Skip to main content

Lab Notebook · GPU-OS + inference

PRISM — a GPU operating system, and datacentre-class inference on a single 3090

PRISM treats a GPU the way an OS treats a CPU: a VRAM page table, model scheduling, custom CUDA kernels. It reached 147.2 tok/s on one RTX 3090. The engine internals are the crown jewel and are protected. The honest negatives — including a retracted ternary speed claim and structural GPU-utilisation limits — are kept on the page.

JR
Jon RossFounder, vocabotics — 15 years building safety-critical systemsLab report · dated 15 March 2026
Verified by a human. Drafted with AI, verified by a human. Jon Ross, 15 Mar 2026
Living document. Reviewed 15 Mar 2026
Entry date
15 March 2026
Category
Inference
Lead over the world
~1–2 years ahead
Access
🔒 Protected
ResearchPartly proven

We built an operating system for the GPU — VRAM as a page table, models as processes — and pushed a real model to 147 tokens a second on a gaming card.

147.2 tok/s
measured on one RTX 3090 (FP16)
2.8–4.5×
speculative decoding speedup
context from INT8 KV-cache (cos 0.9999)
~122K
lines of Rust in the engine + OS

Honest evaluation

Partly proven

The 147.2 tok/s inference-engine result is measured and real; a separate ternary-7B speed claim inside the same project was measured, found false, and retracted.

Runnable proof — see it work

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.

  • An on-disk ternary-7B speed claim was RETRACTED: the generic-kernel ternary path measured ~7× slower than Q4, and one on-disk perplexity figure did not reproduce. The footprint win is real; the speed win, so far, is not.
  • GPU utilisation is structurally limited (~14% on some paths) — a physics check, not a bug to be optimised away cheaply.
  • The wider platform this feeds still carries a large backlog of type errors and is not integrated end-to-end; 'compiles and runs in parts' is not 'shipped'.
  • The aspirational step-time targets remain honestly flagged as targets, not results.
Protected report

The deep body of “PRISM — a GPU operating system, and datacentre-class inference on a single 3090” 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.

  • GPU-as-OS inference147.2 tok/s on one 3090an inference engine that treats GPU/VRAM like an OS treats CPU/process memory — measured at 147.2 tok/s on a single RTX 3090 (engine source protected).

Where this connects

    We use cookies.