P075 Network interlock performance

Can an inline network interlock enforce a signed list of allowed peers at full line rate (800 Gb/s to 1.6 Tb/s)?

Whether an inline network interlock can enforce a signed peer allow-list at full line rate: latency, firmware updates, key setup, in-network-compute switches, NVLink traffic and DMA spoofing.

Status

Started Built, not measured: Amodo report a working prototype in which a central controller leases traffic quotas to NVIDIA BlueField network cards that enforce them, and hardware-accelerated encryption between nodes, with keys the controller can revoke. Membership by key is a form of the allowed-pairs list the problem asks about, but no peer allow-list has been built as such, and no line-rate or latency figures are published. The full interlock design is in flexHEG Part 2 (Petrie & Aarne 2025), with the network card as the retrofit location; encrypting all NVLink traffic is estimated at about 3% of H100 compute (a desk estimate).

Got hardware-accelerated encryption working
AI Security Forum (2025), Requests for Discussion, RFD 029 NIC-Based Bandwidth Limiting & Interconnect Verification (Amodo) (2024-11-27 (RFD created; implementation status as saved)), L1768

Posed by

Also raised in: Petrie & Aarne (2025), Technical Options for Flexible Hardware-Enabled Guarantees (flexHEG Part 2).

Sub-questions

IDSub-questionWhat the sources report
P075.1Can an inline network interlock sustain 800 Gb/s to 1.6 Tb/s with added latency below a stated bound (a target to be agreed)?open
P075.2Can ConnectX-class NICs be firmware-updated to enforce a signed peer allow-list without hardware change (versus costly NIC replacement)?open
P075.3What is the key-setup time for all-to-all session establishment across a stated number of nodes (a target to be agreed)?open

Sub-questions not yet stated precisely enough to check

IDSub-question
P075.4How are switches with in-network compute (e.g. SHARP) brought into the scheme as trusted nodes?
P075.5How does a NIC-based interlock handle NVLink traffic that never passes through the NIC?
P075.6If DMA reads can be spoofed, how much does network-only observation reduce security?

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 P075.1: Can an inline network interlock sustain 800 Gb/s to 1.6 Tb/s with added latency below a stated bound (a target to be agreed)?

For P075.2: Can ConnectX-class NICs be firmware-updated to enforce a signed peer allow-list without hardware change (versus costly NIC replacement)?

For P075.4: How are switches with in-network compute (e.g. SHARP) brought into the scheme as trusted nodes?

Data