A Co-Packaged 5G Radio That Proves Where Your GPU Actually Is

Hardware-rooted distance bounding via 5G URLLC narrows round-trip jitter 10× vs default 5G, rejecting relay attacks from 300 km in simulation. Export-control enforcement gets a physical-layer anchor.

Imagine you're a bouncer checking IDs at a nightclub door. You can ask someone to shout back a passphrase, and you time how long the echo takes to estimate how far away they really are. But if the street is noisy — wind, traffic, other voices — you have to allow a wider margin, and a clever impostor standing further away can exploit that slack by relaying the passphrase over a walkie-talkie. The trick isn't to shout louder; it's to run a private phone line from your ear directly to their mouth, cutting out the street noise entirely. That's what 5G CATS does, except the 'phone line' is a 5G URLLC radio co-packaged on the GPU die, and the 'street noise' is the host operating system, PCIe bus, and baseboard management controller that normally sit between a chip and any network verification signal. The committed claim: by terminating the challenge-response radio link inside the GPU package itself — bypassing the entire untrusted software stack — you can tighten the round-trip time distribution enough to detect fiber-relay attacks from adversaries as close as 300 km, where default 5G cannot. The system replaces permanent kill-switches with short-lived execution leases and cross-checks reported cell identity against the carrier's own records. The ladder here is PowerCATS, a 2024 system from some of the same authors that used power-line communication for the same goal. PowerCATS achieved a safe tracer distance of 619 km against a 3,000 km attacker. 5G CATS pushes that to 1,410 km — a 2.3× improvement in operational radius. Against a closer 300 km relay, PowerCATS's published model cannot guarantee rejection; 5G CATS rejects every round. Default 5G is never safe at any tracer distance against a 3,000 km adversary. These are meaningful, named comparisons with real numbers. Architecturally, this is a distance-bounding protocol layered onto 3GPP Release 16 URLLC slot structure. The key structural bet is co-packaging: the 5G radio, a cryptographic engine, and a hardware interlock all live on the GPU package, so the verification path never touches software the host controls. The simulation runs in ns-3 with the 5G-LENA module, modeling an unloaded cell. The 99.9th percentile round-trip spread narrows to 2.60 ms under the URLLC radio profile, versus roughly 26+ ms for default 5G. That 10× compression is what buys the detection margin. Integrity is the obvious pressure point. Everything here is simulation — ns-3/5G-LENA, not a real base station, not a real co-packaged GPU. The unloaded-cell assumption is generous; real cells carry traffic that would widen jitter. No hardware prototype exists. The authors are straightforward about this scope, but the gap between 'ns-3 says yes' and 'deployed in a Tier-1 carrier network with loaded cells, multipath, and real adversaries' is large. The comparison against PowerCATS uses PowerCATS's published model rather than head-to-head hardware, which is honest framing but not independent validation. The milestone that matters is a hardware demonstrator: a co-packaged 5G radio on an actual GPU substrate, tested against a real URLLC base station under loaded-cell conditions. That's the gate between 'promising architecture' and 'deployable export-control mechanism.' The paper doesn't estimate when that could happen, but co-packaging RF with logic is an active area in advanced packaging (chiplets, 2.5D/3D integration), so the timeline is plausibly 2-4 years for a lab prototype if a motivated actor (a hyperscaler, a government lab) funds it. The obvious experiment not run: loaded-cell simulation with realistic traffic and multipath. ns-3/5G-LENA can model this; the authors chose unloaded conditions. The charitable read is (a) — they scoped to the clean case first to establish the floor, and loaded-cell analysis is the natural next paper. The less charitable read is that loaded-cell jitter could erode the 10× advantage significantly, and they wanted the clean result on record first. Either way, it's the single most important follow-on.