P099 Hardware scope and chip classification
Which chip specifications predict training throughput well enough to set a regime's hardware scope, and where do consumer/edge parts fall relative to that line?
Which chips and specs a regime covers, how chip capability is classified and measured, and where consumer-GPU and edge cases fall.
Terms: H100-equivalent (H100e): a common unit that counts mixed accelerators as multiples of one Nvidia H100 data-centre GPU, so fleets of different chips can be compared on one scale. N: the factor by which a cluster is larger when each chip has 1/N of the throughput.
Status
Analyzed Simulated (a distributed-training cost model calibrated to Covenant-72B, not a measured run): adds a 1,280 GB HBM memory threshold beside the 16-H100-equivalent cluster line for a 10^25 FLOP evasion, showing memory capacity, not only FLOP/s, matters for scope. Consumer/edge clause: the same desk analysis places 4-8 GPU workstations, DGX-class servers and GDDR parts below that threshold, also paper only. No throughput-prediction study; nothing measured. '1025' is the saved text's rendering of 10^25.
imposing the memory limit reduces the maximum unregistered per-node model from 240B to roughly 91B parameters, forces the evader to operate approximately five times as many nodes for equivalent model quality, and raises the cost of illicitly developing a 1025 FLOP model by roughly 50%.
Posed by
- Reuel, Bucknall et al. (2025), Open Problems in Technical AI Governance, compute sections:
how different chip specifications – such as throughput, memory bandwidth, memory capacity, and interconnect bandwidth – affect a chip’s suitability for different AI workloads
(§3.2.1, p.13, L673–674 (also Q13, L650–652)) (source states the problem but not the quantity the question asks for)
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); Scher & Thiergart (2024), Mechanisms to Verify International Agreements About AI Development (MIRI); Aarne & Petrie (2025), International Security Applications of Flexible Hardware-Enabled Guarantees (flexHEG Part 3); Aarne, Fist & Withers (2024), Secure, Governable Chips (CNAS); Shavit (2023), What Does It Take to Catch a Chinchilla? Verifying Rules on Large-Scale Neural Network Training via Compute Monitoring; Scher (2026), Verifying Restrictions on Frontier AI Research (MIRI); Ansari (2026), Hardware-Level Governance of AI Compute: A Feasibility Taxonomy for Regulatory Compliance and Treaty Verification; Koopmanschap & Barten (2026), How to Catch a GPU: A Taxonomy of Verification and Enforcement Mechanisms for International AI Agreements; Kulp et al. (2024), Hardware-Enabled Governance Mechanisms (RAND WR-A3056-1); Moon, Vedula, Geneson & Bar-on (2025), Strategies and Detection Gaps in a Game-Theoretic Model of Compute Governance (RAND RRA3686-1); Rahman (2026), Does Distributed Training Undermine Compute Governance? (MIRI); Dean, Verification Plan (AI 2040).
Sub-questions
| ID | Sub-question | What the sources report |
|---|---|---|
| P099.1 | Which chip specs (FLOP/s, memory BW, memory capacity, interconnect BW) predict training throughput on a frontier-style workload well enough to place thresholds, and which edge-case parts (A100, RTX 5090, CPUs with AI units) fall inside? | partly answered |
| P099.2 | At equal nominal FLOP/s, what training-time and cost penalty does an N× larger cluster of chips with 1/N per-chip throughput incur (slicing / pooling below single-chip thresholds)? | open |
| P099.3 | What share of global AI-relevant FLOP/s sits in data-centre accelerators vs consumer GPUs vs legacy chips vs phone SoCs, and what H100-equivalent fraction does consumer silicon deliver on a frontier training step? | open |
| P099.4 | What cost and wall-clock penalty does a 10^24 to 10^25 FLOP run on consumer GPUs (low-bandwidth interconnect) incur with current low-communication training methods? | open |
| P099.5 | After each export-threshold revision, how many months until a compliant SKU ships with a stated share of the restricted part's training throughput? | implied by a source, not stated as open |
| P099.6 | Does a per-cluster HBM-capacity cap (e.g. 1,280 GB) raise the simulated cost of a 10^25 FLOP evasion by a stated margin when the evader may use memory-heavy chips and pipeline-parallel low-communication training? | partly answered |
| P099.7 | What fraction of non-training cloud instance requests exceed a co-located GPU count, i.e. how specific is cluster size as a training flag? | implied by a source, not stated as open |
| P099.8 | Can teardown / delayering plus clock measurement bound a new accelerator's peak OP/s and interconnect bandwidth to a stated precision without vendor data? | partly answered |
Sub-questions not yet stated precisely enough to check
| ID | Sub-question |
|---|---|
| P099.9 | Should consumer GPUs be covered, AI-limited, or exempted (e.g. if stockpiling is blocked), and is recall feasible? |
| P099.10 | How do hardware trends and non-GPU substrates (neuromorphic, optical, quantum) change the governability of compute? |
| P099.11 | Can hardware configurations fixed at manufacture be adjusted when policy thresholds move? |
| P099.12 | What counts as an AI-relevant chip and as a 'data center' for registration and inspection? |
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 P099.1: Which chip specs (FLOP/s, memory BW, memory capacity, interconnect BW) predict training throughput on a frontier-style workload well enough to place thresholds, and which edge-case parts (A100, RTX 5090, CPUs with AI units) fall inside?
For P099.2: At equal nominal FLOP/s, what training-time and cost penalty does an N× larger cluster of chips with 1/N per-chip throughput incur (slicing / pooling below single-chip thresholds)?
For P099.4: What cost and wall-clock penalty does a 10^24 to 10^25 FLOP run on consumer GPUs (low-bandwidth interconnect) incur with current low-communication training methods?
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