AI Chips Approaching the Reticle Limit
With the continuous evolution of AI chips, advanced packaging has become a well-worn topic. In a recent report, TrendForce assesses the competitive landscape for advanced packaging in the AI chip sector. The report notes that while alternatives like Intel’s EMIB-T are vying for AI chip clients, TSMC’s CoWoS-L—leveraging its technological maturity and superior yield rates—is expected to remain the dominant advanced packaging solution for AI chips through 2028. This outlook suggests that competition in AI chip packaging over the coming years will center on CoWoS-L; for other solutions to gain a foothold in this market, they must demonstrate more compelling performance in terms of both maturity and yield.

Why Are AI Chip Packages Getting So Much "Larger"?
AI chips are growing increasingly large—but this "largeness" primarily refers to the package, rather than the individual compute die itself. As the parameter scale of large models and the complexity of inference tasks continue to rise, AI accelerators demand greater computational power and higher memory bandwidth. Since a single die can no longer meet these requirements on its own, technologies like Chiplets and High-Bandwidth Memory (HBM) have become essential for boosting system performance. The growing need to interconnect compute and storage resources within the chip places significant pressure directly on the packaging stage.
NVIDIA’s Rubin platform serves as a prime example. According to public information, the Rubin GPU incorporates 336 billion transistors and utilizes HBM4, offering up to 288GB of memory and 22TB/s of memory bandwidth per GPU. AMD is also accelerating its development cycle; the Instinct MI400 series employs a Chiplet architecture, with the next-generation MI455X consisting of eight compute dies—designed for the Helios rack-scale system, which comprises 72 GPUs. A common trait among these products is the increasing number of compute dies and HBM stacks integrated into a single package. In terms of physical dimensions, NVIDIA’s Blackwell GPU has reached approximately 3.3 times the "reticle limit" (the maximum size for a single lithographic exposure), and the next-generation Rubin platform is expected to reach 4 times that limit.
As the number of compute dies and HBM stacks grows, the traditional approach—pairing a single compute die with external memory—can no longer satisfy the bandwidth and data-transfer efficiency requirements of modern AI systems. By integrating multiple compute dies and HBM stacks into a single advanced package, engineers can shorten the distance data must travel between components and leverage high-density interconnects to achieve greater bandwidth. Consequently, the "largeness" of AI chips manifests primarily as an expansion in package size. What is truly growing—continuously and significantly—is the total scale of components integrated within an advanced package (including compute dies, Chiplets, HBM, and high-speed interconnects), rather than the lateral surface area of any single logic die.
Data indicates that this trend toward larger package sizes has been underway for nearly a decade. According to a report on advanced packaging technology trends by IDTechEx, the size of the silicon interposer used in TSMC’s CoWoS (Chip-on-Wafer-on-Substrate) technology has evolved from approximately 1.5 times the reticle size (roughly 1,287 mm²) in 2016 to 3.3 times the reticle size (roughly 2,831 mm²) today. Plans are in place to expand this to 5.5 times the reticle size—approximately 4,719 mm²—by 2026. This trend toward larger package sizes is also reflected in manufacturing equipment; for instance, the newly launched FC Bonder 75 from HANMI Semiconductor supports package dimensions of 75 mm x 75 mm, a significant increase over the 20 mm x 20 mm standard for conventional, general-purpose packages.
The push toward larger package sizes extends beyond GPUs to include AI ASICs. Data from TrendForce indicates that shipments of high-end NVIDIA GPUs are projected to grow by approximately 30% year-on-year in 2026, while AMD is set to aggressively promote its MI400 and MI450 series in the second half of 2026. Among cloud service providers (CSPs) developing their own ASICs, Google is expected to show the strongest shipment volume and momentum in 2026, driven primarily by its TPU v7. AWS is slated to begin the gradual rollout of its Trainium v3 chips in the second half of the year, whereas Microsoft and Meta continue to rely predominantly on GPU-based servers, with more modest plans for proprietary ASICs. Across the board, various types of AI chips are moving toward higher levels of integration, necessitating that packaging solutions accommodate an ever-increasing density of computing and memory resources.
Why have package dimensions hit a bottleneck?
As package dimensions continue to expand, new problems arise. For instance, is it even feasible to manufacture an interposer of such a large size?
In traditional 2D packaging, chips are connected to a package substrate via wire bonds, with metal traces and vias within the substrate facilitating signal transmission. This path is quite long, introducing significant resistance and inductance, which leads to bottlenecks in signal latency and power consumption. When multiple chips are packaged separately and placed on a PCB, the physical distance signals must travel between them is substantial; driving signals across this distance significantly increases power consumption and limits the overall communication bandwidth between the chips.
Surface area is another issue. Placing multiple large packages on a PCB consumes a vast amount of physical space—a major drawback for products with strict spatial constraints. More critically, there is a significant disparity between the trace widths on a PCB and the dimensions of the bumps or micro-bumps on a chip. This gap limits the number and density of I/O (input/output) connections, making traditional packaging methods ill-suited to the demands of AI and high-performance computing (HPC), which require high bandwidth, low latency, and miniaturization. With limitations across interconnect length, physical footprint, and I/O density, 2D packaging is increasingly unable to meet the integration requirements of AI chips.
To address these challenges, some manufacturers have introduced 2.5D packaging. The core concept involves incorporating an interposer—a layer upon which multiple chips are placed side-by-side, rather than mounting them directly onto a PCB. The chips connect to the interposer via micro-bumps, and the interposer contains an ultra-fine redistribution layer (RDL) that enables high-density, short-range horizontal communication between the chips.
The interposer is then connected vertically to the underlying package substrate through through-silicon vias (TSVs). This architecture drastically shortens the interconnect distance between chips, improving both signal integrity and bandwidth. Furthermore, it bypasses the reticle (mask) size limit—a constraint that restricts the dimensions of a single monolithic chip—and provides a viable pathway for integrating additional compute chips and high-bandwidth memory (HBM).
It is within this 2.5D packaging paradigm that TSMC’s CoWoS (Chip-on-Wafer-on-Substrate) technology has emerged as a leading, representative solution.
CoWoS-L: A Mainstream 2.5D Solution
TSMC’s CoWoS is a quintessential 2.5D packaging solution, centered on integrating multiple SoCs or chiplets alongside HBM within a single package.
The current CoWoS family consists of three primary variants: CoWoS-S, CoWoS-R, and CoWoS-L. The key distinction among them lies in their respective approaches to the interposer. CoWoS-S utilizes a large-area silicon interposer, which employs routing to establish high-density interconnections, thereby integrating logic chips with HBM. CoWoS-R, by contrast, replaces the silicon interposer with an RDL (Redistribution Layer). Composed of polymer and copper routing layers, the RDL offers superior mechanical flexibility while maintaining high interconnection performance; it is also more easily scalable to accommodate larger package dimensions.
CoWoS-L builds upon the foundations of the previous two, offering a more flexible approach. It utilizes RDL for routing while incorporating small silicon segments known as "Local Silicon Interconnects" (LSI)—the "L" in CoWoS-L. According to TSMC, these LSIs leverage multi-layer, sub-micron copper traces to facilitate high-density die-to-die interconnections. Depending on specific product requirements, they can support various configurations, such as SoC-to-SoC, SoC-to-chiplet, or SoC-to-HBM connections. Meanwhile, the RDL handles signal and power transmission across a broader area. This design ensures that high-density interconnections are concentrated only where fine-pitch connectivity is strictly required, rather than imposing the same high-density routing requirements across the entire interposer.
This structural evolution is the key to CoWoS-L’s ability to support ever-larger package sizes. For ultra-large-scale AI chips, not every region within the package demands the same level of routing density. If a silicon interposer with uniform high-density routing were used throughout, both manufacturing costs and process complexity would escalate significantly as the total surface area increases. By combining LSI (Local Silicon Interconnect) with RDL (Redistribution Layer), CoWoS-L applies high-density interconnects precisely where they are most needed, while leveraging RDL to handle broader-scale connectivity and power delivery. This approach allows for the expansion of the overall package size without compromising interconnect capabilities in critical areas. TSMC has positioned CoWoS-L as a packaging solution for large-scale HPC (High-Performance Computing) products, continuously increasing the size of its interposer to integrate a greater number of silicon dies and HBM (High Bandwidth Memory) stacks.
This strategy—combining localized high-density interconnects with large-area expansion—aligns well with the chiplet architectures used in AI chips. Future AI accelerators may incorporate multiple compute dies, specialized functional chiplets, and several HBM stacks, with varying data flow requirements between them. Some links demand ultra-high bandwidth and minimal connection distance, while others prioritize signal distribution and power delivery across different regions. CoWoS-L can be configured to meet these diverse interconnect needs; it can even integrate embedded deep-trench capacitors (eDTC) beneath the SoC to enhance power management. In essence, packaging is evolving from a simple means of connecting chips into a comprehensive design framework for interconnecting and powering an entire computing system.
In terms of practical scaling, CoWoS-L has moved well beyond the laboratory. TSMC’s first CoWoS-L solution, featuring a 3.5x reticle size, entered mass production in 2024. According to TSMC’s 2025 disclosures, the company has qualified a CoWoS solution with a 5.5x mask/reticle size and plans to begin mass production in 2026 to meet the growing demand for higher performance in AI and HPC applications. TSMC has also indicated that it will continue to expand the size of its CoWoS interposers to accommodate more advanced-node chips and HBM.
This technological roadmap is already being implemented in commercial products. TrendForce notes that AI accelerators from NVIDIA and AMD, as well as Meta’s MTIA 400, utilize the CoWoS-L packaging solution. As cloud service providers continue to develop their own custom ASICs—a process that necessitates the integration of multiple SoCs, chiplets, and HBM within a single package—both Microsoft and AWS are expected to adopt CoWoS-L by 2027. TrendForce anticipates that, thanks to its technical maturity and superior yield rates, CoWoS-L will remain the mainstream advanced packaging solution for AI chips through 2028.
However, CoWoS-L is not the only approach for implementing 2.5D packaging.
EMIB: An Alternative Technical Path
While CoWoS-L focuses on the use of an interposer, Intel’s EMIB (Embedded Multi-die Interconnect Bridge) represents a different 2.5D packaging approach. Instead of employing a large silicon interposer that spans the entire package area, EMIB utilizes smaller silicon bridges embedded directly into the package substrate. These bridges establish high-density interconnections only where they are needed—specifically between chips that require high-density, high-speed data exchange. The chips are placed side-by-side on the substrate, with the silicon bridges facilitating high-speed, high-density data transmission between them.
Compared to a large-area silicon interposer, the EMIB concept is more localized. Silicon bridges are placed only in areas where high-speed communication between chips is required, eliminating the need for a full-coverage silicon interposer across the entire package. According to Intel, the size and type of these bridges can be customized for specific die-to-die connections—whether between logic chips or between a logic chip and HBM—without routing all signals and power through a single, monolithic silicon interposer. This approach not only enhances interconnection density between chips but also allows for a better balance between packaging architecture and cost.
This structure is particularly well-suited for Chiplet architectures. Multiple chiplets—potentially manufactured using different process nodes and serving different functions—can be integrated into a single package and linked via strategically placed EMIB bridges, enabling components like CPUs, GPUs, I/O, and HBM to function as a cohesive system. Intel has already deployed EMIB for large-scale mass production, reporting assembly yields comparable to those of traditional, similarly complex FCBGA (Flip-Chip Ball Grid Array) packages.
However, EMIB does have its limitations. Because the silicon bridges are embedded within the package substrate rather than spanning the entire package as a large-scale interposer, the high-density connectivity between chips remains constrained by the routing capabilities of the substrate itself. Furthermore, as HBM bandwidth continues to increase and the volume of data transmission within AI chips grows, power delivery—in addition to interconnection density—has emerged as a critical challenge. For high-power AI chips, the challenge of delivering sufficient current rapidly and stably within a limited package footprint—while simultaneously minimizing power noise and transmission losses—has become a critical factor for the further evolution of EMIB (Embedded Multi-die Interconnect Bridge).
EMIB-T emerged as an upgraded solution to address this challenge. The "T" stands for "Through-Silicon Via." By incorporating TSVs into the existing local silicon bridge structure, Intel enables power to be delivered vertically from the package substrate directly to the chip above, bypassing the need to route through other interconnect layers. This shortens the power delivery path, increases power delivery density, and enhances power integrity for high-power, high-bandwidth applications. According to Intel’s public documentation, EMIB-T can also integrate multi-layer copper power and ground networks, as well as high-density capacitors, directly within the bridge—further reducing the impedance and noise of the power delivery network.
If EMIB is envisioned as a bridge connecting chips, the primary innovation in EMIB-T is the addition of a vertical channel to that bridge. While the original EMIB focused on facilitating communication between chips, the integration of TSVs allows the package to also handle the direct vertical delivery of power. In the EMIB-T solution showcased by Intel at ECTC 2026, the first-level interconnect bump pitch has been reduced to 25 microns, and the package size has expanded to 120mm × 120mm. It supports the integration of compute and memory chips spanning more than nine reticle areas, while also accommodating high-speed interconnects such as 12 Gb/s HBM4e and 64 Gb/s UCIe.
This signifies that EMIB is evolving from a simple local silicon bridge toward larger dimensions, higher interconnect density, and more robust power delivery capabilities. TrendForce notes that, given ongoing concerns regarding the costs of CoWoS-L packaging and the capacity constraints of TSMC’s advanced packaging, EMIB has emerged as an alternative for some AI chip manufacturers. Google is expected to adopt EMIB-T by 2027, and AWS is also currently testing EMIB technology. Although EMIB-T competes directly with CoWoS-L, the latter’s advantages in terms of technical maturity and yield are expected to keep it the dominant packaging technology for AI chips through 2028. For Intel, EMIB-T—beyond simply increasing the available packaging area—enables local silicon bridges to handle both high-speed signal transmission and vertical power delivery, offering a new packaging pathway for the integration of larger-scale AI chips.
From CoWoS-L to EMIB-T: The Future Path of Advanced Packaging
Comparing CoWoS-L and EMIB-T reveals that the essence of their competition extends well beyond packaging technology itself. In the current era of AI chips, relying solely on transistor miniaturization to drive performance gains is no longer sufficient. As an increasing number of compute dies, chiplets, and HBM (High Bandwidth Memory) modules are integrated into a single package, the packaging process has effectively become an integral part of chip design. CoWoS-L pursues this by expanding the interposer, utilizing RDL (Redistribution Layers) and local silicon interconnects to link various components. EMIB-T, meanwhile, continues to advance through the use of embedded silicon bridges. While their technical approaches differ, their objectives are closely aligned: to enable high-speed collaboration among a greater volume of compute and memory resources within a single package. Consequently, TrendForce anticipates that competition in advanced packaging for AI chips will continue to center on achieving larger package sizes and higher interconnect densities in the coming years.
However, as packages grow larger, 2.5D packaging technology itself is encountering an increasing number of challenges.
Consider interconnects: The volume of data exchange within AI chips has reached such a magnitude that simply adding more compute units is insufficient; the processed data must also be transmitted to the next die or HBM module as rapidly as possible. Factors such as physical distance, trace width, and the number of interconnects all influence both transmission speed and power consumption. As a result, the layout within the package—specifically, the proximity of various chips to one another—has become a critical determinant of overall chip performance. It is no longer accurate to view packaging merely as the final step following the chip fabrication process. In the context of chiplet architectures, the arrangement of dies and the implementation of interconnects must be considered holistically. The industry is increasingly prioritizing "package-level co-design," an approach that integrates electrical, thermal, and mechanical considerations from the very beginning of the chip, package, and system design phases.
Thermal management and yield rates also represent unavoidable hurdles for the continued evolution of 2.5D packaging. Modern AI accelerators are high-power devices; the more chips integrated into a single package, the more concentrated the localized heat becomes. While 2.5D packaging allows for the lateral distribution of GPUs, ASICs, and HBM—thereby providing some space for heat dissipation—this lateral area is ultimately finite. Regarding yield, the value of an individual AI chip is already substantial. When multiple compute dies, HBM (High Bandwidth Memory), and a complex array of interconnects are integrated into a single package, a failure at any point can compromise the entire assembly. As the number of components increases, the processes for manufacturing, bonding, testing, and failure analysis require far greater precision. This is especially true for ultra-large-scale packages, where the focus shifts from merely achieving functionality to ensuring consistent, stable production—all while keeping costs under control.
Given these challenges, why not simply abandon 2.5D packaging in favor of a full transition to 3D packaging? From a technical standpoint, 3D packaging certainly offers higher integration density per unit area. TSMC’s SoIC (System-on-Integrated-Chips) technology already supports 3D integration—whether logic-to-logic or memory-to-logic—with 3nm SoIC stacking entering mass production in 2025. However, for high-power AI logic chips, large-scale 3D packaging faces three major hurdles: thermal management, power delivery, and yield. Heat can become trapped within the stacked structure, and power must be routed through multiple layers—meaning a defect in any single layer could cause the entire package to fail. More critically, 3D packaging is not yet mature in terms of technology or cost; it requires multiple tape-outs and sets of photomasks, leading to development and production costs far higher than those of traditional SoCs. The complex manufacturing processes, combined with lower yields, further drive up the per-unit cost. In contrast, TSMC’s CoWoS (Chip-on-Wafer-on-Substrate) technology has been in mass production for a decade, maintaining a stable yield of over 90%, whereas improvements in 3D packaging yields remain contingent on achieving higher bonding precision. Currently, only a few industry-leading companies possess the capability to develop 3D packaging, and its application is largely confined to high-end markets that are less sensitive to price. Consequently, 2.5D packaging remains the more viable mainstream approach for large-scale AI chip integration at this stage.
Therefore, the next phase of competition in advanced packaging will not be defined solely by package size. While dimensions will continue to grow and interconnect density will keep increasing, these advancements must be balanced with solutions for signal integrity, power integrity, thermal management, yield, and cost—all of which must be addressed in tandem. Chip manufacturers can no longer afford to wait until the chip design phase is complete to decide on a packaging solution; instead, packaging capabilities must be factored in from the very beginning of the design process. Looking ahead, the evolution of HBM (High Bandwidth Memory) will place further demands on packaging. As HBM capacity and bandwidth continue to grow, individual AI packages must accommodate an ever-increasing amount of memory resources. Meanwhile, the rising number of compute dies leads to more frequent data exchange between chips. The interdependencies between packaging, HBM, and compute chips will become increasingly tight—to the point where chip designs might even be modified specifically to accommodate the requirements of the packaging.
Conclusion
CoWoS-L and EMIB-T are, in fact, merely two milestones in this current wave of evolution. They address the most pressing challenges arising from the rapid scaling of AI chips today. The true test for the industry supply chain moving forward lies in its ability to integrate an even greater number of chips—ensuring high-speed performance, effective heat dissipation, reliable power delivery, and stable, large-scale production. As AI chips continue to grow in "size," packaging technologies must inevitably evolve in tandem.
