Research Dossier ID-05 · Candidate · Ships January 2027
A strict black box with handles, an orange accent — and the first machine in this series where the interesting part is not the CPU generation. Two memory pools, one placement engine, and a research question hiding inside: can heterogeneity be an advantage if you stop forcing every part to do the same job?
Two pools, one engine, one PCIe pipe between them. Hover to see what each heart is for — and why their gigabytes must never be summed.
Plate M1 — two hearts, one chassis, overprinted
| In the brochure | What it actually says | |
|---|---|---|
| 192 GB + 32 GB = “224 GB VRAM” | → | Do not sum. Two pools with two bandwidths — 273 GB/s shared versus 640 GB/s private — connected by PCIe. The most honest number on the page is the pair of them, printed separately. |
| “RUNS MoE ACROSS BOTH” | → | Lucebox Engine places hot experts on the R9700 and the rest in Ryzen memory — where they are executed, not just parked. Grouped work, batched transfers, merge moved from CPU to GPU. That's real software engineering. It's also, so far, the manufacturer's own description and numbers. |
| UBUNTU SERVER · SETUP FROM PHONE | → | Initial setup from a laptop or phone, management in a browser, familiar apps pointed at a local API. Signed updates with health checks and automatic rollback. This rhymes directly with the Olares lesson: complex infrastructure inside, one short path for the user outside. Whether it holds after unboxing is the checkable question. |
| 340 × 110 × 320 MM · 6 KG | → | Almost twelve litres with handles — a compact workstation, not a mini cube. Compare it in the workstation group, or the size verdict will lie to you. |
| ~500 W LOAD · 40 W IDLE · 1,000 W PSU | → | Manufacturer figures. The 1,000 W supply is a ceiling, not a diet; 40 W idle — if true — is ≈ £92/yr at the current cap, livable. Every watt of this row is claimed, none is signed. |
| £5,999 → £7,999 · SHIPS JAN 2027 | → | UK prices with VAT and duty included — respect for printing pounds. But +£2,000 for the 128→192 GB step is saffron-grade arithmetic, and the box is a pre-order for a January batch, not stock on a shelf. |
| “FASTER THAN DGX SPARK” | → | The published comparison was run on the previous 395 build — different settings, compression and test modes. Useful signal, wrong generation. Zero 495 gets a clean sheet until its own signed runs. House rule. |
Plate C2 — memory as a space for objects, not a warehouse for weights
This is where the dossier stops asking the buyer's question — “how many GB per pound?” — and starts asking yours: “what other compute system can be built from these possibilities?”
| What we keep in memory | What work it enables |
|---|---|
| Model weights & separate experts | Interpretation, hypothesis-forming, candidate search — the LLM's day job. Lucebox's own pitch: placement of experts across pools, compute travelling with data. |
| Structured facts, roles, relations | Authority checks without re-reading entire history as text every time. Memory spent on representation, not just parameters. |
| Intermediate search states | Branching hypotheses, binding variables, rolling back failed branches — the FOL/PPU playground. This is the LPU question: which primitive operations deserve their own engine? |
| Conclusions + their evidence trails | Reuse results; on revocation, revisit only the affected conclusions. The compliance experiment that matters: fast propagation of “this is no longer allowed” through a delegation chain. |
Keep ALLOW next to the facts, rule version and mandate state it came from. When the mandate dies, revise the touched conclusions — don't replay the whole analysis. The quality bar: reused memory must never become reused stale permission.
“GPU proposes, CPU proves” is a habit, not a law. GPU-native Datalog exists in the literature — mass relational operations, recursive inference on the accelerator. The question: which parts of delegation-chain checking are really bulk relation operations? That's an experiment, not a purchase argument.
One fact base, many deltas: mandate expired / director changed / delegation capped / evidence voided. The useful output isn't “we ran more agents” — it's “the system found the minimal change that makes this action invalid, with a checkable explanation.”
Plate M3 — candidate estimates across two pools and one engine
CAPABILITY MAP STATUS: hypothesis / claim translation. Use the evidence label on each row before reading the adjective.
CONFIGURATOR STATES INCLUDED TAXES, DUTIES AND DELIVERY · FIRST-200 PRICING, BASE PRICE £6,999 QUOTED · SHIPS JANUARY 2027 · FILED AS AN OFFER FROM YOUR BASKET, NOT CONFIRMATION OF PURCHASE
Plate O4 — candidate entry: the Kettle test here is the whole product question
Plate K5 — fine print, blown up to full sheet
Plate Y6 — filed with interest, promoted on evidence
You want a finished AI product, not a parts list: placement engine, browser management, signed rollback-updates — and you accept January. Your workloads are large-MoE residents and studio-bound. The £2,000 memory step doesn't ruin your quarter.
You answer to tok/s per pound (a dedicated GPU wins), you need it now (January), or the vendor-only Engine makes you nervous — the placement magic lives and dies with Lucebox's software. Portable needs? It has handles. It also has 6 kg.
Does it help me do more, or give me more to maintain? If the Engine works as described, this is the first machine in the series where the software is the spec. If it doesn't, it's a lovely black box asking £9,048 for a ROCm apprenticeship. The index waits for owners, not launch videos.
CLOSING — RESEARCH RULE №1 · PLATE F
Every other dossier in this series asks what a machine costs and what it runs.
This one asks what it makes possible: compute travelling with data across two pools,
conclusions stored with their evidence, permissions revisited instead of replayed.
Inside, the architecture may be as complex as it likes.
Outside, a person should get a clear, checked, useful answer — and a system
that can show exactly which change made the answer stop being true.
That's not a product requirement. That's a research programme with a handle on top.