Sector · Semiconductors
Physical AI for semiconductors
A chip is specified, implemented, verified and handed to a foundry across separate tools, so the trace from a requirement to the check that proves it is usually rebuilt by hand. These are the questions engineers ask first, answered from 1 published CodeNinja Atoms reference architecture.
- What does a physical AI system for semiconductors look like?
- Which AI models can a semiconductor company run on its own hardware?
- How much compute and hardware does AI in semiconductors need?
- Is it cheaper to own AI hardware or rent cloud GPUs in semiconductors?
- What ontology or object model does a semiconductors AI system need?
- Who approves the decisions an AI system makes in semiconductors?
What does a physical AI system for semiconductors look like?
A complete physical AI design for semiconductors names what to sense, which existing systems to join, the object model that joins them, the models and hardware, the three-year cost and the person who approves every action. CodeNinja Atoms has published 1 such reference architecture for semiconductors, each free to reuse under CC BY 4.0.
| Design | Country | What it does |
|---|---|---|
| Silicon Ledger | Saudi Arabia | An open reference architecture for system design in physical AI, built for a fabless chip design company in Saudi Arabia that designs silicon in vendor design tool suites and hands tapeouts to an outside foundry, where owning the inference node that runs it costs about 569,000 US dollars over three years. Every requirement is traced to the block, interface and signal that implement it and to the check, test and coverage result that prove it, so the verification lead can ask which requirements have no proof and what one interface change on one block actually touches. The model runs entirely on the company's own hardware under the kingdom's data protection law, export control on the GPU node and trade secret law. |
Which AI models can a semiconductor company run on its own hardware?
Each published semiconductors design names its models and why, and every model is open-weight or no model is used at all, so the operator can run it on hardware it owns.
| Design | Models |
|---|---|
| Silicon Ledger | GLM 5.3 (bespoke licence, 753B at FP8) for the agentic work surface: the spec agent, the check agent and cited answers; BGE-M3 (MIT) for embedding and retrieval so every answer cites its record; PaddleOCR-VL 1.6 (Apache-2.0) for specification parsing at ingestion; no third party model service, nothing trained into weights |
| Design | Choice | What was picked | Why |
|---|---|---|---|
| Silicon Ledger | Work surface model | GLM 5.3, 753B parameters filed, FP8, bespoke licence, served on vLLM on the central node in the company's own server room | Bespoke licence allows commercial use, fine tuning and redistribution, so the company holds its weights outright; frontier class at or above the 300B threshold; carries the spec agent, the check agent and the cited answers |
| Silicon Ledger | Embedding model | BGE-M3, MIT, FP16, about 1.1 GB, on CPU or a modest GPU inside the perimeter | MIT licence with no field-of-use restriction; dense, sparse and multi-vector retrieval from one pass keeps requirement ids and signal names retrievable; multilingual coverage for a bilingual estate |
| Silicon Ledger | Document parsing OCR model | PaddleOCR-VL 1.6, Apache-2.0, about 0.96B parameters, BF16, about 1.9 GB, on a site GPU at ingestion | Apache-2.0 with no field-of-use restriction; table, formula and layout parsing is what a specification page is; Arabic coverage validated on the company's own documents in phase one with a human check stage |
| Silicon Ledger | Frontier hardware | 141 GB HBM GPU class, eight GPUs per node, one node (1,128 GB), roughly ten kilowatts | Holds 753 GB of FP8 weights; a 4A090 item for a D:4 destination, so a BIS licence determination is a phase one checkpoint; the platform runs on the company's existing hardware until the node lands |
| Silicon Ledger | Frontier footprint sizing rule | Frontier model GPU footprint sizing rule, vLLM and SGLang | 753 GB of weights times the 1.2 planning factor is 904 GB against the node's 1,128 GB |
| Silicon Ledger | Node count rule | GLM 5.3 node count by GPU class | One node of eight 141 GB GPUs, sized from the model's filed parameters on the stack card |
| Silicon Ledger | Sensing | None | A design office engagement with no cameras, no machine telemetry and no physical objects to track |
| Silicon Ledger | Pattern: governing | System of Context: engineer-confirmed extraction, tool-neutral assertions in version control, expert lessons as memory, provenance on every fact | Each control stops an error class: misread requirements, vendor lock-in, uninspectable reasoning and uncited statements; nothing is trained into weights |
| Silicon Ledger | Pattern: boundary | The company's own hardware on its own premises in Saudi Arabia; nothing crosses outward, inbound only through the Harbor offline registry | The company's stated constraint of no third party model service, with design data as its core intellectual property |
How much compute and hardware does AI in semiconductors need?
The compute follows from the models: the published semiconductors designs size it as follows, from no new hardware to a full GPU node.
| Design | Part | The design |
|---|---|---|
| Silicon Ledger | Compute | One node of eight 141 GB HBM cards (1,128 GB) in the company's own server room, holding 753 GB of FP8 weights with a 1.2 planning factor, 904 GB, at roughly ten kilowatts; no edge estate. The node is a 4A090 item for a D:4 destination, so a BIS licence determination is a phase one checkpoint, and the platform runs on the company's existing hardware until it lands. Nothing crosses outward |
Is it cheaper to own AI hardware or rent cloud GPUs in semiconductors?
Each semiconductors design prices three years of ownership in its Appendix A, with every price cited, against renting the same capacity from a cloud region at its deepest three-year commitment where hardware is bought.
| Design | Line | Three years |
|---|---|---|
| Silicon Ledger | Three-year cost | About 569,000 US dollars to own the node for three years (a range of 541,000 to 594,000: server, support and power at the Saudi commercial electricity tariff), against about 751,000 to rent the same eight cards on the deepest three-year commitment in AWS Middle East (UAE), so ownership costs about three quarters of it, and about 2.00 million on demand, against which ownership costs a little over a quarter. Renting is a price reference only, because AWS lists no Saudi region on the day read and every rented option would carry design records across the border (Appendix A) |
What ontology or object model does a semiconductors AI system need?
The object model is the part that makes the system an ontology rather than a pipeline: typed objects for the things in the semiconductors operation, with properties, status values and typed links. Every published design ships its object model as hyper-ontology/1 JSON that loads into Hyper Ontology.
| Design | Size | Objects | Download |
|---|---|---|---|
| Silicon Ledger | 16 objects, 13 links | Specification section, Requirement, Design block, Interface, Signal, Design assumption, Verification check, Test, Coverage result, Bug, Waiver, Design revision, Engineer, Verification lead, Expert lesson, Foundry handoff | objects.json |
Who approves the decisions an AI system makes in semiconductors?
In every published semiconductors design a named person makes the decision that changes the physical world; the system prepares it.
| Design | Human control |
|---|---|
| Silicon Ledger | Agents draft and engineers decide: an engineer confirms each extracted requirement before it is used, with the rejection rate as a phase one gate; the verification lead approves every drafted check before it enters the regression, escalating to the senior verification expert after five working days; a reconciliation job after every sync cycle flags stale impact lists and keeps them out of citations |
Every answer on this page is drawn from the papers linked in it: each paper's At a glance table, model register and cost appendix. Full text for agents: llms-full.txt. Designed on Praxis; object models load into Hyper Ontology.
Want one of these designs for your own operation? Talk to us.