Hardware sizing

ReduceWaste runs its local language model entirely on-site. This page is about sizing the box (or boxes) for that model tier — it is not about sensor hardware, which is covered per recipe in Sensors & recipes.

No prices or dates are promised here

Dollar figures below are list prices for third-party hardware, as published by its maker, not a ReduceWaste price. ReduceWaste itself has not published pricing, a beta factory list, or a launch date — treat any figure you see elsewhere as unconfirmed until it appears in an official release note.

Reference class: “GB10-class” unified-memory boxes

The baseline deployment target is a single box built around an NVIDIA GB10-class unified-memory module (128 GB of memory shared between CPU and GPU). Several vendors ship this class of box under their own branding; ReduceWaste does not require or endorse any one of them. As of this writing, published list prices for boxes in this class start in the high-$3,000s to mid-$4,000s (USD), with the exact figure and lineup changing as vendors update their own catalogs — check the vendor’s current listing rather than relying on a number here.

This project’s reference box

There is no GB10 box in this build. The reference CI/eval runner is an x86_64 box with 2× NVIDIA RTX 3090 (24 GB each, 48 GB total) and an i7-14700KF, running the baseline model as a GGUF split across both GPUs with llama.cpp. Discrete-GPU VRAM is summed across cards for sizing (unlike GB10’s single unified pool), not taken from the largest card alone. Throughput measured here is informational only — see spec/SPEC.md §7.1.

Sizing by site

Site	Boxes	Role split
Small (≤20 assets, ≤4 cameras)	1	Everything on one box: historian, vision, local inference, app
Medium (20–100 assets, 5–20 cameras)	2–3	Box A: historian + app + ingest. Box B: local inference. Box C (optional): camera-based vision, or small edge AI units placed at the cameras instead
Large / multi-line	3+	One vision-capable box per camera-heavy line, plus a larger inference pair for bigger models, or extra inference boxes for more concurrent users

Boxes are added by role first (give the model its own box before you need more throughput), and only pooled into a bigger-model configuration once a site’s answer quality needs actually call for it.

Memory planning

The default local model fits comfortably in one 128 GB box with room left over for the embedding model, the historian, and camera-based vision for a handful of cameras. A larger model tier for multi-line or very large sites needs either a pair of boxes linked for pooled memory, or a step up to a dedicated multi-GPU server — sized per project, not by a fixed table, since the right answer depends on how many lines and cameras are live at once.

CPU-only and smaller fallbacks

The demo plant and small installs can run a much smaller model profile on an ordinary 16–32 GB laptop or mini-PC. This is enough to try ReduceWaste and to drive the translation pipeline, but the full multi-step analysis workload needs the baseline tier above.

Camera compute

Camera-based counting and defect detection run as their own worker, independent of the language model. A few cameras run fine on the same box as everything else; a camera-heavy line benefits from its own vision box, or from placing small AI-capable units at the cameras themselves so only counts and events — not raw video — cross back to the core network.