Reference Plant · Reference Fab 2.0 · FabONE
Reference Fab for a technology that has no reference fab.
FabONE is not a conventional CMOS front-end fab. It takes an already-fabricated electronic chip and adds a graphene photonic layer on top, at wafer level (BEOL) — the same graphene forms both modulators and photodetectors, CMOS-compatible, at any node. It is a 300 mm line built by Exyte (pilot line operational 2026, pilot production 2027, volume 2029, full volume 2031). That changes the benchmarking: the facility layer is highly comparable; the graphene layer needs an entitlement model, not an external benchmark.
Reference Fab 2.0
The original Reference Fab (Plieninger, Müller, Ehm & Reczek, Infineon, IEEE/SEMI ASMC 2001) set cost targets top-down and drove sites to catch up — and cut production cost more than 25%. It assumed a reference already existed. A first-of-its-kind fab breaks that assumption. Reference Fab 2.0 keeps the cycle and adds four things: an entitlement model for technology with no benchmark, explicit investment targeting, a continuous AI-assisted loop, and agentic data.
| Reference Fab (2001) | Reference Fab 2.0 | |
|---|---|---|
| Target basis | Best-in-class benchmark — a reference already exists | Entitlement when no reference exists — built from physics, equipment limits, lab and learning curves |
| Scope | Cost: capital · material · personnel · overhead | + Investment / CapEx targeting and ramp economics |
| Cadence | Annual budget cycle, monthly manual reporting | Continuous, AI-assisted; live controlling loop |
| Data | Internal + external benchmarking, gathered by hand | + Agentic ingestion, sparse-data modelling, anomaly detection |
| Output | Site targets + catch-up plan | + Cost per functional device, cash-burn, equipment-add timing |
Source: Original method: Plieninger, Müller, Ehm & Reczek, IEEE/SEMI ASMC 2001. Extensions are this demonstration’s proposal.
Two layers, two methods
Benchmarkable · ~60–70%
Facility, equipment, process flow
Cleanroom, media, equipment availability, MTBF/MTTR, utilization, throughput, WIP & cycle time, automation, energy, cost per wafer pass. Compared against the right universe: specialty, BEOL, MEMS and silicon-photonics fabs — not a 3 nm logic mega-fab.
Entitlement · the graphene core
No external benchmark
Single-crystal graphene, transfer yield, sheet/contact resistance, modulator & photodetector performance, functional optical yield, reliability. Instead of comparing to an “average fab”, you build an entitlement: what is technically possible once the process is stabilized.
The entitlement bridge
The Reference Fab logic (ideal → site-specific → catch-up) in FabONE’s language. The decisive advantage: the “ideal” curve is not hypothetical — it is normalized to 300 mm from real 200 mm platform data (Applied Nanolayers).
Cost per functional device · index
200 mm actual → 300 mm entitlement
Source: Illustrative index (entitlement = 100), not FabONE data. Shows how the methodology sets targets.
Move a lever, see the final cost
The bridge above is a snapshot. This is the live model: change a process lever and the cost per functional device updates — not just the yield — along with the ramp forecast toward the entitlement. The forecast is a learning-curve model today; in production it is a neural network trained on the 200 mm platform and live process data.
Process levers
Graphene transfer to the target wafer (BEOL).
Tungsten damascene contacts.
Modulators + photodetectors that meet spec.
Higher absorbs fixed cost over more wafers (index).
Cost / functional device · index
Ramp forecast · 2027 → 2031
Training the neural network…
Investment & cash-burn · index
Move a lever and everything updates together — cost per functional device, the neural-net floor, the ramp, and the investment it takes to get there. In production the network is trained on the 200 mm platform and live process data; here it trains on synthetic samples so the mechanics are visible.
A bridge chart for the technology itself
The same cost-build-up logic, but on the real graphene-photonics flow — the steps a conventional fab does not have. Each module carries its own efficiency target; the transfer and contact steps are where variability (and cost) concentrate.
Graphene-photonics process · index
Cost builds up module by module → cost / functional device
Source: Flow follows the published 300 mm CMOS process (arXiv:2304.02646): BEOL transfer, hardmask patterning, tungsten damascene contacts. Illustrative index (total = 100).
Economic cascade
Why yield is so costly in a first-of-its-kind line: each yield step amplifies the cost per functional device. At today’s published wafer-scale device yield (~74%), the amplification is steep — which is exactly why the entitlement target and the catch-up plan drive the economics.
Source: Illustrative index (processed wafer = 100). Line yield anchored to published ~74% wafer-scale graphene device yield; functional-optical yield is a placeholder, not FabONE data.
The technology, grounded in the literature
The entitlement is not a guess — it is anchored to what has already been demonstrated. Concrete anchors:
300 mm, today
Wafer-scale graphene modulators in a 300 mm CMOS platform: 50 dB/mm, 15 GHz, reproduced across 400 devices.
Yield anchor
~74% device yield at wafer scale; >95% monolayer uniformity on 150/300 mm wafers.
Headroom
Intrinsic 200 GHz RC-limited bandwidth — the ceiling the entitlement target reaches for.
Differentiator
One graphene layer → modulator + photodetector (57 mA/W, 64% modulation).
arXiv:2304.02646 · 2023
Wafer-scale Graphene Electro-absorption Modulators Fabricated in a 300 mm CMOS Platform
Chenghan Wu et al.
The proof that FabONE-class integration is real and repeatable: 50 ± 4 dB/mm modulation depth and 15.1 ± 1.8 GHz bandwidth, reproduced across 400 devices on 300 mm wafers, with tungsten damascene contacts.
arXiv:1402.7127 · 2014
Multifunctional graphene optical modulator and photodetector integrated on silicon waveguides
Nathan Youngblood et al.
One graphene device does both — 57 mA/W photodetection and 64% modulation. This dual function is Black Semiconductor's core differentiator.
arXiv:1302.3854 · 2013
CMOS-compatible graphene photodetector covering all optical communication bands
Andreas Pospischil et al.
Ultra-wideband, CMOS-compatible photodetection across every telecom band — beyond germanium's reach.
arXiv:2007.00955 · 2020
High-performance integrated graphene electro-optic modulator
Brian S. Lee et al.
12.6 GHz at room temperature with a 200 GHz RC-limited intrinsic ceiling — the headroom an entitlement target aims at.
How AI accelerates the target-setting
The 2001 method took two years of manual work to set up and ran on an annual cycle. For a fab that ramps a brand-new technology, the loop has to be faster and work from thin data. This is where AI earns its place — not as a slogan, but on specific jobs:
Entitlement from sparse data
Derive the 300 mm target from limited 200 mm runs plus physics and published limits — few-shot and transfer learning — instead of waiting for production volume.
Yield-ramp forecasting
Predict the catch-up curve and flag divergence early, turning the monthly controlling loop into a live one.
Process-data anomaly detection
Spot excursions in the transfer and contact steps — the highest-variability parts of the flow — before they become scrap.
Scenario & investment simulation
Run CapEx phasing and capacity-vs-cash-burn what-ifs across the 2027–2031 roadmap in seconds.
Agentic benchmarking & reporting
Continuously ingest internal and public benchmarks to keep targets current, and auto-draft the controlling read-out.
Target derivation
Translate best-demonstrated performance and equipment specs into per-category efficiency targets, consistently and transparently.
Black Semiconductor · cost & investment targeting
Who & what
Black Semiconductor (Aachen) integrates graphene photonics directly on electronic chips — the same graphene layer forms both modulators and photodetectors. In 2025 it acquired Applied Nanolayers, gaining a 200 mm graphene platform and team. FabONE, its first 300 mm line (~15,000 m²), is built by Exyte. Funding: €254.4 M raised, of which €228.7 M public via the EU IPCEI microelectronics programme. Roadmap: pilot line 2026, pilot production 2027, volume 2029, full volume 2031.
How the methodology applies
- • The 200 mm platform anchors the entitlement — the target is derived, not guessed.
- • Investment targeting: phase CapEx to the entitlement ramp — don’t over-equip before the yield entitlement is reached.
- • Cost per functional device drives pricing, capacity and the equipment-add decision.
- • The catch-up curve is the 2027–2031 plan, tracked monthly against budget.
A European consortium: STEP / IPCEI + Gravitnomad
FabONE is already backed by the EU IPCEI microelectronics programme. The EU STEP platform (Strategic Technologies for Europe) is a complementary instrument that channels funding into digital and deep-tech capacity — a natural fit for the operating-modellayer of a first-of-its-kind fab.
The gap a consortium fills
Capital and cleanroom are funded and under construction. What a first-of-its-kind line still lacks is a proven way to set targets, model cost and time investment for a technology with no precedent — the exact problem the Reference Fab was built for, now extended with entitlement modelling and AI.
Gravitnomad’s contribution
- • Reference Fab 2.0 as FabONE’s operating model — entitlement + investment targeting.
- • The AI layer: sparse-data entitlement, yield-ramp forecasting, agentic benchmarking.
- • A controlling loop connecting process performance → yield → cost → cash-burn, live.
- • Co-developed with the fab team (Werner Reczek, method co-author, as technical sponsor).
Proposal, not a claim:this is a collaboration model to discuss — positioning the methodology as FabONE’s target-setting and cost/investment system within a STEP/IPCEI-eligible consortium. No affiliation or agreement is implied.
The Reference Fab for FabONE — four modules
Facility & equipment reference
Footprint, CapEx, utilities, headcount, uptime, tool utilization, maintenance, automation, cost of ownership — a conventional 300 mm fab benchmark.
Process flow reference
Operations, cycle times, batch sizes, queues, rework, sampling, metrology, bottlenecks — modelled on the real line (growth → transfer → capping → patterning → contacts → optical devices → test).
Graphene technology entitlement
No external benchmark exists. Single-crystal growth, transfer yield, sheet & contact resistance, modulator / photodetector performance, functional optical yield, reliability — a target built from physics, equipment specs, the 200 mm platform and learning curves.
Economic & investment model
The decisive link: cost per processed wafer → per good wafer → per functional device, yield loss per step, cash burn, and when to add equipment across the 2027–2031 roadmap.
Basis & honesty
Public facts: Black Semiconductor, Exyte & press (2025–26) — 300 mm FabONE built by Exyte; roadmap pilot line 2026 / pilot production 2027 / volume 2029 / full volume 2031; €228.7 M public funding via EU IPCEI; the 200 mm graphene platform (Applied Nanolayers) acquired in 2025. Device and yield figures are drawn from the published literature, cited above. All cost/index figures are illustrative (entitlement = 100), not FabONE data — they show how the Reference Fab methodology would set targets. The real entitlement is built from the 200 mm platform, equipment specs and lab results.
Source: R. Plieninger, U. Müller, H. Ehm & W. Reczek, “Cost Reduction using Systematic Target Setting of the Reference Fab Methodology,” 2001 IEEE/SEMI ASMC — the underlying methodology. Graphene device data: arXiv 2304.02646, 1402.7127, 1302.3854, 2007.00955.