P066 Guarantee processor and verification co-hardware
Can a guarantee processor (verification co-hardware) be interposed on accelerator I/O at acceptable overhead, with binding that detects its removal?
Guarantee processors and verification co-hardware: overhead, binding, feature set.
Terms: guarantee processor (verification co-hardware): a separate, trusted chip placed alongside an accelerator to watch its inputs and outputs and enforce or attest rules.
Status
Analyzed A design, not a build: the flexHEG Guarantee Processor plus Interlock sitting on the accelerator data path, with four candidate placements and a qualitative latency/throughput trade-off (requirements rise the closer the interlock is to the die). Overhead clause: not analysed anywhere; Ansari 2026 says the performance cost of mediating all chip I/O is 'not yet addressed in the published literature'. Binding clause: only qualitative statements (retrofitted hardware 'may be easier for an attacker to decouple'; flexHEG Part 3 poses how the 'brain' could be integrated so it cannot be easily swapped out as an open question); no pairing scheme, no detection rate. The flexHEG team's power-measurement prototype is a retrofit guarantee processor that makes external measurements and does not sit on the I/O path, so it is a cross-reference, not evidence for this headline.
The Interlock could be implemented in several ways, including 1) as an IP block within the accelerator die, 2) as a chiplet sitting near the accelerator die, 3) as a part of high bandwidth memory (HBM), or 4) as a separate component that operates as a network relay or switch.
Posed by
- Ansari (2026), Hardware-Level Governance of AI Compute: A Feasibility Taxonomy for Regulatory Compliance and Treaty Verification:
performance cost of mediating all chip I/O is an open engineering question not yet addressed in the published literature.
(§3.1 M5, p.6; L306–307)
Also raised in: Harack et al. (2025), Verification for International AI Governance (Oxford AIGI); 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); Shavit (2023), What Does It Take to Catch a Chinchilla? Verifying Rules on Large-Scale Neural Network Training via Compute Monitoring; AI Security Forum, Requests for Discussion; Amodo (2026), AI 2040 Plan A — How You Can Help; Proofworks, Living Verification Documents.
Sub-questions
| ID | Sub-question | What the sources report |
|---|---|---|
| P066.1 | What throughput/latency overhead does a guarantee processor interposed on all chip I/O (NVLink-class) add, i.e. can the accelerator sit outside the trusted computing base at acceptable cost? | partly answered |
| P066.2 | Can a candidate guarantee processor recompute one SM's output stream in real time (throughput ≈1/192 of a Blackwell)? | open |
| P066.3 | Can existing on-accelerator processors (GSP, NIC, BMC, GPU RISC-V cores) serve as guarantee processors via firmware alone, per a published security analysis? | open |
| P066.4 | Has one guarantee-processor design been integrated on accelerators from multiple vendors on short notice? | open |
| P066.5 | Is there a binding method (pairing + tamper evidence) under which swapping or detaching a guarantee chip / mated host CPU is detected, with measured detection rate and time-to-detach in red-team tests? | open |
| P066.6 | Does a guarantee-processor candidate pass a published glitch-injection test suite / FIPS-like certification? | open |
| P066.7 | Can verification logic sit on or near HBM (programmable logic in the HBM stack, HBM guarantee chiplet) at the needed latency/throughput, and where should it sit (HBM, die, NIC)? | implied by a source, not stated as open |
| P066.8 | Has a dual-processor (one per bloc) output-matching design been built and tested to lock down on mismatch, given cross-bloc co-design is judged infeasible? | a source suggests an answer (not checked) |
| P066.9 | Do typical GPU servers have a free PCIe slot or port to attach a >=5 KB/s on-node witness non-invasively? | open |
| P066.10 | Can an SM be audited to interpret kernel code, and what area/power overhead does a per-SM kernel-logging block add? | open |
Sub-questions not yet stated precisely enough to check
| ID | Sub-question |
|---|---|
| P066.11 | Can guarantee-processor cryptography be side-channel-free and quantum-secure? |
| P066.12 | Are DC-SCM-based readings off the main data path spoofable? |
| P066.13 | What is the minimum viable guarantee processor / secure co-processor / verification ASIC feature set (secure boot, lockstep, months of logs, secure NVM)? |
| P066.14 | How can keys for kernel-level encryption be transported safely? |
| P066.15 | Can prover-made and verifier-made hardware be mixed in one node safely? |
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 P066.1: What throughput/latency overhead does a guarantee processor interposed on all chip I/O (NVLink-class) add, i.e. can the accelerator sit outside the trusted computing base at acceptable cost?
For P066.3: Can existing on-accelerator processors (GSP, NIC, BMC, GPU RISC-V cores) serve as guarantee processors via firmware alone, per a published security analysis?
For P066.5: Is there a binding method (pairing + tamper evidence) under which swapping or detaching a guarantee chip / mated host CPU is detected, with measured detection rate and time-to-detach in red-team tests?
For P066.7: Can verification logic sit on or near HBM (programmable logic in the HBM stack, HBM guarantee chiplet) at the needed latency/throughput, and where should it sit (HBM, die, NIC)?
For P066.10: Can an SM be audited to interpret kernel code, and what area/power overhead does a per-SM kernel-logging block add?
For P066.11: Can guarantee-processor cryptography be side-channel-free and quantum-secure?
For P066.13: What is the minimum viable guarantee processor / secure co-processor / verification ASIC feature set (secure boot, lockstep, months of logs, secure NVM)?
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