Author: ICVIEWS Editorial Team
In current industry perception, Chiplet is often categorized under the umbrella of advanced packaging. However, at the recently held HIPI Chiplet Forum, HIPI Vice Chairman Wang Zhimin offered a different perspective. He did not discuss Chiplet within the lineage of advanced packaging but rather as a form of tight-coupled division of labor. Furthermore, instead of measuring chip strength by nanometer nodes, he shifted the evaluation criteria to metrics that more directly reflect product performance, such as integration density, interconnect energy efficiency, and thermal dissipation capabilities. This shift in perspective reveals a complete chain extending from scenario requirements to ecosystem collaboration.
01 Artificial Intelligence Turns a New Page in the Compute Ledger
Every expansion cycle in the chip industry has been driven by a sufficiently large application. The driving force behind this round comes from large models. Wang Zhimin decomposed AI productivity into five interlocking elements: intelligence corresponds to the capability ceiling of large models; computing power corresponds to the supply capacity of AI chips; storage power corresponds to the capacity of high-bandwidth memory; network power corresponds to the connection efficiency of large-scale clusters; and ecosystem power corresponds to the industry's ability to deliver at scale. A bottleneck in any one of these five elements will suppress the actual output of the entire chain. This breakdown explains why capital expenditures are rising with such a steep slope. The global investment scale for AI infrastructure has already been pushed to the $800 billion level, and where the funds flow is precisely at the intersection of these five elements. Chips are no longer just computing units; they are simultaneously storage units, interconnect units, and power/thermal management units. How much capability a single chip can accommodate depends on whether these units can be efficiently assembled together. It is against this backdrop that Chiplet has transitioned from an alternative solution to the primary path.
Currently, there are over 50 fabless design companies in China, along with more than 500 industry-academia-research institutions in the ecosystem chain, which have cumulatively launched over 150 Chiplet-based chips. This number signifies that Chiplets are no longer stuck in proof-of-concept stages but have established a batch of reusable and combinable entities. Wang Zhimin’s medium-term judgment is that by 2030, approximately half of the world's chips will be realized in Chiplet form. If this ratio holds true, the real competition in the next five years will not lie in who can produce the single most powerful Chiplet, but in who can stably, cost-effectively, and at scale organize these Chiplets into systems.
There has long been an expectation in the industry to settle on a universal Chiplet architecture, using the same set of building blocks to construct all products. Wang Zhimin’s judgment contradicts this view: scenarios are diverging, and this divergence itself is the greatest certainty today.
AI training and inference are moving toward super-nodes, pursuing high bandwidth, low latency, and scalability in the sense of vertical scaling. The transistor count of a single super-chip needs to reach the scale of 200 billion. High-density storage is evolving toward hybrid layering of HBM and HBF, with new structures overlaying new integrations, shifting the competitive focus from expanding production capacity to establishing standards. Data center networks and 6G are moving toward integrated communication, sensing, and computing. Ultra-large antenna arrays require localized high computing power on the baseband side, and the commercial window for 6G is becoming clear around 2029. On the network switching side, port capacity is advancing from 51.2T to over 200T, with digital and analog nodes evolving separately, and CPO (Co-Packaged Optics) being included in roadmaps. Autonomous driving and embodied intelligent robots face highly dispersed vehicle model and form factor requirements while being tightly constrained by quality and reliability standards. Reusing and sharing verified Chiplets has become an inevitable choice. AI PCs prioritize cost, power consumption, and local application experience above all else. With shipment volumes of 150 million units, a penetration rate of 52%, and local models with 7 billion parameters, it is determined that this track will not rely on stacking high-cost advanced solutions.
When these five directions are placed side by side, the concept of a universal architecture becomes untenable. AI is willing to pay a power penalty for bandwidth; AI PCs are not. Automotive applications require reliability validation for over ten years, consumer electronics demand generational changes every eighteen months, and AI PC processors are limited to a power window of 15 to 80 watts with an iteration cycle of only 12 months. In this context, excessively pursuing advanced integration is actually a high-risk action. Placing the same Chiplet in different scenarios yields completely different values. Therefore, there is no universal Chiplet, only optimal Chiplet combinations tailored to specific scenarios. These five battlefields are not parallel relationships; their maturity, market size, and customer structures differ significantly. Who succeeds first depends on the timeline of demand-side requirements, not the degree of technological advancement on the supply side. However, divergence comes at a cost: the finer the scenarios, the more Chiplet categories, the messier the interfaces, and the more fragmented the industry becomes. A fragmented Chiplet market affects three parties simultaneously: design companies must repeatedly develop similar Chiplets for different customers; manufacturing and OSAT (Outsourced Semiconductor Assembly and Test) companies face unmergeable process requirements; and system vendors cannot obtain modules available for long-term supply. Ultimately, everyone reverts to single SoCs (System on Chip).
The diversity of scenarios is unavoidable, and system complexity has exceeded the closed-loop capability of any single entity. Evaluation standards must shift from single-point optimization to system-wide optimization, and the method to mitigate fragmentation risks is flexible combination plus an open ecosystem. These four main threads point to the same conclusion: the industry needs a layer of public interface order.
02 The Yardstick for Evaluating Chips Has Changed
Over the past two decades, the yardstick for measuring chip progress has essentially been singular: transistor gate length. As numbers like 7nm, 5nm, and 3nm decreased, performance increased. As process nodes approach physical limits, the marginal returns of this yardstick diminish, while bottlenecks at the system level grow thicker. Wang Zhimin proposed that the competition in chips has shifted from a transistor process race to a system metric race. There are three new yardsticks: how many transistors can be integrated per square millimeter, how many picojoules of energy are consumed per bit transmitted, and how much heat must be removed per watt of power consumption. These three figures correspond to integration density, interconnect energy efficiency, and power/thermal dissipation capabilities, describing whether the entire system can be effectively organized, rather than how advanced a single device is. Based on this judgment, Wang Zhimin condensed technical challenges into five key areas: new integration, optical interconnect, logic folding, electromagnetic-optical-mechanical-thermal multi-physics, and AI enablement. Integration solves how Chiplets are assembled; optical interconnect solves how Chiplets are connected; logic folding solves where density comes from; multi-physics solves whether failures occur after assembly; and AI enablement solves whether humans can still manage the design. These five directions correspond to five processes; if one link is missing, the system gets stuck at that link.
Following these three yardsticks, integration paradigms are progressing sequentially. 2.5D packaging is currently mainstream, with Chiplets laid side-by-side on silicon interposers. 3DI (Three-Dimensional Integration) stacks Chiplets vertically. Hybrid bonding pitch is shrinking from 10 microns to 3 microns, and further to below 1 micron, increasing interconnect density by an order of magnitude. Its feasibility depends on the product of bonding yield, Known Good Die (KGD), and thermal management, with a time window roughly between 2026 and 2030. WPI (Wafer-Level and Panel-Level Integration) pushes substrates from 300mm silicon wafers to 500mm square panels. CoWoS evolves into CoPoS, with economics depending on the product of area utilization, equipment precision, and warpage control, with a window between 2027 and 2032. VIFER (Vertical Integrated Wafer) goes a step further, treating the entire wafer as an integration unit, stacking compute, storage, and interconnect vertically. Yield and repairability are the current bottlenecks.
Glass substrates are a frequently mentioned link in this path. Their coefficient of thermal expansion is close to that of silicon, and they offer high flatness, benefiting optical coupling. Fine lines can achieve2-micron width and 2-micron spacing. The difficulties are concentrated in the hole formation quality of Through-Glass Vias (TGVs) and metallization adhesion. Wang Zhimin indicated an industrialization pace of 2028 to 2033. Viewing these milestones together reveals a trend: chip forms are transitioning from planar to 3D, from single dies to full wafers, and from devices to systems.
Copper interconnect remains the most efficient choice for short distances, but when bandwidth demands rise to a certain level, copper line loss and power consumption become unsustainable. Replacing ultra-short-distance electrical interconnects between Chiplets with optics is the basic logic of this route. Wang Zhimin divided optical interconnects into three stages. NPO (Near-Package Optics) has already seen large-scale deployment in 51.2T switching systems. Optical engines are close to the package and remain pluggable but have not yet entered inside the package. It aims to resolve the contradiction between panel density and maintainability. CPO (Co-Packaged Optics) encapsulates the optical engine and switching chip on the same substrate, reducing link energy consumption. Currently in the trial phase, engineering challenges include external light sources, fiber array coupling, pluggable optical connectors, and opto-electro-thermal coordination. OIO (Optical I/O) pushes optics between Chiplets and Chiplets, representing the final form. Difficulties shift to laser energy efficiency and lifespan, opto-electronic co-design methodologies, and optical path testability. These three stages are not simple replacements but progressively push the boundary of optics deeper into the chip. Each step forward increases the difficulty of maintainability, heat dissipation, and testing.
The material system consequently becomes complex. Indium Phosphide handles light emission, Silicon Photonics undertakes modulation and detection, Thin-Film Lithium Niobate provides high-bandwidth, low-loss modulation capabilities, and micro LED begins to appear in short-range optical interconnect solutions. The coexistence of multiple materials implies increased packaging difficulty, requiring coordination of thermal expansion, stress, and reliability of different materials within the same structure.
Among the five directions, τ Logic Folding is the most easily overlooked. As process advancement becomes increasingly expensive, where else can logic density be squeezed out? Wang Zhimin did not treat it as an isolated technology but placed it in the same lineage as 3DI and VIFER. The common goal of these three is to shorten the distance between logic units, with the difference lying in whether folding occurs at the architectural level or the integration level.
03 Rhythm Judgment Is More Important Than Single-Point Breakthroughs
No matter how long the technical list is, if it does not land in the industry, it remains just a list. Wang Zhimin summarized the core of ecosystem construction as three transformations: transforming manpower into talent, academic achievements into practical technologies, and technical products into mature industries. All three transformations occur at interfaces: the interface between people and positions, papers and engineering, and products and markets. If interfaces do not align, transformation stalls midway. Therefore, he provided a transmission chain: scenarios determine form, form constrains the technology roadmap, and iteration rhythm shapes the industrial ecosystem.
The rhythm chart from 2025 to 2030 unfolds eight directions simultaneously. 3DI moves from mass production of storage and sensors to logic-on-logic stacking. WPI transitions from tight capacity in advanced packaging to panel-level substitution in cost-sensitive categories. Glass substrates move from sample verification to mainstream options for large-size, high-density applications. Optical interconnects evolve from large-scale deployment of Near-Package Optics to Optical I/O entering between Chiplets. Power and thermal management advances from vertical power delivery plus cold plate liquid cooling to 2kW thermo-electric coordination becoming a standard design item. AI4EDA progresses from point-tool efficiency improvements to automatic exploration of the design space becoming the default. New structures move from Gate-All-Around (GAA) mass production to the successive landing of 2D materials. Yield engineering transitions from the initial establishment of KGD systems to mass production driven by yield models and digital twins. This rhythm chart needs to be categorized into five levels: materials, processes, structures, architectures, and EDA. All eight directions find their place within these five layers.
Viewing the table vertically by year, the cell for 2027 to 2028 contains the most information. During this period, 3DI must push hybrid bonding to 3 microns and fully transition HBM to hybrid bonding. Glass substrates move from sample verification to small-batch shipments, requiring synchronized support from equipment and materials. Backside power delivery is introduced during this phase, with two-phase cooling and immersion starting pilot programs. AI4EDA moves from point-tool efficiency to agent-based workflows, and STCO (System Technology Co-Optimization) truly lands in the design flow. New structures advance from GAA mass production to CFET (Complementary FET) engineering verification. Product and yield engineering must solidify Chiplet reliability and certification standards.
From an industry practice perspective, understanding Chiplet as "building blocks" tends to underestimate its weight. It changes the division of labor within chips and, consequently, the mode of industrial collaboration. When evaluation standards shift from nanometers to system metrics, when scenario divergence replaces universal architectures, and when optics and multi-physics coordination become engineering norms, victory no longer depends on which company achieves the strongest single-point technology. Instead, it requires forming a positive cycle among demand, innovation, and ecosystem.