P059 GPU confidential computing and link encryption: availability and overhead
Which shipping accelerators support multi-GPU confidential computing and encrypted links, at what throughput cost and with what key handling?
Which accelerators ship multi-GPU confidential computing / encrypted NVLink and links, at what throughput cost, with what binding and key handling.
Terms: confidential computing: keeping data encrypted even while a chip is processing it. NVLink: Nvidia's high-speed link between GPUs. AES-GCM: a standard method for encrypting and authenticating data in transit.
Status
Started Measured on one Huawei Ascend 910A accelerator, a vendor research prototype of firmware-only confidential computing: 0.028-0.91% added inference latency. Single device and inference only; no multi-device or training overhead has been published. A workshop participant reports an unpublished characterisation of training overhead from 2 to 1,024 GPUs; it is not counted, since no numbers are public. Availability: in January 2026 only the H100 offered working confidential computing; cluster-scale confidential computing on Blackwell at near-unencrypted speed is a vendor claim. Link encryption: roughly 3% of compute to encrypt NVLink traffic is a desk estimate.
introduces minimal performance loss during the inference pass(0.91% in GPT2-neo125M and 0.028% in Llama3-Chat-QA-1.5-8B model
Posed by
- Scher & Thiergart (2024), Mechanisms to Verify International Agreements About AI Development (MIRI):
TEEs on NVIDIA H100s (called “Confidential Computing”) currently only work on single-GPU setups (NVIDIA Trusted Computing Solutions Release Notes, n.d.).
(Evaluations › Verification Approach, p. 39, L1795–1796) (source states the problem but not the quantity the question asks for)
Also raised in: Reuel, Bucknall et al. (2025), Open Problems in Technical AI Governance, compute sections; 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); 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); Ansari (2026), Hardware-Level Governance of AI Compute: A Feasibility Taxonomy for Regulatory Compliance and Treaty Verification; Harack (2026), AI Verification: Infrastructure for Prosperity, Governance, and Peace (Lawfare); AI Security Forum, Requests for Discussion; Proofworks, Living Verification Documents.
Sub-questions
| ID | Sub-question | What the sources report |
|---|---|---|
| P059.1 | What throughput overhead does confidential computing impose on Hopper vs Blackwell for a fixed training and a fixed inference workload? | open |
| P059.2 | Which shipping accelerators (H100, B200, MI300X, TPU, Trainium, Ascend) support GPU CC today, and what fraction of installed data-centre FLOP/s does that cover? | a source suggests an answer (not checked) |
| P059.3 | Can a GPU attestation report be bound to a unique registry-checkable device identity and to a hash of the input data, surviving independent review? | open |
| P059.4 | Can current accelerators emit signed commitments to (code, input, output) hashes per job at line rate, with audited firmware and stated overhead? | open |
| P059.5 | Can an SM2/TCM-rooted Ascend attestation be translated into a DICE/SPDM-verifiable report by a neutral gateway? | open |
| P059.6 | At what sampling rate and aggregation does leakage from GPU power telemetry under CC fall below a stated leakage rate? | a source suggests an answer (not checked) |
| P059.7 | Is there a formally verified or audited hypervisor that hosts GPU AI workloads, at what overhead? | partly answered |
| P059.8 | Can PCIe/DMA reads of HBM by a third-party device be spoofed, and how far can a verifier trust them when CC blocks third-party DMA? | open |
| P059.9 | What is the measured overhead (% compute/power) and added latency (ns) of inline AES-GCM at NVLink-class rates (~900 GB/s)? | partly answered |
| P059.10 | Does InfiniBand support line-rate link encryption for inter-enclave traffic on current hardware? | a source suggests an answer (not checked) |
| P059.11 | For randomized authentication that samples a stated fraction of blocks, with a stated number of blocks modified, what is the detection probability, confirmed empirically? | partly answered |
| P059.12 | What throughput overhead does attested/confidential execution impose for multi-node training at a stated cluster scale (measured, not simulated), and is it production-available? | partly answered |
| P059.13 | What serving-throughput overhead does GPU confidential computing with per-response attestation impose in production? | open |
Sub-questions not yet stated precisely enough to check
| ID | Sub-question |
|---|---|
| P059.14 | How is AES-GCM nonce reuse prevented across a large cluster? |
| P059.15 | How is spoofing contained when one node holding source keys is compromised (and dropped packets handled)? |
| P059.16 | Can randomized authentication hide which packets are authenticated from side channels (always-on crypto, canary probing)? |
| P059.17 | Can devices keep verifiable logs of what results (e.g. weights) they sent to which other devices? |
| P059.18 | How is integrity of the whole hardware/firmware/framework stack (e.g. PyTorch) ensured for cryptographic audit trails? |
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 P059.1: What throughput overhead does confidential computing impose on Hopper vs Blackwell for a fixed training and a fixed inference workload?
For P059.3: Can a GPU attestation report be bound to a unique registry-checkable device identity and to a hash of the input data, surviving independent review?
For P059.5: Can an SM2/TCM-rooted Ascend attestation be translated into a DICE/SPDM-verifiable report by a neutral gateway?
For P059.9: What is the measured overhead (% compute/power) and added latency (ns) of inline AES-GCM at NVLink-class rates (~900 GB/s)?
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