CoWoS vs InFO: what is the difference?
CoWoS and InFO are both part of TSMC's 3DFabric family of back-end packaging technologies. They are often mentioned together because both handle advanced chips, but they target opposite ends of the market: CoWoS maximises interconnect density and memory bandwidth, while InFO minimises thickness, size and cost.
The short answer
The core difference is the interposer. CoWoS mounts dies and HBM stacks on an interposer with very fine wiring and then on a substrate, which gives the bandwidth needed for AI accelerators. InFO (integrated fan-out) embeds the die in mould compound and builds redistribution layers (RDL) directly on it, with no interposer, giving a thinner and cheaper package suited to mobile processors.
Side-by-side comparison
| Aspect | CoWoS | InFO |
|---|---|---|
| Package type | 2.5D: dies on an interposer, interposer on an organic substrate | Fan-out wafer-level package: dies embedded in mould with RDL built on top |
| Interposer | Silicon (CoWoS-S), RDL (CoWoS-R) or RDL with local silicon bridges (CoWoS-L) | None; the RDL itself fans signals out beyond the die edge |
| Die attach | Micro-bumps from each die to the interposer | RDL formed directly over the die pads, no bumps between die and RDL |
| HBM | Designed to place several HBM stacks beside the logic die | Main versions are not aimed at HBM; mobile variants stack LPDRAM above (InFO-PoP) |
| Typical products | Data-centre GPUs and AI accelerators | Smartphone application processors and other compact devices |
| Thickness | Thicker: interposer plus substrate | Thin; can omit the laminate substrate |
| Relative cost | Higher: interposer fabrication and more assembly steps | Lower: fewer layers and wafer-level processing |
| Main strength | Interconnect density and memory bandwidth | Small form factor, good electrical and thermal performance for its size |
How CoWoS works
In CoWoS the logic die and HBM stacks are bonded with micro-bumps onto an interposer while it is still in wafer form. The interposer wiring is made with wafer-fab tools, so it can carry far more lines between the processor and memory than any organic substrate. After thinning and dicing, each interposer is mounted on a package substrate.
This structure is what lets a single accelerator package hold several HBM stacks with very wide interfaces to the processor. Its cost and size are the price paid for that bandwidth.
How InFO works
InFO starts from tested dies that are placed face down on a carrier and embedded in mould compound, forming a reconstituted wafer. Redistribution layers are then built over the die surface. Because the mould extends beyond the die edge, the RDL can fan connections out to a wider area than the die itself, which is the meaning of fan-out.
Without an interposer or, in many cases, a laminate substrate, the package is thin and short in signal path. InFO-PoP adds through-mould vias so that a DRAM package can be stacked on top, a layout used for smartphone processors. Other InFO variants integrate multiple dies side by side for larger designs.
The CoWoS family and its link to InFO
CoWoS-S uses a monolithic silicon interposer. CoWoS-R replaces it with an RDL interposer, which borrows fan-out style wiring, and CoWoS-L combines an RDL interposer with small local silicon interconnect (LSI) bridges where the densest links are needed. CoWoS-R and CoWoS-L exist mainly to reach package sizes beyond what a single silicon interposer can provide.
The boundary between the families is therefore not absolute: RDL technology developed for fan-out packaging also appears inside newer CoWoS variants.
Which to use when
If a design needs HBM or several large dies with very wide die-to-die interfaces, CoWoS is the usual answer. If the priority is a thin, low-cost package for a single processor, especially in a phone, InFO fits better. Similar ideas from other companies include Intel EMIB, which embeds a small silicon bridge in the organic substrate instead of using a full interposer, and fan-out packages from OSATs.
FAQ
What does fan-out mean in InFO?
In a fan-in wafer-level package all connections must fit within the die area. In fan-out the die is surrounded by mould compound, so the redistribution layers can route connections outside the die edge, allowing more I/O than the die area alone would permit.
Why do AI accelerators use CoWoS?
AI accelerators need several HBM stacks connected to the processor through very wide interfaces. The fine wiring on a CoWoS interposer can route those thousands of lines over short distances, which an organic substrate or a standard fan-out package cannot match.
Is InFO a 2.5D technology?
Standard InFO is generally described as a fan-out wafer-level package rather than 2.5D, because it has no separate interposer. Multi-die InFO variants and CoWoS-R blur this line, since they place dies side by side on RDL wiring.
Is CoWoS always more expensive than InFO?
For a comparable product, yes in general, because it adds an interposer and more assembly steps. TSMC does not publish package prices, so exact cost ratios quoted online should be treated as estimates.
Terms compared
- CoWoS (Chip on Wafer on Substrate)TSMC's 2.5D packaging technology that places logic dies and HBM stacks side by side on an interposer; the standard package for AI accelerators.
- Advanced PackagingPackaging technologies that connect several dies densely in one package, such as 2.5D interposers, 3D stacking, fan-out and hybrid bonding.
- HBM (High Bandwidth Memory)A memory made by stacking DRAM dies with through-silicon vias, giving a very wide interface placed next to a GPU or other processor.
Japanese companies involved
More packaging comparisons
Sources
Semiconductor comparisons
- All semiconductor comparisons
- Chiplet vs monolithic: what is the difference?
- Hybrid bonding vs flip chip: what is the difference?
- TSV vs wire bond: what is the difference?
- Wafer test vs final test: what is the difference?
- HTOL vs burn-in: what is the difference?
- SEM vs TEM: what is the difference?
- XPS vs EDX: what is the difference?
Last updated: November 2, 2026