TEST CHAMBER INFRASTRUCTURE

The Hardware

DOC: AFM-2026-HW-009  |  ONE LAPTOP · ONE GPU · 8 GB VRAM  |  STATUS: LOCAL INFERENCE PROVEN

No cluster. No cloud bill. One laptop: an AMD Ryzen 9 and an RTX 4070 Mobile with 8 GB of VRAM — the number everyone says is too small to run a 7-billion-parameter model alongside speech recognition. It isn't. You just can't hold both at once. Whisper transcribes, hands the card over, and the language model summarises. Sequential, not concurrent. The wall is a scheduling problem, not a wall.

8
GB VRAM
7B
Params, Local
2
Models, 1 Card
0
Cloud Required

Live: 8 GB, two models, one card

Try to load both models at once — the card runs out of memory. Switch to sequential (the real design) and the same two models fit fine, one handing off to the other. Sizes are approximate and include overhead; the lesson is exact.

8 GB CEILING
0.0 / 8.0 GB used
Load a model. Then try loading the other.

The rig

Specification
  • CPU AMD Ryzen 9 7845HX 12 cores, mobile
  • GPU RTX 4070 Laptop (AD106M) 8 GB GDDR6
  • RAM 16 GB DDR5
  • OS Windows 11 + WSL2 Docker GPU passthrough
  • PWR G-Helper TDP / fan profiles per task
Proven, not theorised

Docker GPU passthrough works--gpus all verified end-to-end by Vox on 2026-06-05. faster-whisper-medium (Speaches) and qwen2.5:7b (Ollama) both run on this card. Not simultaneously — a VramCooldownSeconds handoff sits between the transcribe phase and the summarise phase so the first model fully releases before the second loads.

1
🎙️
Load Whisper
transcribe all
2
♻️
Release VRAM
VramCooldownSeconds
3
🧠
Load qwen2.5:7b
summarise all
4
📝
Notes to the
Obsidian vault
Batching by phase — all transcriptions, then all summarisations — means the card only ever holds one model. It's the whole reason the "wall" isn't one.

The wall is a myth

"7B-on-8GB is a wall" is outdated.
whisper-medium + qwen2.5:7b run fine on one 8 GB card — sequentially

The belief that a 7B model won't share an 8 GB card with anything is a hosting reflex, not a hardware fact. Don't architect around the wall — architect the handoff. The same instinct that killed the Graveyard's early generations (self-host everything, concurrently, now) is the one this rig quietly disproves.

Planned upgrade: a second, older GPU (a 1070) dedicated to simple CRUD-grade inference — the small, constant, low-stakes calls — so the 4070 is free for the heavy sequential work. Priority scheduling, not more silicon. The rig grows by getting smarter about what it already has.

The subject runs a seven-billion-parameter language model and a speech recogniser on a laptop graphics card designed for video games. It works because he stopped asking the card to do two things at once — a lesson he has, I note, needed to learn in several other chambers as well. — GLaDOS, Facility Governor

One card, one thing at a time — the One Rule, in silicon. See it stated in The Prime Directive, the survivor it runs in The Graveyard, or the smaller builds in The Misc Wing. Back to the Facility Directory.