P038 Facility power models
How large is the error in facility- and distribution-level power models when the operator chooses workloads to mislead them?
Validity of facility- and PDU-level power models under adversarial choice.
Terms: PDU (power distribution unit): the rack-level equipment that feeds and meters power to servers; a facility/PDU power model estimates compute from power readings at those points.
Status
Analyzed Honest operator; facility power modelled, not metered (the NLR dataset combines measured workload profiles with facility power from a bottom-up model of the power-delivery components); 2-16 Kestrel nodes of four H100s each running Llama-2 70B LoRA and Stable Diffusion: a node-count report from mean facility power tolerates about plus or minus 15% uniform calibration error. Adversarial workload choice, the headline's condition, is not modelled at facility or distribution level. Cross-references, not evidence, because the headline names facility scale: Kimpson, Baker & Graham's adversarial power bound is for a single A100 and the paper says it offers no facility-level, multi-device version; Rahman & Tajdari measured wall versus NVML power on one workstation under evasion strategies.
When every power value is multiplied by the same calibration factor, the fixed quantizer remains perfect from −15.5% to +18.5% error; failures begin at −16% and +19%. This sweep measures tolerance to uniform calibration error.
Posed by
- Kimpson, Baker & Graham (2026), Can Power Draw Constrain Covert Compute? Limits of Analogue Verification for AI Governance, arXiv 2610.07476:
How to make the multi-device bound precise is an open question.
(L1895–1896) (source states the problem but not the quantity the question asks for)
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).
Sub-questions
| ID | Sub-question | What the sources report |
|---|---|---|
| P038.1 | What is PDU-level power-model error when the operator chooses workloads to mislead it? | open |
| P038.2 | Has a facility-level power model been validated against metered power at a stated sub-second resolution (current results rest on modelled facility power or 5-min data)? | open |
| P038.3 | At what node count does recovering node allocation from mean facility power lose exactness? | a source suggests an answer (not checked) |
| P038.4 | Does adding wall/rack-level power to NVML raise detection of the hardest evasion strategies above a stated bar? | implied by a source, not stated as open |
| P038.5 | What is the range of the node/GPU energy ratio across AI workloads on one node type? | a source suggests an answer (not checked) |
| P038.6 | Is the gap between measured peak and rated node power stable across nodes and workloads? | a source suggests an answer (not checked) |
| P038.7 | Above what frequency does facility power electronics attenuate per-iteration load swings by a stated amount at the utility meter? | implied by a source, not stated as open |
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 P038.7: Above what frequency does facility power electronics attenuate per-iteration load swings by a stated amount at the utility meter?
* Full text not reviewed; more likely to contain errors.
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