Computer architecture

• Type of physical science: Computation
• Field of study: Computers

Computer architecture refers to the design and organization of a computer’s hardware components and their interaction to support the execution of software. It includes the use of digital circuits to perform logical and arithmetic operations, manage data flow, and store information in memory and storage devices. When hardware components are effectively coordinated, they provide a reliable and programmable platform for running applications and operating systems.

Overview

Computer architecture deals with the design, structure, and behavior of the components of computer systems and their integration. It encompasses the development of hardware technologies and their evolution to improve performance, reliability, and scalability. Computers consist of five primary hardware components: the central processing unit (CPU), random-access memory (RAM), secondary storage devices, input devices, and output devices.

The CPU, sometimes called the brain of the computer, performs data manipulation and executes program instructions. It conducts arithmetic and non-numeric operations in the arithmetic/logic unit (ALU) and control functions in the control unit.

The ALU is a CPU component that performs addition, subtraction, multiplication, and division on several types of numbers, such as integers and floating-point real numbers, as well as logical operations like AND, OR, and NOT. Arithmetic is carried out over binary numbers, or numbers stored as 0s and 1s, based on discrete voltage levels. Binary is used since those two values can be interpreted from the absence or presence of an electric pulse along a wire. Generally, numbers are temporarily stored in the CPU in high-speed storage circuits called registers. New processors typically use 32-bit or 64-bit registers, with some supporting even wider vector registers for parallel processing tasks. In addition to arithmetic, the ALU (often with a floating-point unit, FPU) can evaluate conditions, support branching, and execute a broad range of operations defined by the processor’s instruction set architecture (ISA).

The control unit (CU) portion of the CPU supervises the step-by-step execution of program instructions and manages the flow of data between the CPU’s registers, the arithmetic/logic unit (ALU), memory, and input/output systems. The control unit tracks the current instruction using special-purpose registers such as the program counter and instruction register. Control units occur in two types: hardwired and microprogrammed. In new CPUs, control units are usually hardwired for performance, but microprogrammed control is still used in some cases. The CU ensures that instructions are decoded and executed in the correct sequence, coordinating efficiency in all parts of the processor.

In early computer architectures, control units were often classified as hardwired or microprogrammed. A hardwired control unit uses fixed logic circuits to directly decode and execute instructions, offering faster performance; a microprogrammed control unit stores a set of microinstructions in control memory (typically read-only memory), which translates machine-level instructions into sequences of low-level operations. This design increases flexibility and simplifies the addition of complex instructions but may result in slower execution. While microprogrammed control was common in early complex instruction set computers (CISC), most processors use hardwired control for performance. Some CISC CPUs still use microcode internally for complex operations, with the added ability to patch or update microcode through firmware.

CPUs have a small number of registers complemented by RAM, allowing random access to any memory location in constant time. It is organized into modules with rows, columns, and banks, each location identified by a unique address. In some systems, the CPU connects to RAM using high-speed memory channels (often DDR4 or DDR5 technology). Other components like graphics processing units (GPUs) and storage devices interface with the CPU through dedicated serial interfaces like PCI Express. Program execution follows a structured cycle—the CPU retrieves an instruction from memory, decodes it, retrieves or manipulates data as needed using registers and arithmetic logic units (ALUs), and writes results back to memory. This cycle repeats at speeds governed by the system clock, which generates precise timing signals. Measured in gigahertz (GHz), the system clock coordinates the processor’s and other components’ activities.

Memory occurs in two forms: volatile and non-volatile. Historically, volatile memory such as RAM used silicon chips that stored data temporarily—losing all contents when power dropped to near zero. One early form of nonvolatile memory was battery-backed RAM, which maintained stored data using a constant battery-powered current. Nonvolatile memory technologies largely replaced battery-backed RAM systems, including flash-based storage such as solid-state drives (SSDs), USB flash drives, and NVMe (Non-Volatile Memory Express) devices. These technologies retain data even after power is shut off and are faster, more durable, and more energy efficient. For high-performance accelerators, memory technology also advanced rapidly, with high bandwidth memory (HBM) products such as HBM3E reaching over 1.23 TB/s per stack and theHBM4 standard released in 2025 reaching even higher bandwidth and capacity.

External storage devices—crucial for saving data and programs before use, after execution, and during power-offs—have also evolved. Historically, magnetic media like floppy disks, magnetic tapes, and hard disk drives (HDDs) were standard. These devices stored data by altering magnetic states on disk platters or tape reels. Tape drives were used primarily for long-term backup. Disk drives were often the first component of a computer to crash. To preserve the data on them, data was periodically copied to tapes, a type of removable media that had to be mounted before usage. Since data were copied or read by the tape head as the tapes unwind, data are stored and accessed sequentially as a long stream. Speed was rarely the major issue for tape drives because their usage was periodic. Instead, data density was more crucial. Since several complete backups were commonly needed and disk drives had achieved greater data densities, the shelf space for backups and the effort of mounting many tapes per backup became important. Tape drives offered average tape capacities from 40 megabytes to 5 gigabytes. The situation worsened when large numbers of floppy disks (flexible, removable, low-data-density disks) were used to back up hard drives (a common practice among personal computer users). Another application of floppy disks and tapes was distributing and transporting software or data.

Eventually, other technologies replaced tapes as a backup medium. Besides removable disks, disk drives using lasers to mark data onto treated surfaces (called optical disks) became a cost-effective solution that offered random-access capabilities, though significantly slower than standard magnetic disk drives. CD-ROMs (compact disk storage device read-only memory) also became popular as an alternative distribution media for large amounts of data, but CD-ROMs were useless for backups.

Though magnetic storage remains in use, particularly for high-capacity archival systems, SSDs are now widely adopted for internal and external storage due to their speed and reliability. Older forms of storage had capacities that ranged from a few megabytes to a few gigabytes, but modern consumer-grade storage devices commonly range from 500 gigabytes to over 20 terabytes. Enterprise systems can handle petabytes of data.

Another category of early computer peripherals included input devices such as keyboards, card readers, teletypes, and early graphically oriented pointing devices like mice. These devices often transmitted data via serial cables and relied on interrupts and buffers to manage data flow, reflecting the slower hardware speeds of the time. Several terminals could be connected for input and output in multi-user environments, often using different communication protocols over serial lines. Input methods later expanded to include voice recognition systems, touchscreen interfaces, high-speed scanners, and specialized scientific instrumentation requiring rapid data throughput and processing.

Output devices historically included printers, plotters (used for technical drawings and schematics), and monitors for real-time display. Early printers connected to computers using cables, but later transitioned to USB, Ethernet, and wireless connections like Bluetooth and Wi-Fi. Monitor output was once a complex process: data was sent to a dedicated graphics card that translated digital signals into commands for cathode-ray tube (CRT) monitors, which operated similarly to television sets. Resolution, refresh rates, and color depth were key concerns, and early monochrome text-only cards used only a few kilobytes of memory. In the early 2000s, color graphics cards supporting millions of pixels and 256-color palettes with one or more megabytes of dedicated video RAM became standard. By the 2020s, display technologies such as LED, OLED, and high-resolution 4K and 8K panels dominated the market. Advanced GPUs could have real-time rendering, ray tracing, and artificial intelligence acceleration; many had gigabytes of video RAM and supported gaming, professional design, machine learning, and scientific visualization.

Networking has also evolved significantly over time. Initial communication between systems used modems that converted digital data into analog signals over telephone lines. Early local area networks (LANs) used dedicated network cards and Ethernet cables, often requiring matching protocols and manual configuration. Many computers and devices support high-speed networking through gigabit or 10-gigabit Ethernet, Wi-Fi 6/6E, or even cellular 5G, with plug-and-play standards and sophisticated routing protocols that enable seamless global connectivity.

Applications

The application of technologies to produce microcomputers created the personal computer revolution, raising the state of human mathematical and textual processing capabilities like no other technological revolution. Typewriters and their inherent limitations have made them  historical artifacts. High-speed computation is no longer charged at $20 per minute of mainframe CPU time. The computational aspects of research have become more affordable, especially for individuals. Significant grant funds are no longer devoted to renting CPU time and paying monthly disk storage charges. Similarly, with new input technologies, the frustrations of old punch card decks are gone. Single-user computers with interactive processing are a significant result of the application of advances in hardware.

Users can apply their knowledge of computer architecture in several ways.

With standardized expansion slots and expansion boards, the user can custom-design a computer to meet usage requirements exactly. By learning the disk drive types supported by their specific computers, users can purchase higher-speed or higher-capacity disk drives rather than settle for the standard models that a vendor may have overstocked. Expansion slots serve a central role in the expandability of a computer. Under certain conditions, expandability can avert obsolescence. For example, many owners of the eight-bit IBM personal computer (with the Intel 8088 CPU—a 16-bit processor with an 8-bit bus) were concerned about obsolescence when new software packages were produced exclusively for the advanced 32-bit IBM computer (which used the Intel 80386—a 32-bit processor). Since the eight-bit machines had expansion slots and several manufacturers produced special processor expansion cards (which had Intel 80386 chips on them), users who purchased and installed those cards were able to run the new software without having to discard their eight-bit machines and pay for new 32-bit machines.

It is important to understand computer architecture concepts to produce a feasible system (possesses no incompatibilities among the components), is optimal in cost (produces no over-expenditure on components whose extra capabilities are wasted because of bottlenecks or features ignored by other components), and delivers the highest performance. For example, if a user purchases an inexpensive, low-resolution graphics card and installs a high-resolution, expensive monitor, then the added cost of that higher-resolution monitor is wasted; the video signals will never utilize the higher-resolution modes.

In terms of performance, assume that a user were to buy both a high-resolution monitor and a high-resolution graphics card. If that graphics card does not have a capable GPU and memory subsystem for the monitor’s resolution and workload, then performance will decrease. This decrease occurs because the higher resolution requires more data pixels to be processed.

Without a coprocessor, the CPU will do all the work, resulting in a much slower system.

On a computer system that extensively utilizes disks and whose disk drive is continually active, one solution would be to increase the RAM and use software caching. As data enters the cache from disk, all subsequent accesses of that data are from memory at memory speed, which results in higher throughput and also less wear on disk drives, often the first components to fail. Adding memory may require inserting memory chips or adding a memory expansion card into an expansion slot.

The nature of interrupts, ROM, and bus type has an important role in the expansion of systems. Each type of computer operates with a different bus, and any expansion card for the wrong bus will not fit in the slots. Input/output and external storage device cards must be compatible with the ROM support chips for the CPU. For example, several devices (such as high-density floppy drives) will not operate on many computers compatible with the IBM personal computer, even though the bus slots accept those cards, because the ROM’s stored I/O programs do not support those devices. Even if the card is compatible with the computer, two different expansion cards may conflict in several ways; most importantly, both cards may use the same interrupts, resulting in the CPU passing control to the wrong program or the wrong device (often causing a system to “freeze-up”). To resolve these problems, expansion cards may have switches or metallic jumpers that can be set so that no two cards trigger the same interrupt.

Context

Historically, the creation of the first electronic computer is still an issue of debate because several contributors conceived and developed machines of varying degrees of programmability and computational capabilities. John V. Atanasoff’s work at Iowa State College in the 1930s indirectly contributed to the work of John William Mauchley and John Presper Eckert at the University of Pennsylvania’s Moore School. Mauchley and Eckert constructed ENIAC (Electronic Numerical Integrator and Computer), which is recognized as one of the earliest programmable computers. In 1949, they completed EDVAC (Electronic Discrete Variable Automatic Computer), a computer that stores programs in memory. A team at the University of Cambridge concurrently developed a similar machine based on their principles, called EDSAC (Electronic Delay Storage Automatic Calculator). Although John von Neumann contributed to the EDVAC effort, the concept of storing programs in memory is believed to be incorrectly attributed to him. To further complicate matters, Konrad Zuse worked in Europe on numerous prototype computers with several innovations in the 1930s and 1940s, but his work went unnoticed. In 1951, Eckert and Mauchley built the UNIVAC (Universal Automatic Computer), the first commercial computer, followed by International Business Machines Corporation (IBM) in 1953.

In the earliest computers, relays were used to represent gates. These were replaced by vacuum tubes. Since vacuum tubes were bulky and produced heat, they often burned out, requiring continual replacement by a maintenance staff. Vacuum tubes were replaced by transistor technology. The discovery of transistors represented a major breakthrough in miniaturization and paved the way for successive reductions in circuit size.

During the early 1960s, small-scale integration (SSI) produced integrated circuits with a compression of ten to twelve gates per silicon chip, with leads providing the inputs and outputs of those gates. These were further compressed in “medium-scale integration” (MSI) with up to ninety-nine gates, and beyond one hundred gates with large-scale integration (LSI).

Very large-scale integration (VLSI) achieved compression levels of more than one million transistors (on the Intel 80486 CPU chip) by the late 1980s. Memory chips followed similar reductions. Whereas many minicomputers during the 1970s had 64-256 kilobytes of RAM and mainframes had 1-16 megabytes, many personal computers during the late 1980s supported several megabytes of RAM. Besides smaller circuits, lower assembly costs of circuit boards, greater reliability, and less heat generation are characteristic of this progress. Mainframes and supercomputers could have greater computational abilities on the high end. With the reduced costs, size, and cooling requirements, affordable personal computers and workstations became a reality. Models of a whole computer “on a single chip” were developed.

By the 1990s and early 2000s, scientists continued integrating more transistors onto a single chip. Ultra-large-scale integration (ULSI) allowed billions of transistors to exist on modern processors, which supported the development of multi-core CPUs, in which a single processor chip contains two, four, or even dozens of independent processing units, dramatically improving parallel processing. Memory capacity also improved. Personal computers in the 2000s and 2010s featured several gigabytes of RAM and high-performance systems utilizing terabytes of memory. System-on-a-chip (SoC) architecture revolutionized computing by combining the CPU, GPU, memory controller, and other critical components onto a single chip; compact, power-efficient devices like smartphones and tablets quickly became ubiquitous in daily life.

In the 2020s, computer architecture was concerned with three-dimensional chip stacking, chiplet-based designs, heterogeneous computing architectures, and quantum computing. Developers also emphasized energy efficiency, specialized accelerators, and high-bandwidth interconnects while aiming to meet the growing demands of evolving technology like artificial intelligence and cloud computing. By 2024, computer architecture shifted toward chiplet-based designs and 3D packaging to enhance bandwidth and scalability. This trend continued into 2025 with the release of UCIe 3.0, standardizing chiplet interconnects to improve interoperability across multi-die systems.

Principal terms

BIT: the amount of memory required to store one binary digit

BYTE: a group of bits (often eight) that are processed as a unit

CHIP: circuitry formed on a silicon surface that is embedded in plastic and that has metal leads for connecting to external circuitry

GATE: the circuitry required to compute a single logical operation, such as the OR operation, over two bits

HARDWARE: any physical component of a computer such as circuitry, hard disks, cables, and monitor

PROGRAM: a sequence of ordered instructions that describes a procedure for the computer to perform

Bibliography

Baer, Jean-Loup. Computer Systems Architecture. Computer Science Press, 1980.

BasuMallick, Chiradeep. “Computer Architecture: Components, Types and Examples.” Spiceworks, 11 May 2023, www.spiceworks.com/tech/tech-general/articles/what-is-computer-architecture. Accessed 17 Apr. 2026.

“Blackwell Architecture.” NVIDIA, www.nvidia.com/enus/datacenter/technologies/blackwell-architecture/. Accessed 17 Apr. 2026.

Chattopadhyay, Anupam. Handbook of Computer Architecture. Springer, 2025.

“Computer Organization and Architecture Tutorial.” GeeksforGeeks, 23 Jan. 2025, www.geeksforgeeks.org/computer-organization-and-architecture-tutorials. Accessed 17 Apr. 2026.

Graham, Neill. The Mind Tool. West, 1986.

Greenfield, S. E. The Architecture of Microcomputers. Winthrop, 1980.

Hayes, John P. Computer Architecture and Organization. McGraw-Hill, 1988.

“HBM3E.” SK hynix, product.skhynix.com/products/dram/hbm/hbm3e.go. Accessed 17 Apr. 2026.

“Intel 8088 16-Bit Microprocessor.” Smithsonian Institution, www.si.edu/object/intel-8088-16-bit-microprocessor%3Anmah_713506. Accessed 17 Apr. 2026.

Ledin, Jim, and Dave Farley. Modern Computer Architecture and Organization: Learn X86, ARM, and RISC-V Architectures and the Design of Smartphones, PCs, and Cloud Servers. 2nd ed., Packt Publishing, 2022.

Norton, Peter. Inside the IBM PC. Brady, 1986.

“Press Releases.” UCIe Consortium, www.uciexpress.org/press-releases. Accessed 17 Apr. 2026.

Slotnick, Daniel L., et al. Computers and Applications. D. C. Heath, 1986.

“3DFabric Technology.” TSMC, 3dfabric.tsmc.com/english/dedicatedFoundry/technology/3DFabric.htm. Accessed 17 Apr. 2026.