Skip to main content

Lab Notebook · Inter-Process Wire

NovaP — 15 bytes a token, and why the plumbing mattered more than the demo

NovaP is a binary IPC protocol built for the GPU-inference stack: 15 bytes per token against roughly 200 for an equivalent JSON message, over a lock-free ring buffer. Unglamorous, and load-bearing for every token-throughput number built on top of it.

JR
Jon RossFounder, vocabotics — 15 years building safety-critical systemsLab report · dated 1 February 2026
Verified by a human. Drafted with AI, verified by a human. Jon Ross, 1 Feb 2026
Living document. Reviewed 1 Feb 2026
Entry date
1 February 2026
Category
Inference
Lead over the world
a working baseline, not a lead
Access
🔒 Protected
ResearchPartly proven

No demo, no screenshot — just 15 bytes a token instead of 200, and every token-throughput number upstream of it depends on that number holding.

15 bytes
per token, vs. ~200 bytes for an equivalent JSON message (13.3x)
WITHDRAWN
a '~104x faster than TCP' figure we can no longer source to a harness
24
tests

Honest evaluation

Partly proven

The wire-size result is proven for the specific SPSC workload and JSON baseline measured. The speed multiple that used to sit beside it is withdrawn — we could not locate the harness behind it.

What we learned — a preview

The honest edges are open even when the engine is not. The full report — what it was, what we built, and the measured internals — is protected.

  • We previously published '~104x faster than a comparable TCP-based approach'. Reviewing our own records we could not find the harness, the message size, the hardware or the command behind it, so the figure is withdrawn rather than restated at a lower value. It may well have been real; we cannot show you that it was.
  • The 13.3x is arithmetic on two stated byte counts (200 ÷ 15) and is presented as a computed ratio, not a timing measurement. It is workload-specific and measured against a naive JSON baseline, not a claim of beating every real-time IPC approach in general.
  • A ring-buffer IPC beating JSON-over-TCP on a loopback micro-benchmark is unsurprising to anyone who knows the domain. The number bought us less than it risked, which is part of why it went.
  • A lock-free single-producer/single-consumer ring buffer constrains the topology it can serve — it is not a general multi-producer message bus.
  • There is a second, unrelated project in the archive also named "NovaP" — a sensor wire format (magic bytes NVP1) for the Novaterra wearable hardware family, with no relationship to this GPU-inference IPC protocol beyond the shared name and a house habit of reaching for binary-over-generic on a hot path. See the sensor one at /lab/novap-sensor-wire-protocol if that's what you were looking for.
Protected report

The deep body of “NovaP — 15 bytes a token, and why the plumbing mattered more than the demo” is behind access.

We open the demos, the specs, and the method; we protect the engines, the model internals, and anything that touches commercial, safety-critical work. This report describes an engine — so its details are gated, and no client, company, or project is named. Members can read it in full; if you have a genuine reason to see it, tell us who you are and why.

Proofs & sparks

We demonstrate rather than assert. Each ✅ proof is a visible result with a hard figure.

  • Binary IPC wire protocol15 bytes/token vs ~200a purpose-built binary IPC wire protocol over a lock-free ring buffer — 15 bytes per token against roughly 200 for an equivalent JSON message on the inference hot path. A '~104× faster than JSON/TCP' figure previously published here is withdrawn: we could not locate the harness behind it.

Where this connects

    We use cookies.