FD-SOI

last updated 2026-08-31 · +2 sources in last 30d
Assessmentdraft · unreviewed
Viability
5/5
Drivers
4/5
Novelty
2/5
Diffusion
3/5
Impact
3/5

TimingNow (0-2yr)·ReadUnderrated

FD-SOI is a planar CMOS process family built on a thin silicon layer over a buried oxide, sold by non-leading-edge foundries as a low-leakage, RF- and analog-friendly alternative to bulk CMOS and FinFET, and it has quietly become the default substrate for always-on edge silicon, embedded non-volatile memory and cryogenic quantum control chips.

Summary

FD-SOI (fully depleted silicon-on-insulator) is a manufacturing platform rather than a device concept. The transistor channel sits in a very thin silicon film separated from the bulk wafer by a buried oxide, so the channel is fully depleted without heavy doping and the substrate under the channel remains electrically accessible. The commercial instances that appear in the sources are GlobalFoundries’ 22FDX (22nm FD-SOI) and Samsung’s 28nm FD-SOI, with ST and GlobalFoundries running FD-SOI specialty work at Crolles under EU Chips Act alignment 2024 Eu Chips Act Final Text.

The practical consequences show up in the research literature. Because the PDK exposes transistor substrate terminals, designers can inject and measure substrate noise directly, which is how a 22nm FD-SOI ring VCO was used to characterise through-silicon-via coupling in 3D-IC stacks, yielding a -35.2 dBc sideband spur under a 1 GHz, 0.5 Vpp aggressor. The same platform is the vehicle for leakage-dominated always-on logic research, where feedforward leakage suppression logic is being explored specifically in FDSOI to cut standby power in the always-on domain of event-driven wearables and IoT devices. It is also where low-bit LLM accelerators are being prototyped: the AdaMX adaptive microscaling accelerator was implemented as a 22nm FD-SOI prototype.

Two commercial anchors matter more than the papers. First, embedded non-volatile memory: Samsung has mass-produced eMRAM on 28nm FD-SOI since 2019 and GlobalFoundries put eMRAM into production on 22FDX in February 2020, using perpendicular-MTJ STT-MRAM licensed from Everspin, in 2 Mb to 32 Mb macros with >10-year retention at 125 °C and survival through 260 °C solder reflow ref. GF’s 2026 investor material lists a broad eNVM portfolio (eMRAM, ReRAM, Flash) across FDX, FinFET and BCD, with FDX as the lead platform, framed around “autonomous secure always-on AI-enabled IoT edge devices” and 39B+ connected IoT devices by 2030 ref. Second, cryogenics: GF states its FDX platform delivers the cryogenic CMOS used for sensing, control and readout in quantum systems, built on more than a decade of US Government-partnered work, and has a $375M Department of Commerce letter of intent tied to a ~1% strategic equity stake ref. Academic work corroborates the cryo path: CryoZip was implemented in 22nm FDSOI characterised at 4 K.

The deciding parameters are not transistor physics but platform economics: which nodes carry FD-SOI, which eNVM and RF options are qualified on them, whether the PDK exposes the back-gate/substrate for designers who want it, and whether cryogenic characterisation becomes a supported PDK corner rather than a one-off measurement campaign.

Viability (5/5)

This is not a question of whether it works. Samsung has mass-produced eMRAM on 28nm FD-SOI since 2019 and GlobalFoundries has had eMRAM in production on 22FDX since February 2020, with published macro sizes, retention and reflow specs ref. GF’s own 2026 investor deck lists FDX as a lead platform across its eNVM portfolio ref.

TLDR: In volume production at two foundries with productised embedded memory and multiple independent silicon-level design efforts.

Drivers (4/5)

Supply: FD-SOI is the core of a deliberate “feature-rich, not node-leading” specialty-foundry position at GlobalFoundries, organised in its 2026 investor material around Physical AI and IoT edge devices ref. Public money reinforces it: ST/GlobalFoundries Crolles is listed as an FD-SOI specialty project aligned with the European Chips Act, which mobilised ~€43B public and private against a 20%-of-global-production target for 2030 2024 Eu Chips Act Final Text, and GF’s quantum business carries a $375M Department of Commerce letter of intent with a ~1% equity stake ref.

Demand: three distinct pulls are documented. Embedded non-volatile memory, where 28/22nm is the last node for embedded flash and eNVM (MRAM/ReRAM/FeRAM) is the stated migration path, with FDX carrying both eMRAM and ReRAM ref. Always-on edge devices where leakage in the always-on domain, not active energy, is the bottleneck. And cryogenic control and readout for quantum systems, where GF names Diraq, Equal1, Google Quantum AI, Microsoft, NVIDIA, PsiQuantum and Quantinuum as partners or customers ref.

TLDR: Foundry strategy and state subsidy on the supply side; edge AI, embedded NVM and quantum control on the demand side.

Novelty (2/5)

FD-SOI is not new: 28nm FD-SOI eMRAM has been in mass production since 2019 and 22FDX since 2020. The differentiators visible in the sources are qualitative. The PDK exposes transistor substrate terminals on RF devices, which is what made controlled substrate-noise injection possible in the TSV study. FDSOI is treated as the enabling process for feedforward leakage suppression logic in always-on domains, as against the HVT/UHVT bulk approach. And it is the platform GF asserts delivers cryogenic CMOS for quantum sensing, control and readout ref, corroborated by 22nm FDSOI being characterised at 4 K in academic work.

What the sources do not contain is a number. There is no measured power, area, RF or variability comparison of FD-SOI against bulk 28nm or against a comparable FinFET node in any supplied source, and the large memory ratios that do appear (~1000× faster write, ~1/400 write energy) belong to eMRAM versus eFlash, not to FD-SOI versus anything. The score reflects a mature, differentiated but unquantified platform, not a breakthrough.

TLDR: A decade-old platform with genuinely useful differentiators, but the sources contain no quantified head-to-head against bulk CMOS or FinFET.

Diffusion (3/5)

Diffusion is already underway in the segments where FD-SOI’s options matter: two foundries in production on FD-SOI eMRAM, a productised 22FDX eMRAM offering with automotive-relevant reflow and retention specs ref, European specialty capacity being funded at Crolles 2024 Eu Chips Act Final Text, and an unusually broad set of academic groups designing on 22nm FD-SOI across RF, digital accelerator, always-on logic and cryogenic domains.

The barriers are structural rather than technical. The supplied sources name only a handful of FD-SOI suppliers (GlobalFoundries, Samsung, ST/Crolles), so multi-sourcing is hard; the SPAD scan illustrates the general risk of single-foundry dependence in specialty silicon and treats second-sourcing onto GlobalFoundries as a structural necessity rather than a nice-to-have. GF also spreads its eNVM portfolio across FinFET and BCD as well as FDX ref, and TSMC’s Dresden build is described as 28/22nm plus 16/12nm specialty 2024 Eu Chips Act Final Text, so FD-SOI competes for the same designs against non-SOI specialty options. No source shows an FD-SOI node below 22nm, which caps the platform’s reach into higher-performance sockets.

TLDR: Real production and a live design ecosystem, but a narrow foundry set, no visible node roadmap below 22nm, and competition from bulk and FinFET for the same sockets.

Impact (3/5)

The upside case is one of enablement across several markets rather than a single large one. On the IoT side, GF frames 39B+ connected devices by 2030 as “virtually all requiring eNVM”, with FDX carrying that portfolio and the value proposition being standby/refresh power versus SRAM and I/O power versus external memory ref; the embedded eNVM market anchor in the research synthesis is ~$2.6B by 2029, which is a real but bounded number. On the quantum side, the leverage is disproportionate to the wafer volume: if cryogenic control and readout must be manufactured somewhere, and FDX is the platform GF is positioning for it with US Government backing and a partner list spanning superconducting, spin, ion-trap, photonic and topological modalities, FD-SOI sits on the critical path of a much larger system ref.

The ceiling is set by the platform’s own positioning. It is explicitly a feature-rich, non-node-leading proposition ref, so it will not capture datacentre-class compute value. Scored 3 because the sources support broad, durable relevance without any evidence of a step change in aggregate value creation.

TLDR: High leverage inside specific sockets (IoT eNVM, always-on edge, cryo control) with no claim on the leading edge.

Timing Now (0-2yr)

The core platform and its flagship option have been in production for years: 28nm FD-SOI eMRAM since 2019 and 22FDX eMRAM since February 2020 ref. ReRAM is a named multi-platform GF offering with 22FDX as lead ref. Nothing here requires waiting.

The newer directions are earlier but not distant. Cryogenic CMOS on FDX is being commercialised now, with a business launched in May 2026 and a $375M Department of Commerce letter of intent ref, while the supporting circuit work is at the characterised-prototype stage as of mid-2026. Edge-AI accelerator and always-on logic work on 22nm FD-SOI is likewise at prototype stage in 2026. European specialty capacity tied to Chips Act projects lands later, with TSMC Dresden opening 2027 2024 Eu Chips Act Final Text.

TLDR: Already shipping; the open question is the pace of the cryogenic and edge-AI extensions, not the platform.

Overrated or underrated? Underrated

FD-SOI reads as boring specialty silicon and is usually discussed as a node-race also-ran. The sources show something more useful: a single production platform that is simultaneously the lead vehicle for embedded MRAM and ReRAM as embedded flash hits its 28/22nm wall ref, the process being characterised at 4 K for quantum error-correction hardware, the platform a foundry is putting behind a state-backed quantum manufacturing business ref, and the process of choice for both RF substrate-noise work and always-on leakage suppression. Platforms that sit under several independently growing demand curves at once tend to be undervalued relative to whichever single application is fashionable.

The caveat is that this is a judgement about strategic position, not about measured superiority. No supplied source quantifies FD-SOI against bulk 28nm or a FinFET node on power, area or RF performance, and none shows a node below 22nm. If the differentiators turn out to be replicable on non-SOI specialty processes, notably GF’s own FinFET and BCD eNVM lines ref, the platform argument weakens to a wafer-cost argument. Anyone underwriting FD-SOI should be underwriting the sockets (eNVM, always-on, cryo control), not the substrate.

Prediction

By 31 December 2028, at least one of GlobalFoundries’ named quantum partners (Diraq, Equal1, Google Quantum AI, Microsoft, PsiQuantum, Quantinuum) will publicly disclose a cryogenic control or readout chip fabricated on GF’s FDX FD-SOI platform.

Evidence base

Open questions


Assessment drafted 2026-08-31 from up to 10 KB sources using the technology-scorecard framework; scores are a draft read pending review.

Recent mentions

Frontier questions