P027 Enforcing and attesting cluster composition and interconnect limits
Can hardware fix which devices may link, cap inter-device bandwidth, and attest cluster size against bridging devices, at what training slowdown?
Hardware that fixes which devices may link, caps inter-device bandwidth, and attests cluster size/configuration against bridging devices.
Status
Started Amodo's own report of its own work, not independently checked: a central controller leases bandwidth and transfer-volume quotas to BlueField DPUs over an encrypted tunnel, the DPUs enforce them in software with the Linux traffic controller, and the system runs on one research server; hardware acceleration on BlueField-2/3 is listed as in progress. Not shown: a hardware-fixed set of permitted peers, attestation of cluster size against a bridging device or reconfiguration, holding a cap at line rate (the RFD lists 'can software tc keep up?' as open), and any training slowdown. Fixing which devices may link and attesting cluster size: paper designs only (RAND 2024 proposes restricting interconnect to a handful of preauthorized chips and says the mechanism 'does not yet exist'; flexHEG proposes verifying cluster configuration and outgoing bandwidth). Training slowdown when traffic is forced off NVLink: no work (RAND's 'might not present a significant performance penalty' is speculation).
"Phase 1: Software implementation (complete)" … "DPUs enforce quotas via Linux traffic controller (tc)" … "Running on Amodo research server"
Posed by
- Reuel, Bucknall et al. (2025), Open Problems in Technical AI Governance, compute sections:
It may be possible to build such a system on top of existing features such as trusted platform modules (Hosseinzadeh et al., 2019), or it may require new protocols and new hardware-level features – all of which are open problems.
(§6.2.3, p.36, L1989–1991) (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); Petrie, Aarne, Ammann & Dalrymple (2025), Overview of Flexible Hardware-Enabled Guarantees (flexHEG Part 1); Petrie & Aarne (2025), Technical Options for Flexible Hardware-Enabled Guarantees (flexHEG Part 2); 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; Ansari (2026), Hardware-Level Governance of AI Compute: A Feasibility Taxonomy for Regulatory Compliance and Treaty Verification; Kulp et al. (2024), Hardware-Enabled Governance Mechanisms (RAND WR-A3056-1); Rahman (2026), Does Distributed Training Undermine Compute Governance? (MIRI); AI Security Forum, Requests for Discussion; Proofworks, Living Verification Documents.
Sub-questions
| ID | Sub-question | What the sources report |
|---|---|---|
| P027.1 | Has a peer-attestation scheme been implemented in which compliant accelerators refuse links to non-attesting peers (blocking mixed clusters with untracked older chips), at what overhead? | open |
| P027.2 | Has a hardware-enabled export condition (location or cluster-size limit) been implemented on a shipping chip and survived a published circumvention attempt? | open |
| P027.3 | Has a model-tenancy ledger (append-only record of loaded model hashes) been implemented on an accelerator and shown resistant to rollback/erasure? | open |
| P027.4 | Has a governance-enforced bandwidth-capped configuration been demonstrated that holds inter-GPU bandwidth below a stated cap, and what training slowdown does it impose? | open |
| P027.5 | Can inspectors verify a link or cable is low-bandwidth (e.g. <=1 Gb/s vs 400 Gb/s) and that the cap cannot be raised without visible modification, including against supply-chain tampering? | open |
| P027.6 | Can a set of devices attest an upper bound on cluster size and outbound bandwidth that survives a red-team bridging device or reconfiguration (and is not broken by compromising a few weak devices)? | open |
| P027.7 | Can a pod's high-bandwidth link be extended to an unauthorized device via an interposer or retimer without failing attestation? | open |
| P027.8 | Is there a shipped attestation standard (e.g. SPDM) by which GPU, CPU and NIC in one server mutually attest identity? | open |
| P027.9 | Can a verifier determine after the fact (logs, attestations, inspection) whether chips were interconnected above a stated bandwidth, with what error rate? | open |
| P027.10 | What throughput penalty does a training job incur when NVLink traffic is forced onto PCIe/Ethernet with a communication-aware parallelism plan? | open |
| P027.11 | Can a periodic verifiable high-bandwidth window (e.g. once per 48 h) be implemented with a bound on the training traffic it admits? | open |
| P027.12 | Can a one-way (forward-only) interconnect be built and verified that blocks backward-pass traffic between pipeline stages? | open |
| P027.13 | Is there a switch bandwidth/port-count threshold below which training a 10^24 FLOP run on consumer GPUs becomes impractical (>10x cost)? | implied by a source, not stated as open |
| P027.14 | Up to what line rate does software (tc) quota enforcement on a DPU such as BlueField-3 hold? | partly answered |
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 P027.4: Has a governance-enforced bandwidth-capped configuration been demonstrated that holds inter-GPU bandwidth below a stated cap, and what training slowdown does it impose?
For P027.6: Can a set of devices attest an upper bound on cluster size and outbound bandwidth that survives a red-team bridging device or reconfiguration (and is not broken by compromising a few weak devices)?
For P027.7: Can a pod's high-bandwidth link be extended to an unauthorized device via an interposer or retimer without failing attestation?
For P027.14: Up to what line rate does software (tc) quota enforcement on a DPU such as BlueField-3 hold?
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