Advanced Semiconductor Packaging — 2.5D/3D Integration and TSV
1. Overview
A. Definition
Advanced semiconductor packaging is the umbrella term for back-end integration technologies that use fine-wiring techniques such as silicon interposers, redistribution layers (RDL), through-silicon vias (TSV), and hybrid bonding to integrate dies of different processes and functions horizontally (2.5D) and vertically (3D), realizing at the package level the performance, bandwidth, and integration density that cannot fit on a single chip.
Traditionally, the value of a semiconductor came from front-end miniaturization — the race to make transistors smaller. However, as process nodes reached 3nm and 2nm and the performance gain per unit cost of miniaturization slowed sharply, the center of gravity for performance gains shifted to the packaging domain, to "how well can we connect and stack dies." Packaging of the past was closer to "packaging" that protected a finished chip from external shock and moisture and connected it electrically to the substrate, but advanced packaging has been elevated to a design and process technology that itself governs system performance. NVIDIA's AI accelerators H100 and B200 can fit HBM memory and the GPU die into a single package, and Apple's M series achieves high performance by connecting several dies — both thanks to advanced packaging.
B. Background and Necessity
Three pressures overlap behind advanced packaging's rise to the industry's core. The first is the Memory Wall. While compute performance has risen steeply, memory bandwidth has not kept up, making data supply the bottleneck in AI and HPC workloads. Attaching memory right beside the compute die (2.5D) or on top of it (3D) with ultra-high-density wiring dramatically shortens wire length, so bandwidth rises tens of times and transfer energy drops greatly. The second is the economic limit of miniaturization: as the chiplet strategy of combining function-divided dies in a package — instead of placing every block on the most advanced process — became mainstream, the packaging that connects these dies at high density became an essential foundation. The third is the reticle limit (about 858㎟): an AI chip demanding ultra-large compute exceeding the area a lithography tool can expose in one pass cannot be made as a single die in the first place, so package integration of multiple dies is the only solution. Thus advanced packaging has become not an option but the key to breaking the bottleneck of semiconductors in the AI era. Indeed, the fact that much of the supply constraint on AI accelerators stems not from compute dies but from packaging capacity such as CoWoS plainly shows that packaging has become both the bottleneck and a strategic resource for the whole industry.
C. Characteristics
The nature of advanced packaging can be summarized in three points. First, heterogeneous integration. By placing dies of different process nodes and devices (logic, memory, analog, photonic) in one package, it realizes combinations impossible on a single die. Second, bandwidth-centric design. Because the goal is to physically reduce wire length and minimize the distance data travels, the performance metric shifts from compute volume to bandwidth and transfer energy (pJ/bit). Third, the front-end-ization of back-end processes. TSV etching, interposer wiring, and hybrid bonding demand front-end-level precision, so packaging has been reshaped from low-value assembly into a high-value technology domain.
2. Integration Structure and Core Components
Advanced packaging is divided hierarchically by the direction and wiring density with which dies are connected. The key is "how short and dense the wiring between die and die can be made"; the shorter this distance, the higher the bandwidth and the lower the transfer energy. The structural diagram below shows the evolution of integration from traditional packaging through 2.5D to 3D.
flowchart TB
subgraph Trad["Traditional packaging (wire bonding / FC-BGA)"]
T1["Single die"]
T2["Substrate"]
T1 --- T2
end
subgraph D25["2.5D integration (horizontal)"]
L1["Logic die"]
H1["HBM stack"]
IP["Silicon interposer + TSV"]
L1 --- IP
H1 --- IP
end
subgraph D3["3D integration (vertical)"]
U1["Top die (cache/memory)"]
U2["Bottom die (logic)"]
U1 -->|"TSV / hybrid bonding"| U2
end
Trad --> D25 --> D3
Let us examine the core components with a focus on principles. The through-silicon via (TSV) is a copper pillar that vertically penetrates the silicon die; it is the foundation of 3D integration, stacking dies top and bottom so electrical signals pass directly. Because wiring penetrates straight through the die rather than routing around the package edge, the signal path is minimized and bandwidth maximized. HBM (High Bandwidth Memory) delivers ultra-wide bandwidth through thousands of I/Os precisely because it vertically stacks DRAM dies with hundreds to thousands of TSVs.
The TSV process is divided into Via-First, Via-Middle, and Via-Last by "when the hole is drilled," and in current mass production the Via-Middle approach — etched right after transistor formation and before the wiring layers — is mainstream. Vertically etching deep, thin holes (high aspect ratio) and filling them with copper can induce micro-cracks in the die from stress and thermal-expansion differences, so the design rules are demanding, including a KOZ (Keep-Out Zone) that forbids device placement around the TSV. For this reason TSVs are treated as a resource that "is essential for performance but carries area and stress costs," and their number and placement must be carefully optimized.
To organize the relationship among these components: the TSV is the vertical signal path, the interposer and RDL are the horizontal wiring plane, and bumps and hybrid bonding are the joining means — each with a distinct role. In 2.5D the interposer and bumps are the protagonists, in 3D the TSV and hybrid bonding, and in fan-out the RDL. In other words, which elements are combined determines the identity of the packaging method, and this freedom of combination is why advanced packaging is called a "design technology."
The silicon interposer is an intermediate substrate on which several dies are placed and interconnected by fine wiring underneath. Whereas an organic substrate has wiring widths on the order of tens of ㎛, the silicon interposer is fabricated by semiconductor processes and can realize tens of thousands of sub-1㎛ ultra-fine wires, forming the core of 2.5D integration. The redistribution layer (RDL) is a fine metal wiring layer that repositions a die's I/O pads to desired locations; it is the foundation of fan-out packaging, which realizes high-density wiring with organic materials without an expensive silicon interposer. Finally, the micro-bump and C4 bump are the joining points that electrically connect die, interposer, and substrate, forming a hierarchical structure where the pitch is narrow and dense going up (die↔interposer) and wider going down (substrate↔board). Because how far this joint pitch can be narrowed determines the ceiling of integration density, the transition to hybrid bonding is ultimately a trend toward eliminating these bumps.
Here it is worth noting the meaning of the name "2.5D." In that the dies lie on one plane it is 2D, but in that the interposer connecting them includes a vertical signal path via TSVs it has 3D elements. That is, it is called 2.5D in the sense of being between full vertical stacking (3D) and planar placement (2D). This distinction goes beyond mere naming, revealing a continuous spectrum of integration difficulty in which heat, wiring density, and design difficulty rise steeply from 2D→2.5D→3D. Thus in practice, hybrid integration — using 2.5D for bandwidth-critical memory and 3D for cache even within a single product — is becoming common.
A. 2.5D Integration — Side by Side Horizontally
2.5D places the logic die and memory (HBM) die horizontally side by side on an interposer and connects them with the interposer's ultra-fine internal wiring. Because they are not stacked vertically, thermal management is relatively easy, and bandwidth-critical memory like HBM can be placed tight against the compute die. The representative technology is TSMC's CoWoS (Chip-on-Wafer-on-Substrate), with which most of NVIDIA's and AMD's data-center AI accelerators integrate GPU and HBM.
CoWoS's competitiveness lies in enlarging the interposer to hold more compute dies and HBM. TSMC is mass-producing a CoWoS of 5.5× reticle size with yields reported to exceed 98%, and has presented a roadmap targeting mass production in 2028 of a 14× reticle-size package (about 10 compute dies + about 20 HBM stacks). CoWoS capacity has nearly doubled each year over the past three years yet still cannot fully meet AI demand, and memory and ABF substrate supply are cited as further bottlenecks.
A variant that lowers cost by embedding a localized silicon bridge (e.g., Intel EMIB, Embedded Multi-die Interconnect Bridge) into an organic substrate instead of a silicon interposer is also widely used. A full silicon interposer grows in area as dies increase, raising fabrication cost and yield burden, whereas EMIB inserts a small silicon bridge locally only at the junction where die meets die, solving only where high-density wiring is needed and leaving the rest to a cheaper organic substrate. This is a difference in design philosophy — "lay ultra-high-density wiring everywhere, or only where needed" — forming an important fork in 2.5D integration in terms of cost and scalability.
B. 3D Integration — Stacking Vertically
3D stacks dies vertically with TSVs, obtaining one-dimension-shorter wiring and higher integration density than 2.5D. Placing cache or memory on top of the compute die lets data move straight up and down without crossing a plane, so bandwidth, latency, and energy efficiency improve dramatically. If in planar (2D) wiring the distance between two blocks reaches several mm, in vertical stacking it shrinks to tens of ㎛, so signal delay and drive power can decrease by factors of tens to hundreds. This is the fundamental motivation for placing memory "on top" rather than "beside." AMD's 3D V-Cache, which stacks L3 cache on top of the CPU die to boost gaming and cache-intensive workload performance, is a representative commercial case. Early generations placed the cache die above the compute cores, making it hard to extract heat from the cores and forcing sacrifices in operating frequency, but later generations redesigned the stack structure itself, moving the cache below the cores to improve the thermal path. As this shows, stacking worsens the thermal density problem of trapped heat, so which die to place on top and how to extract heat become the core design challenges, and a single stacking order determines the performance ceiling.
C. Hybrid Bonding — Eliminating Bumps
Traditional stacking connects dies with small solder balls called micro-bumps, but there is a physical limit to narrowing the bump spacing (pitch). Hybrid bonding is a next-generation technology that directly bonds copper-to-copper (Cu-Cu) without bumps, narrowing the pitch to a few ㎛ or less. Wiring density becomes up to 10× higher than bumps, dramatically improving bandwidth and power efficiency. TSMC provides SoIC (System on Integrated Chips) based on hybrid bonding, presenting a roadmap from about 6㎛ pitch today to 4.5㎛ pitch by 2029, and Apple's M5 Pro and M5 Max are reported to have adopted SoIC-based packaging. Intel's Foveros Direct and Samsung's SAINT aim in the same direction.
Looking a bit deeper into hybrid bonding's difficulty, eliminating bumps means flattening the two dies' copper pads and surrounding oxide to the atomic level, pressing them together, then diffusing and bonding the copper by annealing (heat treatment). Even particles or roughness of a few nanometers on the surface cause bonding defects (voids), so ultra-clean, ultra-flat processes and nanometer-level alignment precision are essential. This very process difficulty is why hybrid bonding is grouped as "next-generation," and why securing mass-production yield becomes the battleground of industry competition.
D. Fan-Out Packaging — A Low-Cost, High-Density Compromise
Not every application needs an expensive silicon interposer. Fan-out packaging repositions dies on a reconstituted wafer or panel and then spreads the I/O widely outward (fan-out) with RDL alone to secure more inputs and outputs. The silicon interposer and substrate can be omitted, so thickness decreases, cost falls, and electrical characteristics improve. A representative case is TSMC's InFO (Integrated Fan-Out), applied to Apple's A-series APs to boost mobile form factor and power efficiency. Fan-out is not as ultra-high-density as 2.5D, but it fills the middle ground of the advanced-packaging spectrum by providing a cost-performance balance point for high-volume applications "where bandwidth is not urgent enough to warrant an interposer."
3. Comparison of Major Technology Types
The table below compares integration methods and representative technologies. However, the table is merely an aid to organization; the essence of the choice lies in the triangular trade-off of bandwidth demand, heat, and cost. AI accelerators where HBM bandwidth is key use 2.5D CoWoS, CPUs where cache capacity is key use 3D stacking, and cost-sensitive mobile and network chips use fan-out (InFO) — thus workload characteristics determine the packaging choice.
| Category | Integration direction | Wiring means | Wiring density | Representative tech | Main use |
|---|---|---|---|---|---|
| Traditional FC-BGA | Single plane | Solder bump | Low | Flip chip | General SoC |
| Fan-Out | Planar expansion | RDL | Medium | TSMC InFO | Mobile AP |
| 2.5D | Horizontal | Silicon interposer/bridge | High | CoWoS, EMIB | AI accelerator + HBM |
| 3D | Vertical | TSV + micro-bump | Very high | 3D V-Cache | Cache/memory stacking |
| 3D hybrid bonding | Vertical | Cu-Cu direct bonding | Highest | SoIC, Foveros Direct | Next-gen logic stacking |
Recasting the table's five types onto one axis, it is ultimately a continuous spectrum between wiring density (bandwidth) and cost/process difficulty. Toward the right (hybrid bonding) performance rises, but yield becomes harder to secure and investment grows. Thus there is no "best packaging"; it becomes a problem of choosing the most suitable point in the combination of bandwidth, power, form factor, and cost a product requires. This view matters because excessive integration can instead raise total cost of ownership through yield loss and heat.
Looking at the difference between fan-out and 2.5D down to the reasons: fan-out uses no expensive silicon interposer and repositions I/O with RDL alone to lower cost, but wiring density is limited, while 2.5D gains ultra-high density via a silicon interposer but is expensive because the interposer itself is made by semiconductor processes. That is, the question "is bandwidth truly necessary, or is cost more important?" separates the two. The difference between 3D and hybrid bonding is likewise: micro-bump stacking is mature and stable but limited in pitch reduction, while hybrid bonding narrows the pitch to a few ㎛ or less to maximize bandwidth but demands atomic-level surface flatness, cleanliness, and alignment precision, so process difficulty and investment are far higher. One investment estimate holds that SoIC's capital cost per unit capacity is markedly higher than CoWoS's.
The point where this comparison takes on practical meaning is consistency with the product roadmap. For example, a budget mobile AP using hybrid bonding is over-investment, while conversely a top-tier AI accelerator staying with fan-out is a choice that invites bandwidth shortage. Thus the packaging choice should be treated not as an engineering judgment for a single chip but as a strategic decision that also considers the product's market position, volume, and life cycle.
A. Next-Generation Substrates — Glass Substrates and Panel-Level Packaging
As much as wiring density, the limits of the substrate itself are becoming a bottleneck. Current 2.5D and large packages use organic ABF substrates, but as the package grows the substrate warps (warpage) and fine wiring becomes hard to lay stably. To overcome this, the glass substrate has emerged as a strong next-generation alternative. Glass has excellent flatness and dimensional stability and easily matches its thermal expansion to silicon, enabling larger packages with finer wiring. Also, panel-level packaging (PLP), which processes in square panel units instead of round wafers, is an attempt to lower cost per unit area by processing a larger area at once, reducing the edge waste of wafer-level (WLP). These still have challenges in mass-production maturity, but coupled with the trend of package enlargement, they are cited as the next battlefield of advanced packaging.
As substrates, processes, and integration methods evolve simultaneously in this way, advanced packaging must be understood not as a single technology but as a composite roadmap in which several technology tracks advance in concert. Because the structure is one in which packaging fills the gap left as miniaturization (front-end) slows, the picture in which the innovation pace of front-end and back-end crosses and complements to drive the overall performance curve is likely to persist for a while.
4. Application Cases and the Packaging Process Flow
Looking at the actual packaging of an AI accelerator reveals why advanced packaging is called "the front-end-ization of the back-end." The diagram below simplifies the assembly flow of a 2.5D CoWoS-based AI accelerator.
flowchart LR
A["GPU logic die<br/>(most advanced process)"] --> C["Bond to silicon interposer<br/>(Chip-on-Wafer)"]
B["HBM stack<br/>(TSV vertical stacking)"] --> C
C --> D["Mount interposer on substrate<br/>(on-Substrate)"]
D --> E["Molding / underfill / heat-dissipation structure"]
E --> F["Completed package (AI accelerator)"]
A notable point in the process flow is that every step accumulates yield multiplicatively. The logic die, HBM stack, and interposer must each be good, and losses add at every bonding, mounting, and molding step, so the final package yield falls close to the product of the individual process yields. Because a defect at the final step after gathering expensive dies causes enormous loss, densely placing intermediate inspection and KGD screening at each step is the core of cost management.
As a concrete case, NVIDIA's H100 integrates several HBM stacks around a single GPU die in 2.5D on a CoWoS interposer, securing terabyte-per-second (TB/s)-class memory bandwidth. The HBM stack itself is a 3D structure that vertically stacks 8–12 tiers of DRAM dies with TSVs, so within a single AI accelerator 2.5D and 3D coexist, overlapping. In another case, AMD's server CPU uses 3D V-Cache to stack an L3 cache die on top of the CPU die by hybrid bonding, boosting cache-intensive workload performance at the same power. In mobile, TSMC's InFO (Integrated Fan-Out) is used in Apple's A-series APs to capture both power efficiency and form factor with a thin, inexpensive fan-out package. Recently, Apple's M5 Pro and M5 Max adopted SoIC-based 2.5D packaging, showing the trend of hybrid-bonding integration spreading even to high-performance consumer products.
The common thread running through these three cases is the idea of "bringing data closer" rather than "making compute faster." The H100 improves bandwidth by attaching memory beside compute, V-Cache improves latency by stacking cache above compute, and InFO improves integration density by spreading I/O widely. All three are results not of making transistors smaller but of redesigning the placement and connection of dies, showing that advanced packaging has become a new axis of performance improvement. Numerically, too, 2.5D HBM integration raises bandwidth tens of times over traditional board-level memory while greatly lowering per-bit transfer energy (pJ/bit), directly contributing to the energy efficiency of AI training and inference.
This transition also ripples into design methodology. Whereas in the past packaging was considered after chip design finished, now System Technology Co-Optimization (STCO) — solving die partitioning, interfaces, heat, power, and package structure together from the start of design — is becoming the standard. In other words, the decisive power over performance has passed from a single transistor to the integrated design of the whole system.
The decisive difference from traditional packaging in this flow is that processes such as interposer fabrication, TSV etching, and ultra-fine bonding demand semiconductor front-end-level precision. Thus foundries (TSMC, Samsung) and OSATs (outsourced packaging firms) came to compete directly in the advanced-packaging domain, and packaging was reshaped from a low-value "back end" into a core axis of foundry competitiveness.
Thermal management is likewise being integrated as part of package design. As dies are stacked vertically and package area grows, heat per unit area surges, hitting the limits of conventional heat spreaders alone. Accordingly, direct liquid cooling that takes heat right above the die, microfluidic cooling that carves fine channels inside the package, and improvements in highly thermally conductive TIM (Thermal Interface Material) have been incorporated as essential elements of the packaging roadmap. That is, the question "how to integrate" can no longer be separated from the question "how to extract heat," and co-design that solves power, heat, and wiring together decides the success of advanced packaging.
5. Deep Dive — Recent Trends and the Industrial Competitive Landscape
Advanced packaging is currently the hottest investment area in the semiconductor industry. As AI accelerator demand explodes, how well dies are integrated — rather than the miniaturization of compute dies — has become the supply bottleneck and the measure of competitiveness, and foundries are pouring investment into packaging no less than into the front end. The direction can be summarized in three points.
First, the enlargement of packages. TSMC is mass-producing 5.5× reticle CoWoS and intends to enlarge package area to 14× in 2028 and beyond in 2029, to hold more compute dies and HBM in one package. This is a strategy to push a single accelerator's memory capacity and bandwidth to the extreme in step with the explosive growth of AI models. However, as the package grows, substrate warpage, uneven heat distribution, and yield decline worsen, so it is sustainable only in parallel with next-generation foundation technologies such as glass substrates and PLP.
Second, the transition to hybrid bonding. As micro-bumps reach their pitch limit, TSMC SoIC, Intel Foveros Direct, and Samsung SAINT are all moving to Cu-Cu direct bonding. TSMC SoIC targets 4.5㎛ by 2029 from about 6㎛ pitch today, and Intel Foveros Direct aims for a sub-10㎛ pitch for wiring density up to 10× that of micro-bumps. However, these figures are estimates based on vendor and industry announcements, and the actual range of application may vary with cost, yield, and heat, so definitive claims are difficult. The important implication is that as pitch narrows, die-to-die I/O density rises so that memory bandwidth and power efficiency improve together, and this directly affects the way the base die and logic are combined in next-generation HBM (HBM4 and later).
Third, ecosystem bottlenecks and supply-chain realignment. In a situation where CoWoS demand is not fully met even as capacity doubles each year, HBM memory and ABF substrates are cited as further bottlenecks. Accordingly, investment spreads to OSATs and materials and equipment firms, and domestically (in Korea) too, Samsung Electronics and SK hynix are competing with HBM stacking, hybrid bonding, and packaging-integration capability as strategic tasks.
It is also worth gauging the likely exam direction. Short-answer forms are likely such as "the difference between 2.5D and 3D packaging," "the principle and role of TSV," and "comparison of CoWoS, SoIC, Foveros," while essay forms are likely to connect the background problem (memory wall, miniaturization limits) to the technical solution, such as "discuss the role and challenges of advanced packaging in improving AI accelerator performance" or "the memory wall problem and packaging-based solutions." Since related past and similar topics — [[chiplet]], [[ucie-chiplet-interconnect]], [[high-bandwidth-memory]], and [[backside-power-delivery]] — tend to be examined together, understanding advanced packaging as a higher-level concept running through these is advantageous for constructing an answer.
In sum, competition in advanced packaging converges on a triple contest of "who integrates bigger, denser, and more reliably." While TSMC holds the lead with CoWoS and SoIC, Intel (Foveros, EMIB) and Samsung (SAINT, I-Cube) pursue, and the memory firms holding HBM are also entering the packaging value chain through base dies and custom HBM. The outcome of this competition depends on comprehensive capability spanning standards, ecosystem, and supply chain beyond mere process technology.
6. Considerations and Implications
Advanced packaging is, beyond a mere back-end technology, a strategic variable that reshapes the value chain and competitive structure of the semiconductor industry. The key considerations from a professional engineer's perspective when adopting and setting strategy are as follows.
- Thermal and power-integrity (PI/TI) design: The more dies are stacked, the higher the thermal density and the harder power delivery becomes. Thermal design (TIM, microfluidic cooling) and the power delivery network (PDN) must be integrated (co-designed) from the start of package design, and heat-vulnerable dies must have their stacking position chosen carefully. In particular, as the path supplying power to the top die in a 3D stack lengthens and IR drop worsens, combination with backside power delivery technologies such as [[backside-power-delivery]] is necessarily examined.
- Design-process integration (Chiplet + Package Co-Design): A full-cycle design system that simultaneously optimizes die partitioning, interface standards (UCIe, HBM), interposer wiring, and heat/power is needed. A design organization that looked only at the front end must expand into capability spanning packaging.
- Cost-yield trade-off: Hybrid bonding and large interposers maximize bandwidth but sharply raise process difficulty and investment. By comparing the bandwidth level a workload demands with the target cost, one must choose right-sizing rather than excessive integration.
- Supply-chain and KGD risk: Because packaging combines several known-good dies (KGD), a single defective one discards all the expensive dies. Per-die screening tests, multi-sourcing, and stable securing of bottleneck resources such as HBM and ABF substrates are strategically important.
- Organizational and capability realignment: The traditional organizational and supply-chain structure that separated front-end (design, device) from back-end (packaging) cannot meet the demands of co-design. Having an integrated engineering system spanning design, front-end, packaging, and test, and a close collaboration model among foundry, OSAT, and materials firms, is the premise of mid- to long-term competitiveness.
- Standardization and ecosystem outlook: The maturation of D2D standards like UCIe and of hybrid-bonding processes is the key to enabling heterogeneous-vendor chiplet integration. In the long run packaging is a core axis of foundry competitiveness, and whether advanced-packaging capability is secured is likely to determine leadership in AI semiconductors.
- Test and reliability-verification system: Because internal defects are hard to access and repair after dies are integrated, a DFT (Design for Test) strategy spanning both pre-bond testing of individual dies and post-bond verification of interconnects is needed. Because a single micro-defect among thousands of TSVs and bumps can degrade the whole package's reliability, reliability verification such as burn-in and thermal cycling must be redefined at the package level.
References
- TrendForce, "TSMC's 5.5-reticle CoWoS reportedly tops 99% yield" — https://www.trendforce.com/news/2026/08/11/news-tsmcs-5-5-reticle-cowos-reportedly-tops-99-yield-flags-memory-abf-as-ai-bottlenecks/
- TrendForce, "TSMC speeds up AP7/AP8 build-outs, targets doubling SoIC capacity" — https://www.trendforce.com/news/2025/03/26/news-tsmc-reportedly-speeds-up-ap7-and-ap8-build-outs-targets-doubling-soic-capacity/
- EENewsEurope, "TSMC SoIC roadmap 2029 chip stacking" — https://www.eenewseurope.com/en/tsmc-soic-roadmap-2029-chip-stacking
- Intel, "Foveros Direct 3D Tech Brief" — https://www.intel.com/content/dam/www/central-libraries/us/en/documents/2025-11/foveros-direct-3d-tech-brief.pdf
In one line: Advanced semiconductor packaging integrates dies in 2.5D and 3D with TSVs, interposers, and hybrid bonding to surpass the memory wall and miniaturization limits; represented by CoWoS, SoIC, and Foveros, it has risen as a core axis of foundry competitiveness in the AI era.