01 / 07 · THE OUTPUT · CLIP
Two desk-side AI systems, ninety days, ten things to prove.
The question: what can a 128 GB box under a desk do for a creative pipeline, measured, not assumed?
This room is a Gaussian splat trained on the boxes in 39 minutes, from 3,220 renders of a fictional interior.
02 / 07 · LIVE · SPLAT
The same room, running in this page.
Load it and drag to look around. It holds across the room, not at one camera position.
Reduced from 2.35 million splats for the web. Soft floaters remain in the upper corners.
03 / 07 · WHAT ONLY THIS BOX DOES · PLATE
An 80-billion-parameter model, on one box.
Peak use was 121.2 of 121.7 GiB. No quantization of this model fits a 32 GB workstation card.
That is the case for the hardware: not speed, but models that do not otherwise run locally.
04 / 07 · ONE PHOTOGRAPH IN · CLIP
A room from a single photograph. At one spot.
Photograph to panorama to camera moves to splat: 2 h 24 min and 139 Wh on one box, unattended.
It reads as a room only near where the photograph was taken. Move a few metres and it smears. Reported that way.
05 / 07 · DAILIES · CLIP
A still becomes a camera move overnight.
81 frames at 720p in 12 minutes, drawing 90 W.
A power-capped workstation GPU does the same job 2.1 to 2.7 times faster. The box wins on memory and energy, not on speed.
| RI | What it does | Reading | State |
|---|---|---|---|
| 01 | Overnight versioning and conform | 23 ms per file | Installed |
| 02 | World from a photograph | 2 h 24 min, 139 Wh | One station only |
| 03 | Product re-renders | 4 candidates in 46 s | Label text fails |
| 04 | Retailer treatments + compliance gate | 40 ms per file | Installed |
| 05 | Resident local assistant | 19 / 19 chores, ~52 tok/s | Takes the whole box |
| 06 | Decisions watcher | demo run only | Not installed |
| 07 | Batch quality pre-review | 40 images in 9 min | Unvalidated |
| 08 | Detail and upscale | 4× in 135 s at 52 W | Installed |
| 09 | Dailies-grade video | 720p clip in 12 min | Installed |
| 10 | Model evaluation lane | 4 of 5 task classes pass | One class fails |
06 / 07 · THE MENU · LOG
Stand it up, meter it, say what it is.
Each one was installed from a written recipe, run on a fixed job, and logged: time, memory, watts, and whether the output earned its headline.
This is the customer-engineering job in miniature.
- 9 / 10reference implementations installed
- 121.2 GiBpeak, one model, one box
- 2–4×slower where a workstation card fits
| Item | Reading | State |
|---|---|---|
| Reference implementations | 9 of 10 installed and metered | Done |
| Large-memory models | 80B model, 121.2 of 121.7 GiB | Runs |
| Energy, same job | 101 W peak against 466 W | Measured |
| Speed, where a 32 GB card fits | 2–4× slower than the workstation | Measured |
| Thermal headroom | 95.9 °C peak, 96 °C abort line | No margin |
| World from a photograph | holds at the photographed station only | Open |
| Real-time version of the room | 6.4 / 10 realism against an 8.5 target | Stopped by rule |
07 / 07 · BENCH LOG · LOG
What works, and what does not yet.
The findings go out with the misses in them. A recommendation is only useful if the reader can see where it stops.