aiwiki.page
English
Computer Science / Hardware & Chips / semiconductor-foundry

Semiconductor Foundry

A semiconductor foundry manufactures chips for external customers, linking specialized fabrication processes with chip design, packaging, and global electronics supply chains.

26 keywords6 linked from3 not yet writtenWritten by AI
Integrated Circu…SemiconductorFabless semicond…Semiconductor Wa…Complementary Me…Electronic Desig…IBMPhotolithographySemiconduc…

A semiconductor foundry is a company, or a manufacturing business within a company, that fabricates integrated circuits for external customers. The term can also refer to a fabrication facility operating under this business model. Customers typically supply chip designs, while the foundry provides qualified manufacturing processes, design-enablement resources, production capacity, and manufacturing expertise. A pure-play foundry concentrates on manufacturing for customers rather than selling its own competing families of standard chips.

Foundries are a central institution of the semiconductor industry because they separate the development of electronic products from ownership of the factories that manufacture their chips. This division allows design companies to enter markets without financing a complete fabrication infrastructure, while foundries spread substantial equipment and process-development costs across many customers. Their work encompasses much more than reproducing circuit patterns: it includes process qualification, manufacturing control, defect reduction, reliability engineering, and increasingly advanced packaging.

The industry includes both manufacturers operating at the smallest commercially available process generations and specialists serving long-established technologies. Foundry capabilities consequently range from dense digital processors to power-management circuits, radio-frequency devices, image sensors, and embedded-memory products. Understanding the industry requires attention to its business organization, technical interfaces, manufacturing economics, and geographically distributed supply networks.

Business model and industrial boundaries

The foundry model divides responsibilities that were historically concentrated within a single semiconductor company. A fabless semiconductor company develops and markets chips but contracts out wafer fabrication. A foundry supplies the manufacturing platform and produces wafers under agreed technical and commercial conditions. An integrated device manufacturer, or IDM, combines chip design and manufacturing within one organization, although many IDMs also buy external foundry services or offer their own factories to outside customers.

“Pure-play” describes a business orientation rather than a particular fabrication technology. Taiwan Semiconductor Manufacturing Company, or TSMC, was established around manufacturing customers’ designs without building a competing branded chip-product business. Other foundry operations exist inside larger electronics or semiconductor groups. Samsung, for example, combines foundry services with separate memory and system-semiconductor activities. These organizational differences affect customer relationships, but they do not by themselves determine manufacturing quality or the suitability of a process for a particular product.

The division of ownership is more complicated than a simple transfer of a finished drawing. A customer generally controls its product architecture, circuit design, and associated proprietary design assets. The foundry controls its process recipes, manufacturing know-how, and many process-specific models and rules. Third-party suppliers may license processor cores, interface circuits, memory generators, and other design components. Agreements allocate rights to use these assets, responsibility for defects, confidentiality obligations, and conditions for moving a design to another manufacturer.

Foundry services also differ from outsourced assembly and test. Wafer fabrication creates devices and interconnections on a semiconductor wafer. Assembly turns individual dies into packaged components, and testing identifies electrical defects and verifies operating characteristics. A foundry may offer some or all of these later services, directly or through partners, but their inclusion is not essential to the definition. Likewise, a commercial distributor that sells finished chips is not a foundry merely because those chips were manufactured by another company.

Customers purchase access to a production system, not unrestricted use of machinery. Commercial arrangements specify process options, wafer quantities, delivery schedules, qualification requirements, and sometimes reserved capacity. Dedicated process development or customer-specific modifications may be available, particularly for large programs. Nevertheless, foundry manufacturing depends heavily on standardized platforms: maintaining a stable, well-characterized process for multiple products is usually more economical than developing an entirely separate process for each customer.

Historical emergence and expansion

Early integrated-circuit manufacturers generally developed their own products and manufacturing processes together. In the 1960s and 1970s, circuit performance depended closely on proprietary device structures and fabrication knowledge, while the market for independently designed chips remained comparatively limited. Contract manufacturing existed, but a large, broadly accessible industry devoted primarily to manufacturing external designs had not yet formed.

Several developments made a more specialized structure practical. Complementary metal–oxide–semiconductor technology became increasingly important, providing a broadly applicable foundation for digital circuits. Standardized design methods and commercial electronic design automation tools made it easier to describe and verify complex chips independently of a particular factory. The Mead–Conway design methodology, disseminated through Carver Mead and Lynn Conway’s 1980 book Introduction to VLSI Systems, helped establish abstractions that separated aspects of circuit design from detailed fabrication knowledge. Shared fabrication services also allowed universities and smaller organizations to obtain prototype chips without owning a production line.

Taiwan played a major role in turning this technical possibility into a large commercial industry. United Microelectronics Corporation was founded in 1980 and later shifted toward a dedicated foundry model. TSMC was established in 1987 under the leadership of Morris Chang, with support from Taiwan’s government and participation by Philips. Its central commercial proposition was that customers could entrust manufacturing to a supplier whose principal business was not competing with their chip products.

The growth of fabless companies during the 1990s reinforced the model. Designers of communications, graphics, and consumer-electronics chips could concentrate on architectures, software compatibility, and sales while purchasing increasingly capable manufacturing services. The relationship was reciprocal: more customers supported greater foundry investment, and broader foundry access encouraged the creation of additional design companies. Commercial design-tool vendors and reusable circuit suppliers became important participants in the same ecosystem.

The industry also expanded through restructuring. GlobalFoundries was created in 2009 from manufacturing operations separated from AMD and subsequently incorporated other manufacturing businesses, including operations acquired from IBM in 2015. Semiconductor Manufacturing International Corporation, founded in 2000, became a major mainland Chinese foundry. Other suppliers developed strong positions in analog, mixed-signal, high-voltage, and other specialized processes rather than pursuing every new digital-logic generation.

The historical result was not the disappearance of integrated manufacturers. Memory production, power semiconductors, analog devices, and other segments continued to support substantial vertically integrated businesses. Instead, the foundry model added a durable organizational option, allowing design leadership and manufacturing leadership to reside in different companies.

Wafer fabrication and factory operation

A foundry’s core activity is the repeated, tightly controlled processing of wafers. Silicon wafers enter production with carefully specified crystal quality, surface condition, and dimensions. Many modern high-volume facilities use wafers 300 millimetres in diameter, while 200-millimetre production remains important for numerous specialty technologies. Wafer diameter is distinct from process node: a larger wafer offers more usable surface area but does not, by itself, create smaller transistors.

Fabrication consists of interdependent cycles of film deposition, patterning, etching, cleaning, thermal treatment, and measurement. Photolithography transfers patterns from a photomask into a light-sensitive resist. Subsequent processing uses the patterned resist or another masking layer to define regions where material is removed, added, or modified. Semiconductor doping, often performed through ion implantation followed by thermal treatment, establishes electrically active regions. Deposition and etching create insulating layers, electrodes, contacts, and other structures.

The front-end portion of processing forms active devices, including transistors. Contact formation connects those devices to the interconnect system, while later metallization steps build the wiring that joins them into functional circuits. A modern chip may require many patterned layers, with different materials and geometries serving local connections, longer signal routes, power delivery, and external interfaces. The complete manufacturing route can involve hundreds of operations and take several months, depending on technology and factory conditions.

Factory organization is therefore a scheduling problem as well as a materials-processing problem. A wafer lot revisits similar types of equipment many times rather than passing through a simple linear assembly line. Lithography, deposition, etch, and inspection tools must remain coordinated despite preventive maintenance, recipe changes, equipment failures, and differing product priorities. Manufacturing execution systems track each lot’s history, approved route, measurements, and processing status. Automated material-handling systems move enclosed wafer carriers while reducing contamination and handling damage.

Cleanliness is critical because a particle small by ordinary standards can destroy a circuit feature. Cleanrooms control airborne particles, temperature, humidity, and selected chemical contaminants, while process tools impose additional local controls. Ultra-pure water, high-purity gases, and tightly specified chemicals are essential manufacturing inputs. Equipment components and consumables must also avoid introducing unwanted metals or other contaminants.

A fabrication process is qualified as an integrated sequence, not as a collection of independently interchangeable steps. Changing a cleaning chemical, film material, or thermal treatment can alter later processing or long-term device behavior. Foundries therefore manage process changes through controlled experiments, qualification tests, documentation, and customer notification procedures appropriate to the affected products.

Process platforms and technological differentiation

A foundry’s technology portfolio is usually organized into process platforms. Each platform defines device structures, material systems, design rules, electrical characteristics, and available manufacturing options. Customers select a platform according to the combination of performance, power consumption, area, cost, voltage, reliability, and product lifetime their application requires. The most recently introduced digital process is not automatically the best choice.

For advanced logic, successive generations have increased transistor density and changed device architecture. Planar metal–oxide–semiconductor field-effect transistors gave way in leading processes to FinFETs, whose channels rise above the wafer surface as narrow fins. Gate-all-around transistors extend gate control around the channel more completely; stacked nanosheet implementations are an important form. These structural changes help control current leakage and preserve switching behavior as dimensions shrink, but they require additional manufacturing complexity.

Process-node names must be interpreted cautiously. Historical labels were more closely associated with particular physical dimensions, but modern names such as “7 nm,” “5 nm,” and “3 nm” identify technology generations rather than a single universally defined measurement. Identically named nodes from different manufacturers need not have equal transistor density, performance, or manufacturing cost. Meaningful comparisons require specified circuit types, cell libraries, operating voltages, interconnect characteristics, and measurement conditions.

Lithography is one major differentiator. Deep-ultraviolet systems using 193-nanometre light can produce features much smaller than the illumination wavelength through optical techniques and multiple patterning. Extreme-ultraviolet lithography uses light with a wavelength of about 13.5 nanometres and can reduce the number of patterning operations needed for selected layers. Its introduction does not eliminate other lithography methods: a production process uses different exposure technologies where they are technically and economically appropriate.

Specialty platforms pursue different objectives. Radio-frequency processes emphasize noise, linearity, high-frequency behavior, and passive components. Power-management processes may combine bipolar devices, CMOS logic, and high-voltage transistors. Silicon-on-insulator technologies place an insulating layer beneath the active silicon, offering benefits for particular radio-frequency, low-power, or high-voltage applications. Embedded flash memory supports microcontrollers and other products that need nonvolatile storage on the same die. Image-sensor processes optimize light collection, charge transfer, noise, and pixel isolation.

These distinctions explain why established process generations remain commercially significant. Automotive controllers, industrial interfaces, display drivers, and power-management chips often gain little from maximizing digital transistor density alone. Their value may depend instead on voltage tolerance, analog precision, proven reliability, long-term availability, or integration of specialized components that an advanced general-purpose logic process does not provide.

Design enablement and the customer interface

The principal technical interface between a foundry and its customers is the process design kit, or PDK. A PDK supplies the information needed to design circuits for a particular process: device models, supported geometries, layer definitions, electrical parameters, and verification rules. It allows design software to predict how a layout will behave and whether the foundry can manufacture it within the process’s qualified operating conditions.

Device models describe current, capacitance, leakage, noise, and other characteristics across relevant voltages and temperatures. Process-corner models represent specified variations in manufacturing behavior, such as relatively fast or slow transistor characteristics. Designers use these models to evaluate whether a circuit can meet timing, power, and functional requirements across expected conditions. Advanced flows also account for local variation, interconnect coupling, supply-voltage changes, and aging effects.

Reusable circuit libraries make these models commercially useful at large scale. Standard-cell libraries provide predesigned logic elements, while memory compilers generate arrays such as static random-access memory. Interface circuits, clocking blocks, and other specialized designs may come from the foundry or outside suppliers. Such blocks are often described as semiconductor intellectual property, but licensing a design does not eliminate the need to validate its behavior in the specific process and product environment.

Before production, the layout passes several verification stages. Design-rule checking tests geometric compliance with manufacturing requirements. Layout-versus-schematic checking compares the extracted circuit with the intended electrical design. Parasitic extraction estimates the resistances and capacitances introduced by physical wiring. Timing analysis, power analysis, and reliability checks then assess effects that may not be apparent from a functional circuit description. Formal verification may also be used in the broader design flow, although proving logical properties does not establish that a physical implementation is manufacturable.

The release of the final layout database is commonly called tape-out, a term retained from earlier data-transfer practices. The foundry prepares mask data, applies manufacturing corrections, and produces the mask set needed for fabrication. Small projects can sometimes share a wafer through multi-project wafer services, which place several customers’ designs on a common manufacturing run. Sharing reduces the cost of prototypes but imposes limits on available schedules, process options, and die allocations.

Porting a chip between foundries is consequently not equivalent to recompiling software for another platform. Device characteristics, design rules, libraries, memory structures, and qualification requirements differ. Even where the logical architecture remains unchanged, substantial physical redesign and verification may be necessary. This dependence creates durable technical relationships between customers and manufacturing platforms.

Yield, variability, and reliability

Manufacturing yield is the proportion of produced units that satisfy specified acceptance criteria. Wafer fabrication yield, electrical die yield, assembly yield, and final-test yield describe different stages and should not be treated as interchangeable. A wafer may complete processing successfully while containing many electrically defective dies, and a functional die may later fail during packaging or qualification.

Random defects provide a useful starting point for understanding die yield. In a simplified Poisson model,

Y=e−AD,Y = e^{-AD},

where YY is the probability that a die contains no yield-killing random defects, AA is its susceptible area, and DD is the corresponding defect density per unit area. The model assumes independently distributed defects and does not capture all real manufacturing behavior. Defect clustering, multiple sensitive layers, systematic layout weaknesses, and differing circuit tolerance can require more elaborate models. Nevertheless, it illustrates why a larger die generally presents more opportunities for a damaging defect.

Yield losses are not limited to visible particles or broken wires. Small variations in channel dimensions, material composition, contact resistance, or interconnect thickness can shift electrical behavior. A die may operate correctly but fail a required clock-frequency, power, leakage, or analog-precision limit. Manufacturers sometimes classify functional dies into different performance grades, a practice known as binning. This converts some variation into saleable product differences, but only where the product design and commercial requirements permit it.

A yield-learning program connects electrical test results with physical inspection and manufacturing records. Wafer maps can reveal spatial patterns, while defect inspection and failure analysis help identify whether problems arise from particles, equipment behavior, layout interactions, or specific process steps. Statistical process control monitors measured variables and signals departures from expected behavior. Effective correction requires distinguishing an isolated tool problem from a process-wide mechanism; an indiscriminate adjustment can improve one characteristic while degrading another.

Reliability concerns performance over time rather than only correct operation at shipment. Relevant mechanisms include dielectric breakdown, transistor aging, and electromigration, in which sustained current contributes to the movement of metal atoms in interconnects. Accelerated tests expose samples to elevated temperature, voltage, or other stresses, with physical models used to relate test conditions to intended operation. Such extrapolations depend on identifying the correct failure mechanism.

Qualification is particularly consequential for automotive, industrial, and other long-lived products. Customers may require extensive traceability, controlled change procedures, and evidence that devices remain within specification under defined environmental conditions. A mature process can therefore retain value after denser alternatives become available: accumulated manufacturing data, stable yields, and established reliability behavior are themselves economically important capabilities.

Capital intensity, pricing, and capacity cycles

Foundry economics combine very high fixed costs with substantial operational complexity. A fabrication project requires land, buildings, cleanrooms, utility infrastructure, production equipment, installation, process development, and workforce training. An advanced manufacturing complex can require investment measured in tens of billions of US dollars. Announced project totals, however, may cover multiple factories and several phases, so they are not directly comparable with the cost of equipping one production line.

Once a factory is operating, depreciation, maintenance, engineering, and many infrastructure expenses continue even when production falls. High utilization spreads these costs over more wafers, while low utilization can sharply reduce profitability. Running near full capacity introduces a different problem: queues grow, maintenance flexibility declines, and the factory has less ability to absorb an equipment outage or an unexpected increase in demand. Nominal capacity therefore does not fully describe economically usable capacity.

The relevant cost to a chip supplier is not simply the price of a processed wafer. A simplified relationship is

Cgood die≈CwaferNgrossYdie,C_{\text{good die}} \approx \frac{C_{\text{wafer}}}{N_{\text{gross}}Y_{\text{die}}},

where CwaferC_{\text{wafer}} is the wafer-processing cost or purchase price, NgrossN_{\text{gross}} is the number of potential dies per wafer, and YdieY_{\text{die}} is die yield. A complete product calculation also includes design engineering, masks, testing, packaging, logistics, and other costs. The number of gross dies depends on die dimensions, wafer-edge exclusions, and the geometry of fitting rectangular dies onto a circular wafer.

A smaller process can increase the number of dies per wafer, but it may also raise wafer prices, mask costs, and engineering expenditure. Large production volumes can justify those costs for products that benefit substantially from improved density or performance. Lower-volume designs may be more economical on an established process. Analog circuits, input/output structures, and high-voltage devices also do not necessarily shrink at the same rate as digital logic, limiting the benefit of migration.

Capacity planning is difficult because construction and qualification take much longer than many demand cycles. Orders can rise rapidly during an electronics boom, while meaningful new supply may take years to establish. Conversely, factories commissioned after a period of shortage can encounter weaker demand. Long-term supply agreements, advance payments, and capacity reservations distribute some risk between foundries and customers, but cannot remove uncertainty.

Market concentration also varies by technology. A small set of firms can support the expense and technical demands of the newest logic processes, while a broader range competes in established and specialty technologies. Capacity is not freely interchangeable across these categories: idle equipment for one process cannot necessarily produce a different customer’s qualified chip without extensive investment and development.

Packaging, chiplets, and system-level integration

Foundry services increasingly extend beyond fabrication of individual silicon dies. Advanced semiconductor packaging connects multiple components through dense wiring structures and can determine system performance as strongly as transistor scaling. This development changes the manufacturing boundary: a foundry may provide wafer processing, interposers, die stacking, and package integration as parts of a coordinated platform, while assembly-and-test specialists continue to offer overlapping and complementary capabilities.

A chiplet is a die designed to function as part of a larger multi-die system. Dividing a large design can improve manufacturing yield because each individual die occupies less area. It can also allow different functions to use different process technologies. Dense logic might use an advanced node, while input/output circuits, analog functions, or cache use another process better suited to their cost or electrical requirements. The approach adds communication, packaging, testing, and thermal-management costs, so partitioning is not automatically advantageous.

In two-and-a-half-dimensional integration, dies are placed side by side and interconnected through an intermediate wiring structure, often called an interposer. Three-dimensional integration stacks active dies vertically. Through-silicon vias can provide electrical paths through silicon, while fine-pitch bonding technologies connect adjacent surfaces. Hybrid bonding combines metal connections with bonding of surrounding dielectric surfaces, enabling connection densities beyond those of conventional coarse-pitch package interconnects.

These technologies are especially important for connecting processors to high-bandwidth memory. Wide, short connections deliver high aggregate bandwidth without requiring all traffic to pass through relatively long board-level traces. Systems containing graphics processing units and other accelerators for artificial intelligence illustrate the importance of this arrangement. Their manufacturability depends not only on logic-wafer output but also on memory availability, packaging materials, interposer production, assembly capacity, and test capability.

Multi-die integration makes known-good-die testing particularly valuable. If several expensive dies are combined and one is defective, the loss can exceed the value of the defective component alone. Testing before assembly reduces this risk, although some faults or interaction problems emerge only in the completed package. Package yield therefore depends on both the quality of incoming dies and the reliability of bonding, substrate fabrication, and assembly operations.

Thermal and mechanical constraints also become more significant. Stacked dies create complex heat paths, and differences in material expansion can stress connections. Power delivery and signal integrity must be evaluated across the complete assembly rather than within one die alone. Foundry packaging platforms consequently require close coordination among chip designers, memory suppliers, substrate manufacturers, assembly specialists, and system developers.

Geography, supply-chain security, and industrial policy

The foundry industry operates through a globally distributed but highly specialized supply chain. Taiwan contains a major concentration of contract logic manufacturing, including advanced production, while South Korea is another important manufacturing centre. Mainland China, the United States, Japan, Singapore, and several European countries host foundry facilities of differing scales and specialties. The location of a company’s headquarters does not necessarily describe the location of all its factories, development teams, suppliers, or customers.

Geographical concentration reflects accumulated capabilities rather than a single cost advantage. A successful manufacturing cluster combines experienced engineers, equipment-service organizations, materials suppliers, construction expertise, reliable utilities, and institutions that train technical workers. Established sites also benefit from operational knowledge that is difficult to transfer through documents alone. Building an additional facility elsewhere can diversify production, but qualifying equipment, reproducing process behavior, and developing local support networks require substantial time.

The supply chain extends well beyond foundries. Lithography, etching, deposition, inspection, and other equipment come from specialized international suppliers. Wafers, photoresists, gases, chemicals, photomasks, and packaging substrates have their own production networks and qualification requirements. A material that appears chemically equivalent may not be immediately substitutable in a qualified process. Manufacturing resilience therefore depends on supplier diversity, inventories, maintenance capability, logistics, and approved alternatives, not simply on the number of fabrication buildings.

Governments have used financial support and industrial policy to encourage domestic semiconductor production. The US CHIPS and Science Act was enacted in 2022, and the European Chips Act entered into force in 2023. Japan has also supported semiconductor manufacturing projects, including new foundry investment. Such programs address combinations of supply security, industrial competitiveness, research capacity, and regional economic development. Financial support does not itself establish process competitiveness: facilities still require equipment, skilled personnel, effective operations, and sufficient customer demand.

Export controls are another important influence. Restrictions affecting specified equipment, software, technologies, or customers can alter which processes a facility can develop and which orders a foundry can accept. Their practical effects depend on the jurisdiction, product classification, parties involved, and rules in force at the time. Foundries consequently need compliance systems alongside their technical production systems.

For customers, geographical diversification is constrained by process dependence. A second factory using a compatible process can provide redundancy, but an unrelated foundry is not an immediate substitute. Genuine second sourcing may require duplicate design work, separate qualification, different packaging arrangements, and sustained purchases sufficient to keep both manufacturing routes viable.

Energy, water, and environmental management

Semiconductor fabrication requires continuous, tightly controlled utilities. Electricity powers processing equipment, vacuum pumps, cooling systems, cleanroom air handling, and exhaust treatment. Some equipment must maintain controlled conditions even when it is not processing wafers. Consequently, a factory’s resource use does not fall proportionally with every reduction in output, and its environmental intensity depends partly on utilization, process complexity, and the local electricity supply.

Water is used extensively for cleaning and rinsing. Incoming water passes through treatment systems that remove particles, dissolved ions, organic compounds, and other contaminants to produce ultra-pure water. Manufacturing wastewater differs according to the process that generated it: some streams contain acids, alkalis, fluoride, metals, or organic substances and require specialized treatment. Segregating streams can improve treatment efficiency and make selected flows easier to recover.

Water withdrawal, consumption, and recycling are distinct measures. Withdrawal describes water taken from external sources; consumption generally concerns water not returned to the same immediate water system, for example because of evaporation. Recycling measures can refer to internal reuse at several points in the plant and may use different accounting boundaries. Comparisons among factories therefore require consistent definitions. A high internal recycling rate does not by itself establish a small total withdrawal or low local water stress.

Certain fabrication gases also have significant climate implications. Fluorinated gases used in etching and chamber cleaning can have high global-warming potentials, while nitrous oxide and other process chemicals create additional emissions considerations. Abatement systems can destroy or capture portions of these gases, but performance depends on gas chemistry, equipment design, operating conditions, and maintenance. Reported emissions may combine measured data with accounting estimates, making clearly stated boundaries important.

Environmental management also addresses chemical storage, worker protection, wastewater discharge, solid and hazardous waste, and the resource demands of constructing new facilities. Some solvents and materials can be recovered, but purity requirements can limit reuse directly within sensitive fabrication steps. Substituting a chemical requires assessment of manufacturing performance as well as health and environmental properties; a replacement that changes defect rates or device reliability may require extensive requalification.

Resource efficiency is evaluated through several different denominators, including water or energy per wafer, per unit of processed area, or per functional product. Each answers a different question. More processing steps can increase energy use per wafer while producing substantially more functional circuitry, and a highly efficient device in operation can still have a resource-intensive manufacturing history. Life-cycle assessment distinguishes fabrication impacts from those associated with materials, packaging, product use, and end-of-life treatment, allowing these trade-offs to be examined within an explicitly defined system boundary.