound kernels (e.g. long-context attention)? | open |
| P019.22 | Has a frontier-competitive model been pre-trained mainly in 8-bit (int8/FP8), breaking the fp16/bf16 assumption of energy floors? | implied by a source, not stated as open |
| P019.23 | What MFU / energy-estimate bias do accelerator-level power models achieve under adversarial workload choice (target <=0.03; <0.01 hard)? | partly answered |
Sub-questions not yet stated precisely enough to check
| ID | Sub-question |
|---|---|
| P019.24 | How long would mass-producing chip-activity sensors for millions of AI chips take? |
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 P019.1: Can analog measurement attest the chip itself (integrity, identity) — has a physics-based attestation (PAtt-style) scheme been built for a GPU?
For P019.2: Can a prover controlling the GPU make active-probe telemetry under-report true utilisation when host and device are both untrusted?
For P019.3: Can a verifier learn a GPU's operating clock (and hence voltage) without trusting the operator - e.g. an antenna beside a populated rack recovering SM clock to a stated precision?
For P019.5: How much do energy-per-op bands vary across DVFS range, vendors (A100 vs MI300) and operand patterns (does any pattern exceed the checkerboard's energy per op by a stated margin)?
For P019.6: At what unit cost and volume can side-channel suppression components be produced, and does look-through jamming suppress leakage by a stated amount while leaving the inspector's measurement usable?
For P019.7: Using external sensors only, what error (%) is achieved in inferring a chip's operations over an interval, and how well does power track HFU under adversarial padding?
For P019.8: For a fixed workload type, can OP/W be predicted from power, temperature and clock to a stated precision?
For P019.9: How does the error in inferred AI-chip energy/compute scale with metering granularity (per chip interposer vs server vs PDU/rack)?
For P019.11: What is the error (+-%) of a cumulative-FLOP estimate from facility energy at the grid connection, across accelerator generations and utilisation regimes?
For P019.12: How much do real meter noise and the feed-side conversion/cooling draw widen a single-device bound (beta*, delta-P0) relative to a perfect on-die meter?
For P019.14: Is there a retrofit per-server or per-chip power sensor with tamper evidence and stated accuracy in production?
For P019.17: What is the largest covert load a meter-accepted schedule can hide (closing the 0.41-1.16 gap)?
For P019.22: Has a frontier-competitive model been pre-trained mainly in 8-bit (int8/FP8), breaking the fp16/bf16 assumption of energy floors?
Data
- problems.csv — one row per problem
- problem_status.csv — status, scope and evidence per problem
- problem_sources.csv — the posed-by quote and other documents per problem
- leaves.csv — all sub-questions, including those without a checkable form
- leaf_sources.csv — every source row behind each sub-question