MENU

Heterogeneous Integration Explained

Back to Technology & Company Guides

TECHNOLOGY EXPLAINER

What Heterogeneous Integration Is
— not a technique for splitting things up, but for putting unlike things together

The term is often used interchangeably with "chiplets", but it covers a wider field — and from a materials standpoint, a far harder one. This article looks at what it actually means to put compound semiconductors, and light itself, into the same package as silicon.

Built from primary sources published by IEEE EPS (HIR), imec, TSMC and SEMI / Last updated September 2026

Conceptual image of several dies with different colours and textures mounted side by side on a single package
Conceptual image (AI-generated). An impression of components of different origins integrated into one package. It does not represent a real product's configuration, materials or layout.
What this article covers
  1. What heterogeneous integration is (the short version)
  2. How it differs from chiplets — two terms that get conflated
  3. "Different" comes in four degrees
  4. Case 1: putting compound semiconductors on silicon (imec)
  5. Case 2: electronics and photonics in one stack (TSMC COUPE)
  6. A materials engineer's view: three mismatches
  7. What eight years did to bonding accuracy
  8. What is still hard
  9. Glossary / Primary sources / Claim-to-source audit
How claims are labelled in this article

Sourced = a value stated in published material from a research institute, standards body or manufacturer (link given)
Our calculation = a figure this article derived, with the assumptions spelled out
Not yet confirmed = a plan or target with no confirmed production record

1. What heterogeneous integration is (the short version)

Heterogeneous integration has a definition the industry actually refers to. The Heterogeneous Integration Roadmap (HIR), led by the IEEE Electronics Packaging Society (EPS), puts it this waySourced.

The HIR definition

"Heterogeneous Integration refers to the integration of separately manufactured components into a higher-level assembly that, in the aggregate, provides enhanced functionality and improved operating characteristics." (IEEE EPS, Heterogeneous Integration Roadmap)

Three things in that sentence do the work.

  • Separately manufactured: these parts were not made together, in one plant, in one process flow.
  • Into a higher-level assembly: the package is where they become one thing.
  • In the aggregate: the goal is more than the sum of the parts.

The HIR is supported jointly by IEEE EPS, SEMI, the IEEE Electron Devices Society (EDS), the IEEE Photonics Society and the ASME EPPD, and it sets out packaging, test and interconnect technology requirements over a 15-year horizon (25 years for emerging research areas)Sourced. Its stated mission is to identify technical challenges with enough lead time that they do not become barriers to progressSourced.

2. How it differs from chiplets — two terms that get conflated

The two words are often used as if they meant the same thing. In fact they point in opposite directions.

ChipletsHeterogeneous integration
Direction of thoughtDividingCombining
Starting pointA chip that has grown too large to build as one pieceFunctions or materials that cannot be built as one piece at all
Main motivationYield, reticle limit, costCapability that a single die cannot deliver, or cannot deliver economically
Typical exampleSplitting a compute die into severalPutting compound semiconductors or photonic devices onto silicon
RelationshipChiplets are one form of heterogeneous integration, not a separate thing

The programme manager of imec's 3D system integration programme told SEMI that "we see heterogeneous integration as a scaling booster for functional partitioning, and as a way to create systems that would be impossible, or economically unviable, as a single chip"Sourced (SEMI, 2 October 2018).

"Impossible as a single chip" is the dividing line. Chiplets split up something that could, in principle, be made as one die. Heterogeneous integration deals with things that could never have been one die in the first place.

Fig. 1 · How monolithic, chiplet and heterogeneous integration relate
Conceptual illustration of How monolithic, chiplet and heterogeneous integration relate
Fig. 1 Conceptual image (AI-generated). This shows how the terms nest inside one another. It is not a depiction of physical structure.

3. "Different" comes in four degrees

"Heterogeneous" covers a wide range. How hard the job is depends entirely on what is different and by how much. Re-sorting the technologies this series has covered along that axis makes the landscape much clearer.

Fig. 2 · Four degrees of difference in heterogeneous integration
Conceptual illustration of Four degrees of difference in heterogeneous integration
Fig. 2 Conceptual image (AI-generated). Our own grouping, not a standardised industry classification. The ordering by difficulty is conceptual.

Levels 1 and 2 are silicon meeting silicon. Thermal expansion, cleaning chemistries and handling know-how are shared. At levels 3 and 4, the materials themselves differ. That is where "heterogeneous" starts meaning what it says.

4. Case 1: putting compound semiconductors on silicon (imec)

On 11 June 2026, imec announced a 300 mm RF silicon interposer platform that combines III-V materials such as InP, GaAs and GaN with the scalability and cost efficiency of Si-CMOS technologySourced (imec).

Why put compound semiconductors on silicon at all?

Silicon is not good at everything. Amplification at millimetre-wave and sub-THz frequencies, and emitting light, are things III-V semiconductors such as InP and GaAs do well. High volume, low cost and large-scale integration, meanwhile, belong to silicon. The idea is to combine each material where it is strongest, inside one package.

The three building blocks as published

ElementWhat was published
High-density embedded capacitorsA new MIMCAP structure gives 10 to 100 times the capacitance density of typical on-chip capacitors in III-V technology, pairing high-k aluminium hafnium oxide with a 3D oxide stud structure in the BEOL
Passive component modellingA framework validated to roughly 300 GHz (sub-THz), able to predict circuit performance accurately when geometry changes
Laser-assisted bondingAlignment accuracy below 600 nm and rotational misalignment below 0.05 degrees across 43 devices, while holding reflection below -15 dB in the 110 to 170 GHz band

All rows Sourced (imec announcement, 11 June 2026). These are results from a research platform and do not represent production-level figures.

What a materials engineer should notice here

Look at how the metric is framed: rotational misalignment below 0.05 degrees across 43 devices. Translational placement error (600 nm) is reported separately from rotation.

That split is specific to high-frequency circuits. At millimetre-wave and sub-THz frequencies, small differences in conductor length or relative angle turn directly into performance spread. What matters is not whether the joint is connected, but how precisely it was placed and oriented — bonding has stopped being a mechanical step and become a step that builds the electrical characteristics.

Conceptual image of small compound semiconductor dies of a different hue placed precisely on a large silicon interposer
Fig. 3 Conceptual image (AI-generated). An impression of dies of different material systems placed on silicon. It does not show real colours, dimensions or layout.

5. Case 2: electronics and photonics in one stack (TSMC COUPE)

One level harder is putting two different kinds of physics — electrical and optical — into the same package.

TSMC's COUPE (COmpact Universal Photonic Engine) uses SoIC-X die-stacking technology to stack an electrical die on top of a photonic dieSourced. TSMC states that this arrangement achieves the lowest impedance at the die-to-die interface and delivers better energy efficiency than conventional stacking approachesSourced (TSMC, 24 April 2024).

The timeline is published too: qualification for small pluggables in 2025, and integration into CoWoS in 2026 as co-packaged optics (CPO), bringing optical connectivity inside the packageSourced.

Why "not mechanically weak" is worth saying out loud

TSMC's technical paper says of COUPE that "for both grating coupler and edge coupler, the structure is solid without any cavity or mechanically weak region, and therefore can achieve low insertion loss without contamination or mechanical concerns"Sourced (TSMC Research, 2021).

Optical components are sometimes built with a cavity so that light has a clear path. But a cavity is a mechanical weak point and a place for contamination to collect. The structure that serves the optics undermines the reliability of the assembly — and that sentence is TSMC saying it sidestepped the conflict by building solid. In heterogeneous integration, the logic of performance and the logic of assembly collide constantly.

Conceptual image of an electrical die stacked on a die carrying optical waveguides, with optical fibres entering from the side
Fig. 4 Conceptual image (AI-generated). An impression of a photonic die with an electrical die stacked on it. It does not accurately show real optical structures, coupling schemes or dimensions.

6. A materials engineer's view: three mismatches

Seen from the materials side, everything that makes heterogeneous integration hard reduces to three mismatches.

Fig. 5 · Three mismatches in heterogeneous integration
Conceptual illustration of Three mismatches in heterogeneous integration
Fig. 5 Conceptual image (AI-generated). A summary of general relationships in materials engineering and process integration. It does not state property values or process conditions for any specific material.

The materials SEMI and imec single out

In the SEMI interview quoted earlier, imec's programme manager said that heterogeneous integration is pushing materials to their limits, and named these specificallySourced:

  • TBM (temporary bond materials) — adhesives that hold a thin die for handling and then release it
  • WLUF (wafer-level underfill)
  • Photosensitive polymers for fine line and space (covered in the RDL article)

On die-to-wafer assembly, the same interview said that reaching pitches well below 10 µm remained a challenge in terms of alignment and cleanlinessSourced. That statement dates from 2018. The next section looks at what happened afterwards.

7. What eight years did to bonding accuracy

Whether heterogeneous integration moves from concept to product depends on how accurately unlike things can be placed. Laying the published numbers out in order makes the trajectory visible.

Bonding accuracy, as published over time October 2018 (SEMI / imec) Die-to-wafer pitch well below 10 um: not yet reached Alignment and cleanliness were the blockers May 2024 (imec) D2W hybrid bonding demonstrated at 2 um pitch Overlay error below 350 nm; plasma dicing used June 2026 (imec) III-V chiplets joined by laser-assisted bonding Alignment below 600 nm; rotation below 0.05 deg In eight years, what was called out of reach moved into demonstration
Fig. 6 Drawn from published figures. Each entry reflects what was published at that time and does not represent production-level capability. The three results differ in purpose and measurement conditions, so they cannot be read as one continuous progression.
How to read this figure

The three numbers come from different technologies measured under different conditions (D2W hybrid bonding and laser-assisted bonding are not the same joining method). This is not a like-for-like comparison. Treat it as an illustration that demonstrations are now reaching territory once described as a challenge.

8. What is still hard

(1) Heat — parts with very different power densities share one lid

In heterogeneous integration, components with different power consumption and different thermal behaviour end up in one enclosure. The compute die runs hot, photonic devices are sensitive to temperature change, and compound semiconductors have their own temperature characteristics. One cooling design has to satisfy all of them, and that is getting harder.

(2) Design — which function belongs on which layer

Design was the first item raised in the SEMI interview: "if you break up a large chip you have to decide how to reassemble it, and which function belongs at which level of the hierarchy. We would like to do that with a tool set that supports the designer, rather than by hand"Sourced.

(3) Test

Once you are gathering parts made in different plants and assembling them, the advantage disappears unless every part can be confirmed good before assembly. And screening a photonic or high-frequency device needs different equipment from screening logic.

Conceptual image of a hot die and a cool die sitting next to each other on the same package, giving an uneven temperature distribution
Fig. 7 Conceptual image (AI-generated). A qualitative impression of the uneven temperature distribution that results when components with different thermal behaviour share a package. It is not a thermal simulation result and shows no specific temperatures.

9. Glossary

Heterogeneous integration
Integrating separately manufactured components into a higher-level assembly that, in the aggregate, improves function and operating characteristics.
HIR
Heterogeneous Integration Roadmap. An industry roadmap led by IEEE EPS and partners.
Monolithic
Building every function into a single piece of silicon. The opposite of heterogeneous integration.
III-V semiconductors
InP, GaAs, GaN and similar. Strong at high-frequency operation and light emission, where silicon is weak.
InP
Indium phosphide. Used for light-emitting and light-detecting devices in optical communication.
GaN
Gallium nitride. Used in high-frequency and high-power devices.
MIMCAP
Metal-insulator-metal capacitor. A capacitor built into the interconnect layers.
high-k
An insulator with high permittivity, used to get large capacitance from a small area.
BEOL
Back End Of Line. The wiring steps that follow transistor formation.
Laser-assisted bonding
Joining by heating locally with a laser, which limits the thermal load on everything nearby.
Silicon photonics
Building optical circuits on silicon.
CPO
Co-Packaged Optics. Optical components housed in the same package as the processor.
EIC / PIC
Electronic integrated circuit / photonic integrated circuit. COUPE stacks one on the other.
Grating coupler / edge coupler
Two ways of connecting an optical fibre to a waveguide on the chip.
Insertion loss
Signal lost in passing through a component. Lower is better.
TBM
Temporary Bond Material. An adhesive used to hold a thin die during handling.
WLUF
Wafer Level UnderFill. Underfill applied while still at wafer level.

10. Primary sources

  1. IEEE EPS "Heterogeneous Integration Roadmap" — eps.ieee.org
  2. imec "imec unlocks system-level III-V chiplet integration on Si-CMOS, advancing its 300mm RF silicon interposer", 11 June 2026 — imec-int.com
  3. imec "imec demonstrates die-to-wafer hybrid bonding with Cu interconnect pad pitch of 2um", 29 May 2024 — imec-int.com
  4. SEMI "3D Heterogeneous Integration Drives Demand for New Materials and Integration Solutions", 2 October 2018 (interview with Gerald Beyer of imec) — semi.org
  5. TSMC "TSMC Celebrates 30th North America Technology Symposium", 24 April 2024 (includes the COUPE description) — pr.tsmc.com
  6. TSMC Research "Heterogeneous Integration of a Compact Universal Photonic Engine for Silicon Photonics Applications in HPC", 2021 — research.tsmc.com
  7. TSMC CoWoS technology page — 3dfabric.tsmc.com

11. Claim-to-source audit

Claim in the textBasisLabel
The definition of heterogeneous integration: integrating separately manufactured components into a higher-level assembly that in the aggregate provides enhanced functionality and improved operating characteristicsIEEE EPS / HIR[Source 1] https://eps.ieee.org/technology/heterogeneous-integration-roadmap/Sourced
The HIR is supported by IEEE EPS, SEMI, IEEE EDS, the IEEE Photonics Society and ASME EPPD; it sets out packaging, test and interconnect requirements over 15 years (25 for emerging areas); its mission is to identify challenges before they become barriersIEEE EPS / HIR[Source 1] https://eps.ieee.org/technology/heterogeneous-integration-roadmap/Sourced
imec: heterogeneous integration is a scaling booster for functional partitioning and a way to create systems that would be impossible or economically unviable as a single chipSEMI interview, 2 October 2018[Source 4] https://www.semi.org/en/blogs/technology-trends/materials-and-processes-for-3d-heterogeneous-integrationSourced
imec: a 300 mm RF silicon interposer combining III-V (InP, GaAs, GaN) with Si-CMOS; MIMCAP giving 10 to 100 times the capacitance density using high-k aluminium hafnium oxide with a 3D oxide stud; modelling validated to about 300 GHz; laser-assisted bonding with alignment below 600 nm, rotational misalignment below 0.05 degrees across 43 devices, and reflection below -15 dB in the 110 to 170 GHz bandimec press release, 11 June 2026[Source 2] https://www.imec-int.com/en/press/imec-unlocks-system-level-iii-v-chiplet-integration-si-cmos-advancing-its-300mm-rf-siliconSourced
TSMC COUPE: SoIC-X used to stack an electrical die on a photonic die; lowest impedance at the die-to-die interface and better energy efficiency than conventional stacking; qualification for pluggables in 2025 and integration into CoWoS as CPO in 2026TSMC press release, 24 April 2024[Source 5] https://pr.tsmc.com/english/news/3136Sourced
COUPE uses a solid structure with no cavity or mechanically weak region, for both grating and edge couplers, achieving low insertion loss without contamination or mechanical concernsTSMC Research paper, 2021[Source 6] https://research.tsmc.com/page/on-chip-interconnect/14.htmlSourced
TBM, WLUF and photosensitive polymers for fine line and space are being pushed to their limits; die-to-wafer pitches well below 10 um were not yet reached on alignment and cleanliness grounds as of 2018; the wish to support design with a tool set rather than by handSEMI interview, 2 October 2018[Source 4] https://www.semi.org/en/blogs/technology-trends/materials-and-processes-for-3d-heterogeneous-integrationSourced
imec D2W hybrid bonding at 2 um pitch, overlay below 350 nm, using plasma dicingimec press release, 29 May 2024[Source 3] https://www.imec-int.com/en/press/imec-demonstrates-die-wafer-hybrid-bonding-cu-interconnect-pad-pitch-2mmSourced
The four degrees of "different" used to organise section 3Commentary by this article. Not a standardised industry classificationCommentary
The three mismatches (thermal expansion, process temperature, contamination control); the point that rotational accuracy drives high-frequency performance; the point that differing thermal behaviour complicates cooling designCommentary based on general relationships in materials engineering and process integration. No numerical claim is made about any specific productCommentary
The three values in Fig. 6 are not a like-for-like comparisonStated in the text and the figure note, because the sources differ in technology and measurement conditions. This article's own caveatCommentary

Last updated 20 September 2026. Sources are limited to primary material (official announcements and technical publications from standards bodies, research institutes and manufacturers). All figures are explanatory concept graphics. Visual-overview figures are shown as AI-generated conceptual images, matching the Japanese edition. Vector drawings are retained where they carry quantitative values or precision-critical technical labels. AI-generated images do not depict real equipment, products, facilities, dimensions or exact cross-sections; replaced source SVG overview drawings remain in the HTML but are hidden.

🌐 Japanese