{
  "id": "chiplet",
  "title": "Chiplet",
  "lang": "en",
  "source_lang": "en",
  "summary": "A chiplet is a separately fabricated semiconductor die designed to operate with other dies as part of a larger integrated system.",
  "category": "computer-science",
  "subcategory": "hardware",
  "type": "concept",
  "aliases": [
    "Chiplet die"
  ],
  "generated_at": "2026-11-03",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/chiplet/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/chiplet/",
      "json": "https://aiwiki.page/zh/chiplet.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "A chiplet is a separately fabricated semiconductor die designed to perform a defined function within a larger assembly of interconnected dies. Instead of implementing an entire processor or system on one piece of silicon, designers divide it into components—such as compute arrays, memory controllers, input/output circuits, or cache—and connect those components inside a package. The resulting assembly can function as a single integrated circuit from the perspective of the surrounding computer, even though its internal construction is physically distributed.\n\nChiplets have become important because improvements in semiconductor manufacturing do not benefit every circuit equally, and very large dies are expensive to design and manufacture. A chiplet architecture can combine advanced-process computing logic with less expensive peripheral circuitry, reuse components across products, and assemble systems larger than a single exposure field used in photolithography. These advantages come with additional communication, packaging, thermal, and testing requirements. A chiplet system is therefore not simply a conventional chip cut into pieces: its architecture must be designed around the physical and economic consequences of crossing die boundaries."
    },
    {
      "heading": "Definition and relationship to other integration approaches",
      "text": "A semiconductor die is an individual piece of processed wafer containing electronic circuitry. “Chiplet” describes its role in a larger design rather than a fixed size or manufacturing technology. A chiplet can be relatively small, but some components called chiplets are substantial dies containing billions of transistors. Conversely, a small die is not necessarily a chiplet: a standalone sensor or discrete component does not become one merely because it occupies little silicon area.\n\nThe defining idea is functional decomposition combined with close integration. A processor might contain several compute chiplets and one input/output die; another system might combine programmable logic, communications interfaces, and specialized accelerators. These components generally communicate through die-to-die interfaces that are designed for the short electrical distances and dense wiring available within a package. They may share memory addressing, power-management mechanisms, and system-control functions, although not every chiplet architecture provides all of these features.\n\nChiplets overlap with the older concept of the multichip module, an assembly containing multiple semiconductor dies on a common substrate. There is no universally enforced boundary between the two terms. “Multichip module” principally describes physical construction, whereas “chiplet” usually emphasizes architectural partitioning, component reuse, and interfaces between functional building blocks. Some historical multichip modules would fit a broad present-day definition of a chiplet system even though their designers did not use that term.\n\nA system-on-chip integrates functions such as processors, memory controllers, and peripheral interfaces into a single chip design, traditionally on one die. A chiplet-based assembly can implement comparable system functionality across several dies. “System-in-package” is broader still: it can include memory, sensors, passive components, and other devices that do not form a tightly coupled digital architecture. Thus, a system-in-package may contain chiplets without every device in the package being described as a chiplet.\n\nChiplets also differ from separate packaged chips mounted on a circuit board. Board-level connections normally traverse longer conductors and package interfaces, with different electrical loading and signaling requirements. Within one package, closely spaced dies can support much wider links at lower signaling energy, although the benefit depends on the implementation. They do not automatically match connections inside a monolithic die. Package wiring, transmitter and receiver circuits, buffering, clock crossings, and link protocols all introduce costs that an architect must account for."
    },
    {
      "heading": "Historical development and commercial adoption",
      "text": "The physical integration of multiple dies predates modern chiplet terminology by decades. Ceramic multichip modules were used in high-performance computing and other demanding applications because they could bring chips closer together than conventional circuit boards allowed. IBM developed several generations of dense packaging for mainframes, including thermal conduction modules used in the IBM 3081 generation introduced in 1980. Such systems established important techniques for interconnect density, cooling, and assembly, although they were not organized around today’s standardized chiplet interfaces.\n\nThe modern approach developed alongside increasing processor complexity and the rising cost of advanced manufacturing. A notable transitional example was Intel’s first generation of quad-core desktop processors in 2006, which placed two dual-core dies in one package. This arrangement demonstrated an economic advantage of assembling a larger processor from smaller dies, but it did not provide the degree of heterogeneous functional partitioning associated with many later chiplet designs.\n\nProgrammable logic was another early commercial application of dense multi-die integration. Xilinx’s Virtex-7 2000T, announced in 2011, used several FPGA dies connected through a silicon interposer. The assembly provided a larger programmable device than could conveniently be implemented as one conventional die. Its importance lay not only in combining silicon area but also in demonstrating that a package-level interconnect fabric could support closely cooperating logic components in a commercial product.\n\nAMD made chiplets central to high-volume processor families. First-generation EPYC server processors, introduced in 2017, combined four eight-core dies in one package. In 2019, the Zen 2 generation separated computing cores from much of the input/output circuitry. Desktop Ryzen products combined 7-nanometer compute dies with a 12-nanometer input/output die, while second-generation EPYC processors used a 14-nanometer input/output die. This made the choice of manufacturing process an explicit architectural decision rather than a property shared by every circuit in the processor.\n\nOther products extended the approach in different directions. Intel’s Lakefield processors, introduced in 2020, used its Foveros technology to stack compute and base dies. AMD’s Instinct MI200 accelerator family, announced in 2021, combined two graphics compute dies, while the Radeon RX 7900 series, launched in 2022, separated a graphics compute die from memory-and-cache dies. Apple’s M1 Ultra, also introduced in 2022, connected two M1 Max dies through a high-bandwidth package-level interconnect. These examples illustrate distinct motivations—manufacturing economics, product scaling, three-dimensional integration, and bandwidth—rather than a single standard chiplet architecture."
    },
    {
      "heading": "Manufacturing economics, yield, and process selection",
      "text": "One motivation for chiplets is the relationship between die area and manufacturing yield. A wafer contains a finite number of possible die sites, and defects can make individual dies unusable. As die area grows, fewer dies fit on the wafer, and each die covers more area in which a damaging defect might occur. Actual yield depends on defect clustering, circuit redundancy, manufacturing maturity, and design rules, but a simple model helps explain the basic effect.\n\nUnder a Poisson model, the probability that a die contains no randomly distributed fatal defects is\n\n\\[Y=e^{-AD_0},\\]\n\nwhere \\(A\\) is die area and \\(D_0\\) is the average density of fatal defects per unit area. If \\(D_0=0.1\\) defects per square centimetre, a die with an area of \\(8\\ \\mathrm{cm}^{2}\\) has an illustrative yield of about 45%, whereas a \\(2\\ \\mathrm{cm}^{2}\\) die has a yield of about 82%. These are model outputs, not predictions for any particular manufacturing process.\n\nThe economic advantage comes from fabricating, testing, and selecting smaller dies before assembly. Four smaller dies taken from a pool of tested parts can form a system even when other dies elsewhere on their source wafers are defective. This differs from requiring all four regions of a large monolithic die to work together at one fixed wafer location. However, smaller dies require additional interfaces, and the final assembly can fail during packaging. The relevant comparison is the cost of a functioning packaged system, not merely the yield of its constituent dies.\n\nA simplified expression is\n\n\\[C_{\\text{system}} \\approx \\frac{\\sum_i n_i C_i+C_{\\text{assembly}}+C_{\\text{test}}} {Y_{\\text{assembly}}},\\]\n\nwhere \\(n_i\\) is the number of chiplets of type \\(i\\), \\(C_i\\) is the cost of a tested usable chiplet, and \\(Y_{\\text{assembly}}\\) represents yield after assembly and final qualification. Real cost models also include substrate losses, rework, inventory, development expenses, and the economic value of partially functional products.\n\nProcess selection provides another major benefit. Dense digital logic often gains substantially from smaller transistors, whereas analog circuits, external interfaces, and some power-related structures may scale less effectively. A manufacturer can place compute logic on an advanced process while retaining input/output circuitry on an older, less expensive process. Reusing one input/output design across several processor generations can also reduce engineering expense.\n\nChiplets do not invariably lower cost. At low production volumes, custom packaging and multiple design flows may outweigh silicon savings. For a small monolithic device with high yield, splitting the design can add unnecessary area and complexity. The economic result depends on product volume, die size, process maturity, package technology, and how widely components can be reused."
    },
    {
      "heading": "Architectural partitioning and system organization",
      "text": "The first architectural decision is where to place die boundaries. A useful boundary minimizes expensive communication while separating functions that benefit from different manufacturing processes or development schedules. Repeated computing blocks are attractive candidates because a common design can be replicated to create products with different core counts. Input/output circuitry is another candidate because it can remain relatively stable while computing logic changes.\n\nA common processor arrangement places clusters of CPU cores and local cache on compute chiplets, with memory controllers and external connectivity concentrated in a separate input/output die. This central die can simplify communication with memory and peripheral devices. However, requests from a core may have to leave its compute die before reaching main memory, introducing additional link traversal and arbitration. The resulting latency and bandwidth must be evaluated against the savings in fabrication and design reuse.\n\nOther systems distribute memory controllers among compute dies. This can increase aggregate bandwidth and avoid a single centralized bottleneck, but access time may depend on which die owns the requested memory. Such arrangements are often described using non-uniform memory access, or NUMA. Software may need to place threads and data near the relevant memory resources. A package can appear as one processor socket while still containing important internal locality differences.\n\nMaintaining cache coherence across chiplets creates another architectural challenge. If several caches can hold copies of the same memory location, the system must coordinate their states so that software observes the required memory behavior. Coherence messages consume interconnect capacity in addition to ordinary data traffic. Directory placement, cache-line ownership, and the mapping of addresses to memory controllers can therefore have substantial effects on performance.\n\nNot every workload requires a coherent shared-memory design. A graphics processing unit or specialized accelerator may divide work into tiles that exchange larger blocks of data less frequently. This can make a chiplet boundary easier to tolerate. Conversely, workloads with fine-grained synchronization or unpredictable data sharing may expose the latency of crossing between dies. Effective partitioning depends on communication patterns, not simply on the number of arithmetic units.\n\nThe interconnect topology must also fit the workload. A central hub offers straightforward connectivity but can become a congestion point. Rings, meshes, and direct links distribute traffic differently and require different routing resources. Designers must consider peak traffic, average traffic, failure behavior, and the physical placement of interfaces. Adding more compute chiplets does not guarantee proportional speedup if memory bandwidth, interconnect capacity, or parallelism in the workload becomes the limiting factor."
    },
    {
      "heading": "Package structures and three-dimensional integration",
      "text": "The package provides the physical connections that allow chiplets to operate together. Conventional organic substrates can support multi-die assemblies, particularly when interface pitch and bandwidth requirements are moderate. They are comparatively economical and offer established manufacturing methods, but their wiring density is lower than that achievable with fine-patterned silicon structures. Dense interfaces may therefore require additional integration technologies.\n\nIn 2.5-dimensional integration, dies are commonly placed side by side on an interposer that provides fine-pitch routing. A silicon interposer can carry many short connections between logic dies and memory stacks. It may be largely passive, containing wiring and vertical connections rather than active computing circuitry. Its size, manufacturing cost, and susceptibility to defects affect the economics of the complete package.\n\nLocal silicon bridges provide an alternative to a full interposer. Intel’s Embedded Multi-die Interconnect Bridge, or EMIB, places small silicon routing structures beneath adjacent die edges within a package substrate. This concentrates fine-pitch wiring where it is needed rather than placing a large silicon sheet beneath the entire assembly. Fan-out redistribution technologies offer another family of approaches, using patterned metal layers to connect embedded or mounted dies. These methods differ in routing density, assembly sequence, achievable dimensions, and production infrastructure.\n\nThree-dimensional integration places active dies above one another. Connections may use solder microbumps or hybrid bonding, in which closely aligned metal contacts and dielectric surfaces are bonded without conventional solder bumps at each connection. Vertical electrical paths can include through-silicon vias, or TSVs, which pass through silicon to connect different levels of the stack. TSVs and hybrid bonding address different parts of the integration problem and are not interchangeable terms.\n\nStacking can provide very high connection density because interfaces are no longer restricted to neighboring die edges. A cache die, for example, can connect vertically to the logic beneath it through a large number of short links. AMD’s 3D V-Cache products, first commercialized in 2022, illustrate the use of vertically integrated cache to increase capacity without enlarging the compute die’s planar footprint. The performance effect depends on whether a workload benefits from the additional cache.\n\nThe most compact arrangement is not necessarily the easiest to manufacture or cool. Dies differ in thickness, thermal expansion, and mechanical behavior. Assembly must control alignment, warpage, bond quality, and stress through repeated temperature changes. A defect in a high-value stack may render several otherwise functional dies unusable. Package design therefore involves a coupled mechanical, electrical, manufacturing, and thermal problem rather than only a choice of interconnect density."
    },
    {
      "heading": "Die-to-die communication and interface standards",
      "text": "A die-to-die interface contains several logically distinct elements. Its physical layer defines signaling, electrical behavior, timing, and the arrangement of connections. Above that, link functions may provide initialization, error detection, retry, and flow control. Higher-level protocols determine what is transported: memory transactions, coherent cache messages, streaming data, or another traffic type. Two chiplets can share electrical characteristics without understanding one another’s transactions.\n\nThe physical design is shaped by the package. Short connections can use relatively low voltage swings and dense parallel signaling, reducing energy per transferred bit. Longer or more electrically difficult paths may require different signaling and equalization. The designer must balance bandwidth density, power, reach, clocking complexity, and the silicon area occupied by interface circuitry. Comparisons between interfaces are meaningful only when these conditions are specified.\n\nFor a simple unidirectional link, raw bandwidth can be expressed as\n\n\\[B_{\\text{raw}}=\\frac{NR}{8},\\]\n\nwhere \\(N\\) is the number of data lanes and \\(R\\) is the signaling rate in bits per second per lane. A 64-lane link operating at \\(16\\ \\mathrm{Gbit/s}\\) per lane therefore provides \\(128\\ \\mathrm{GB/s}\\) of raw bandwidth in that direction. Useful payload bandwidth is lower when framing, error-control information, idle cycles, or other protocol overhead is included. Adding the capacities of both directions produces an aggregate bidirectional figure, which should not be confused with one-direction throughput.\n\nUniversal Chiplet Interconnect Express, or UCIe, was announced in March 2022 by an industry consortium. Its original specification defined a standardized die-to-die interconnect and supported transport of protocols including PCI Express and Compute Express Link, as well as streaming operation. UCIe 1.1 followed in 2023, and UCIe 2.0 in 2024 expanded the specification to include features for three-dimensional integration and a system architecture addressing manageability, debug, and testing.\n\nAnother interface family, Advanced Interface Bus, or AIB, emerged from Intel technology and was used in the context of the U.S. Defense Advanced Research Projects Agency’s heterogeneous-integration initiatives. Proprietary links also remain important because manufacturers can optimize them for a controlled set of dies and packages.\n\nStandardization does not make arbitrary chiplets interchangeable. Components must agree on package geometry, lane organization, power supplies, thermal limits, initialization, transaction semantics, and software-visible behavior. Even two conforming interfaces may support different optional capabilities. Practical interoperability therefore requires a validated system specification in addition to a common link standard."
    },
    {
      "heading": "Memory, cache, and bandwidth-intensive applications",
      "text": "Memory is central to chiplet design because moving data can limit performance even when computing capacity is abundant. Separating compute and memory functions allows each to use a suitable manufacturing technology, but also introduces physical links that must carry the workload’s traffic. The key questions are how much data crosses each boundary, how often it crosses, and whether transfers can overlap with useful computation.\n\nHigh-bandwidth memory, or HBM, is frequently integrated alongside large processors and accelerators. HBM uses stacked memory dies and a wide interface to deliver high aggregate bandwidth within a compact footprint. Packages commonly connect the memory stacks and processor through a silicon interposer or another dense routing structure. HBM stacks are often discussed within the broader chiplet ecosystem, although terminology varies: some descriptions reserve “chiplet” for the logic components and identify memory stacks separately.\n\nThe wide, relatively short connections to HBM differ from the connections used by conventional memory modules on a motherboard. They can deliver substantial bandwidth without requiring every individual wire to operate at an exceptionally high signaling rate. The trade-off includes specialized packaging, limited package area, and the cost of assembling and testing multiple high-value components. Memory capacity and bandwidth cannot be increased indefinitely merely by adding more stacks, because routing, power, controller area, and cooling also impose limits.\n\nCache chiplets address a different part of the memory hierarchy. A larger cache can reduce the number of accesses that reach external memory, particularly when a workload repeatedly uses a working set that fits within the expanded cache. However, additional cache has little benefit for data that is streamed once and not reused. Its organization also matters: access latency, associativity, bandwidth, and coherence behavior can be as important as nominal capacity.\n\nBandwidth-intensive machine-learning workloads have made these issues particularly visible. Training and inference involve repeated movement of model parameters, activations, and intermediate results. A multi-die accelerator can divide matrix operations among compute chiplets, but communication between them may become expensive if the partition requires frequent exchanges of large intermediate tensors. Efficient execution depends on matching the computational partition to the physical links and available memory.\n\nThe roofline model provides a useful first approximation:\n\n\\[P\\leq \\min(P_{\\text{peak}},B_{\\text{memory}}I),\\]\n\nwhere \\(P\\) is attainable computational throughput, \\(P_{\\text{peak}}\\) is peak computing capacity, \\(B_{\\text{memory}}\\) is memory bandwidth, and \\(I\\) is arithmetic intensity, measured as operations per byte transferred. A chiplet system can have several relevant bandwidth limits, including local memory, die-to-die links, and external communication. Increasing peak arithmetic throughput alone does not remove these limits."
    },
    {
      "heading": "Power delivery, heat removal, and operating reliability",
      "text": "Chiplet integration changes where power is consumed and how heat leaves the system. A multi-die package may spread computing activity across a larger area, reducing some local heat concentrations. It may also place several power-hungry dies close together, increasing the total heat that a cooling assembly must remove. The result depends on die placement, workload, package construction, and operating limits rather than on the use of chiplets alone.\n\nDie-to-die communication consumes energy in transmitters, receivers, clock distribution, and associated logic. A useful first-order relationship is\n\n\\[P_{\\text{link}}=E_{\\text{bit}}R_{\\text{bit}},\\]\n\nwhere \\(E_{\\text{bit}}\\) is energy per transferred bit and \\(R_{\\text{bit}}\\) is the transfer rate. For example, transferring \\(1\\ \\mathrm{TB/s}\\), equivalent to \\(8\\times10^{12}\\) bits per second, at \\(1\\ \\mathrm{pJ/bit}\\) consumes 8 watts in the portion of the communication system covered by that energy figure. Published energy values require careful interpretation because some include only physical-layer circuitry while others include more of the link.\n\nPower delivery must supply high current while maintaining acceptable voltage at each die. Resistance causes voltage drop, and rapid changes in current interact with package inductance and capacitance. Closely placed decoupling capacitors and carefully designed power-distribution networks help control these effects. Different chiplets may require different supply voltages or separate power domains, increasing the number of rails and the complexity of package routing.\n\nThree-dimensional stacks add thermal constraints. An upper die can lengthen the heat-flow path from circuitry below it, and temperature variations can alter leakage current and timing. A simplified thermal relation is\n\n\\[\\Delta T=P\\theta,\\]\n\nwhere \\(\\Delta T\\) is temperature rise, \\(P\\) is dissipated power, and \\(\\theta\\) is effective thermal resistance. Actual packages require spatially resolved models because neighboring chiplets heat one another and because the thermal path is not uniform. Some architectures deliberately place lower-power functions in particular stack positions to manage these constraints.\n\nMechanical reliability is closely connected to temperature. Silicon, organic substrates, solder, and other materials expand at different rates. Repeated heating and cooling can stress connections, while large package dimensions can make warpage more difficult to control. Very fine-pitch connections also place demanding requirements on surface preparation and assembly precision.\n\nOperating control must account for the assembly as a whole. Independent voltage and frequency domains can improve efficiency, but activity on one chiplet may raise the temperature or alter the power budget of another. Sensors, throttling mechanisms, and firmware coordinate behavior across the package. A chiplet system’s sustainable performance can therefore differ from the sum of the isolated peak capabilities of its components."
    },
    {
      "heading": "Testing, verification, and the limits of interchangeability",
      "text": "Testing is especially important because a completed package can contain several expensive components. The industry term known good die describes a die that has passed an appropriate set of tests before assembly. It does not imply absolute certainty that the die is defect-free. Tests have finite coverage, and some faults appear only under particular temperatures, voltages, operating frequencies, or interactions with other components.\n\nWafer probing can test internal logic, memory arrays, and selected interface functions before dies are separated. Scan structures and built-in self-test provide access to circuitry that would otherwise be difficult to exercise. However, a dense die-to-die interface may not be fully testable at its final operating conditions while the die is still on the wafer. Additional testing after bonding is needed to detect missing connections, shorts, alignment problems, and failures that emerge only when the assembled system runs.\n\nThree-dimensional integration makes test access more difficult because some dies and connections become physically inaccessible. Test architectures must provide routes through the stack for control and observation. IEEE 1838-2019 addresses test access for three-dimensional stacked integrated circuits. Such standards help organize access, but manufacturers still need product-specific test patterns, diagnostic methods, and acceptance criteria.\n\nDesign verification is a separate problem from manufacturing test. A perfectly fabricated assembly can still contain architectural or protocol errors. Engineers must verify link initialization, clock-domain crossings, reset sequences, power-state transitions, coherence behavior, and recovery from communication errors. Formal verification can establish selected properties of interface logic, while simulation and hardware-based methods explore broader system behavior. Verification becomes especially demanding when chiplets are developed by different organizations with different assumptions.\n\nSoftware integration also influences interchangeability. A substituted compute or input/output chiplet can change memory locality, available bandwidth, device enumeration, or power-management behavior. The operating system, compiler, and runtime software may need to understand these differences. A common electrical interface cannot ensure that a replacement component preserves application performance or even exposes the same system functions.\n\nA commercial ecosystem also requires agreements about design information, qualification, security, liability, and product lifetime. Vendors must establish which party diagnoses a failed package and how faults are attributed among dies, assembly processes, and system design. Authenticity and supply-chain controls matter when components come from multiple sources, while test and debug interfaces must not create unintended access to protected functions.\n\nThese requirements distinguish three levels of modularity: reuse within one manufacturer, interoperability among selected partners, and broadly interchangeable components available from independent suppliers. The first is already common in major processor families. The latter two require progressively stronger interface definitions, qualification procedures, and commercial coordination. Standardized die-to-die links provide part of that foundation, but compatibility ultimately belongs to the complete physical, architectural, and operational system."
    }
  ],
  "links": [
    "semiconductor",
    "integrated-circuit",
    "photolithography",
    "transistor",
    "ibm",
    "central-processing-unit",
    "cpu-cache",
    "non-uniform-memory-access",
    "cache-coherence",
    "graphics-processing-unit",
    "parallel-computing",
    "high-bandwidth-memory",
    "machine-learning",
    "neural-network-inference",
    "roofline-model",
    "formal-verification",
    "operating-system",
    "compiler"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}