Hubbry Logo
Semiconductor memorySemiconductor memoryMain
Open search
Semiconductor memory
Community hub
Semiconductor memory
logo
7 pages, 0 posts
0 subscribers
Be the first to start a discussion here.
Be the first to start a discussion here.
Contribute something
Semiconductor memory
Semiconductor memory
from Wikipedia

Semiconductor memory is a digital electronic semiconductor device used for digital data storage, such as computer memory. It typically refers to devices in which data is stored within metal–oxide–semiconductor (MOS) memory cells on a silicon integrated circuit memory chip.[1][2][3] There are numerous different types using different semiconductor technologies. The two main types of random-access memory (RAM) are static RAM (SRAM), which uses several transistors per memory cell, and dynamic RAM (DRAM), which uses a transistor and a MOS capacitor per cell. Non-volatile memory (such as EPROM, EEPROM and flash memory) uses floating-gate memory cells, which consist of a single floating-gate transistor per cell.

Most types of semiconductor memory have the property of random access,[4] which means that it takes the same amount of time to access any memory location, so data can be efficiently accessed in any random order.[5] This contrasts with data storage media such as CDs which read and write data consecutively and therefore the data can only be accessed in the same sequence it was written. Semiconductor memory also has much faster access times than other types of data storage; a byte of data can be written to or read from semiconductor memory within a few nanoseconds, while access time for rotating storage such as hard disks is in the range of milliseconds. For these reasons it is used for primary storage, to hold the program and data the computer is currently working on, among other uses.

As of 2017, sales of semiconductor memory chips are $124 billion annually, accounting for 30% of the semiconductor industry.[6] Shift registers, processor registers, data buffers and other small digital registers that have no memory address decoding mechanism are typically not referred to as memory although they also store digital data.

Description

[edit]

In a semiconductor memory chip, each bit of binary data is stored in a tiny circuit called a memory cell consisting of one to several transistors. The memory cells are laid out in rectangular arrays on the surface of the chip. The 1-bit memory cells are grouped in small units called words which are accessed together as a single memory address. Memory is manufactured in word length that is usually a power of two, typically N=1, 2, 4 or 8 bits.

Data is accessed by means of a binary number called a memory address applied to the chip's address pins, which specifies which word in the chip is to be accessed. If the memory address consists of M bits, the number of addresses on the chip is 2M, each containing an N bit word. Consequently, the amount of data stored in each chip is N2M bits.[5] The memory storage capacity for M number of address lines is given by 2M, which is usually in power of two: 2, 4, 8, 16, 32, 64, 128, 256 and 512 and measured in kilobits, megabits, gigabits or terabits, etc. As of 2014 the largest semiconductor memory chips hold a few gigabits of data, but higher capacity memory is constantly being developed. By combining several integrated circuits, memory can be arranged into a larger word length and/or address space than what is offered by each chip, often but not necessarily a power of two.[5]

The two basic operations performed by a memory chip are "read", in which the data contents of a memory word is read out (nondestructively), and "write" in which data is stored in a memory word, replacing any data that was previously stored there. To increase data rate, in some of the latest types of memory chips such as DDR SDRAM multiple words are accessed with each read or write operation.

In addition to standalone memory chips, blocks of semiconductor memory are integral parts of many computer and data processing integrated circuits. For example, the microprocessor chips that run computers contain cache memory to store instructions awaiting execution.

Types

[edit]

Volatile memory

[edit]
RAM chips for computers usually come on removable memory modules like these. Additional memory can be added to the computer by plugging in additional modules.

Volatile memory loses its stored data when the power to the memory chip is turned off. However it can be faster and less expensive than non-volatile memory. This type is used for the main memory in most computers, since data is stored on the hard disk while the computer is off. Major types are:[7][8]

RAM (Random-access memory) – This has become a generic term for any semiconductor memory that can be written to, as well as read from, in contrast to ROM (below), which can only be read. All semiconductor memory, not just RAM, has the property of random access.

  • DRAM (Dynamic random-access memory) – This uses memory cells consisting of one MOSFET (MOS field-effect transistor) and one MOS capacitor to store each bit. This type of RAM is the cheapest and highest in density, so it is used for the main memory in computers. However, the electric charge that stores the data in the memory cells slowly leaks out, so the memory cells must be periodically refreshed (rewritten) which requires additional circuitry. The refresh process is handled internally by the computer and is transparent to its user.
    • FPM DRAM (Fast page mode DRAM) – An older type of asynchronous DRAM that improved on previous types by allowing repeated accesses to a single "page" of memory to occur at a faster rate. Used in the mid-1990s.
    • EDO DRAM (Extended data out DRAM) – An older type of asynchronous DRAM which had faster access time than earlier types by being able to initiate a new memory access while data from the previous access was still being transferred. Used in the later part of the 1990s.
    • VRAM (Video random access memory) – An older type of dual-ported memory once used for the frame buffers of video adapters (video cards).
    • SDRAM (Synchronous dynamic random-access memory) – This added circuitry to the DRAM chip which synchronizes all operations with a clock signal added to the computer's memory bus. This allowed the chip to process multiple memory requests simultaneously using pipelining, to increase the speed. The data on the chip is also divided into banks which can each work on a memory operation simultaneously. This became the dominant type of computer memory by about the year 2000.
      • DDR SDRAM (Double data rate SDRAM) – This could transfer twice the data (two consecutive words) on each clock cycle by double pumping (transferring data on both the rising and falling edges of the clock pulse). Extensions of this idea are the current (2012) technique being used to increase memory access rate and throughput. Since it is proving difficult to further increase the internal clock speed of memory chips, these chips increase the transfer rate by transferring more data words on each clock cycle
        • DDR2 SDRAM – Transfers 4 consecutive words per internal clock cycle
        • DDR3 SDRAM – Transfers 8 consecutive words per internal clock cycle.
        • DDR4 SDRAM – Transfers 16 consecutive words per internal clock cycle.
      • RDRAM (Rambus DRAM) – An alternate double data rate memory standard that was used on some Intel systems but ultimately lost out to DDR SDRAM.
      • SGRAM (Synchronous graphics RAM) – A specialized type of SDRAM made for graphics adaptors (video cards). It can perform graphics-related operations such as bit masking and block write, and can open two pages of memory at once.
      • HBM (High Bandwidth Memory) – A development of SDRAM used in graphics cards that can transfer data at a faster rate. It consists of multiple memory chips stacked on top of one another, with a wider data bus.
    • PSRAM (Pseudostatic RAM) – This is DRAM which has circuitry to perform memory refresh on the chip, so that it acts like SRAM, allowing the external memory controller to be shut down to save energy. It is used in a few game consoles such as the Wii.
  • SRAM (Static random-access memory) – This stores each bit of data in a circuit called a flip-flop, made of 4 to 6 transistors. SRAM is less dense and more expensive per bit than DRAM, but faster and does not require memory refresh. It is used for smaller cache memories in computers.
  • CAM (Content-addressable memory) – This is a specialized type in which, instead of accessing data using an address, a data word is applied and the memory returns the location if the word is stored in the memory. It is mostly incorporated in other chips such as microprocessors where it is used for cache memory.

Non-volatile memory

[edit]

Non-volatile memory (NVM) preserves the data stored in it during periods when the power to the chip is turned off. Therefore, it is used for the memory in portable devices, which don't have disks, and for removable memory cards among other uses. Major types are:[7][8]

  • ROM (Read-only memory) – This is designed to hold permanent data, and in normal operation is only read from, not written to. Although many types can be written to, the writing process is slow and usually all the data in the chip must be rewritten at once. It is usually used to store system software which must be immediately accessible to the computer, such as the BIOS program which starts the computer, and the software (microcode) for portable devices and embedded computers such as microcontrollers.
    • MROM (Mask programmed ROM or Mask ROM) – In this type the data is programmed into the chip when the chip is manufactured, so it is only used for large production runs. It cannot be rewritten with new data.
    • PROM (Programmable read-only memory) – In this type the data is written into an existing PROM chip before it is installed in the circuit, but it can only be written once. The data is written by plugging the chip into a device called a PROM programmer.
    • EPROM (Erasable programmable read-only memory or UVEPROM) – In this type the data in it can be rewritten by removing the chip from the circuit board, exposing it to an ultraviolet light to erase the existing data, and plugging it into a PROM programmer. The IC package has a small transparent "window" in the top to admit the UV light. It is often used for prototypes and small production run devices, where the program in it may have to be changed at the factory.
      4M EPROM, showing transparent window used to erase the chip
    • EEPROM (Electrically erasable programmable read-only memory) – In this type the data can be rewritten electrically, while the chip is on the circuit board, but the writing process is slow. This type is used to hold firmware, the low level microcode which runs hardware devices, such as the BIOS program in most computers, so that it can be updated.
  • NVRAM (Non-volatile random-access memory)
  • Flash memory – In this type the writing process is intermediate in speed between EEPROMS and RAM memory; it can be written to, but not fast enough to serve as main memory. It is often used as a semiconductor version of a hard disk, to store files. It is used in portable devices such as PDAs, USB flash drives, and removable memory cards used in digital cameras and cellphones.

History

[edit]

Early computer memory consisted of magnetic-core memory, as early solid-state electronic semiconductors, including transistors such as the bipolar junction transistor (BJT), were impractical for use as digital storage elements (memory cells). The earliest semiconductor memory dates back to the early 1960s, with bipolar memory, which used bipolar transistors.[9] Bipolar semiconductor memory made from discrete devices was first shipped by Texas Instruments to the United States Air Force in 1961. The same year, the concept of solid-state memory on an integrated circuit (IC) chip was proposed by applications engineer Bob Norman at Fairchild Semiconductor.[10] The first single-chip memory IC was the BJT 16-bit IBM SP95 fabricated in December 1965, engineered by Paul Castrucci.[9][10] While bipolar memory offered improved performance over magnetic-core memory, it could not compete with the lower price of magnetic-core memory, which remained dominant up until the late 1960s.[9] Bipolar memory failed to replace magnetic-core memory because bipolar flip-flop circuits were too large and expensive.[11]

MOS memory

[edit]

The advent of the metal–oxide–semiconductor field-effect transistor (MOSFET),[12] invented by Mohamed M. Atalla and Dawon Kahng at Bell Labs in 1959,[13] enabled the practical use of metal–oxide–semiconductor (MOS) transistors as memory cell storage elements, a function previously served by magnetic cores in computer memory.[12] MOS memory was developed by John Schmidt at Fairchild Semiconductor in 1964.[14][15] In addition to higher performance, MOS memory was cheaper and consumed less power than magnetic-core memory.[14] This led to MOSFETs eventually replacing magnetic cores as the standard storage elements in computer memory.[12]

In 1965, J. Wood and R. Ball of the Royal Radar Establishment proposed digital storage systems that use CMOS (complementary MOS) memory cells, in addition to MOSFET power devices for the power supply, switched cross-coupling, switches and delay-line storage.[16] The development of silicon-gate MOS integrated circuit (MOS IC) technology by Federico Faggin at Fairchild in 1968 enabled the production of MOS memory chips.[17] NMOS memory was commercialized by IBM in the early 1970s.[18] MOS memory overtook magnetic core memory as the dominant memory technology in the early 1970s.[14]

The term "memory" when used with reference to computers most often refers to volatile random-access memory (RAM). The two main types of volatile RAM are static random-access memory (SRAM) and dynamic random-access memory (DRAM). Bipolar SRAM was invented by Robert Norman at Fairchild Semiconductor in 1963,[9] followed by the development of MOS SRAM by John Schmidt at Fairchild in 1964.[14] SRAM became an alternative to magnetic-core memory, but required six MOS transistors for each bit of data.[19] Commercial use of SRAM began in 1965, when IBM introduced their SP95 SRAM chip for the System/360 Model 95.[9]

Toshiba introduced bipolar DRAM memory cells for its Toscal BC-1411 electronic calculator in 1965.[20][21] While it offered improved performance over magnetic-core memory, bipolar DRAM could not compete with the lower price of the then dominant magnetic-core memory.[22] MOS technology is the basis for modern DRAM. In 1966, Dr. Robert H. Dennard at the IBM Thomas J. Watson Research Center was working on MOS memory. While examining the characteristics of MOS technology, he found it was capable of building capacitors, and that storing a charge or no charge on the MOS capacitor could represent the 1 and 0 of a bit, while the MOS transistor could control writing the charge to the capacitor. This led to his development of a single-transistor DRAM memory cell.[19] In 1967, Dennard filed a patent under IBM for a single-transistor DRAM memory cell, based on MOS technology.[23] This led to the first commercial DRAM IC chip, the Intel 1103, in October 1970.[24][25][26] Synchronous dynamic random-access memory (SDRAM) later debuted with the Samsung KM48SL2000 chip in 1992.[27][28]

The term "memory" is also often used to refer to non-volatile memory, specifically flash memory. It has origins in read-only memory (ROM). Programmable read-only memory (PROM) was invented by Wen Tsing Chow in 1956, while working for the Arma Division of the American Bosch Arma Corporation.[29][30] In 1967, Dawon Kahng and Simon Sze of Bell Labs proposed that the floating gate of a MOS semiconductor device could be used for the cell of a reprogrammable read-only memory (ROM), which led to Dov Frohman of Intel inventing EPROM (erasable PROM) in 1971.[31] EEPROM (electrically erasable PROM) was developed by Yasuo Tarui, Yutaka Hayashi and Kiyoko Naga at Japan's Ministry of International Trade and Industry (MITI) Electrotechnical Laboratory in 1972.[32] Flash memory was invented by Fujio Masuoka at Toshiba in the early 1980s.[33][34] Masuoka and colleagues presented the invention of NOR flash in 1984,[35] and then NAND flash in 1987.[36] Toshiba commercialized NAND flash memory in 1987.[37][38]

Applications

[edit]
MOS memory applications
MOS memory type Abbr. MOS memory cell Applications
Static random-access memory SRAM MOSFETs Cache memory, cell phones, eSRAM, mainframes, multimedia computers, networking, personal computers, servers, supercomputers, telecommunications, workstations,[39] DVD disk buffer,[40] data buffer,[41] nonvolatile BIOS memory
Dynamic random-access memory DRAM MOSFET, MOS capacitor Camcorders, embedded logic, eDRAM, graphics card, hard disk drive (HDD), networks, personal computers, personal digital assistants, printers,[39] main computer memory, desktop computers, servers, solid-state drives, video memory,[40] framebuffer memory[42][43]
Ferroelectric random-access memory FRAM MOSFET, Ferroelectric capacitor Non-volatile memory, radio-frequency identification (RF identification), smart cards[39][40]
Read-only memory ROM MOSFET Character generators, electronic musical instruments, laser printer fonts, video game ROM cartridges, word processor dictionary data[39][40]
Erasable programmable read-only memory EPROM Floating-gate MOSFET CD-ROM drives, embedded memory, code storage, modems[39][40]
Electrically erasable programmable read-only memory EEPROM Floating-gate MOSFET Anti-lock braking systems, air bags, car radios, cell phones, consumer electronics, cordless telephones, disk drives, embedded memory, flight controllers, military technology, modems, pagers, printers, set-top box, smart cards[39][40]
Flash memory Flash Floating-gate MOSFET ATA controllers, battery-powered applications, telecommunications, code storage, digital cameras, MP3 players, portable media players, BIOS memory,[39] USB flash drive,[44] digital TV, e-books, memory cards, mobile devices, set-top box, smartphones, solid-state drives, tablet computers[40]
Non-volatile random-access memory NVRAM Floating-gate MOSFETs Medical equipment, spacecraft[39][40]

See also

[edit]

References

[edit]
Revisions and contributorsEdit on WikipediaRead on Wikipedia
from Grokipedia
Semiconductor memory refers to electronic data storage devices constructed using semiconductor-based integrated circuits, primarily silicon, that enable random access to digital data bits stored as electrical charges or states. These devices are broadly classified into two categories: volatile memory, which loses stored data upon removal of power supply (such as dynamic random-access memory or DRAM and static random-access memory or SRAM), and non-volatile memory, which retains data without continuous power (such as read-only memory or ROM, electrically programmable ROM or EPROM, electrically erasable programmable ROM or EEPROM, and flash memory). The development of semiconductor memory began in the late , replacing earlier systems that dominated . Key milestones include the invention of the one-transistor DRAM cell by Robert Dennard at in 1967, which dramatically reduced cell size and cost compared to prior multi-transistor designs, and the commercial release of Intel's 1103 DRAM chip in 1970, the first single-chip DRAM with 1,024 bits of storage. Non-volatile advancements followed, with the floating-gate concept proposed by and at in 1967, leading to Intel's NOR flash in 1988 and Toshiba's NAND flash in 1989, enabling high-density data retention for portable devices. By the 1990s, CMOS scaling under propelled semiconductor memory into widespread use, with DRAM densities increasing exponentially to support personal . Today, semiconductor memory underpins modern electronics, serving as main memory (DRAM) for fast data access in processors, cache memory (SRAM) for high-speed operations, and persistent storage (flash) in smartphones, solid-state drives, and embedded systems. Volatile types like DRAM offer high density and low cost but require periodic refreshing to maintain data, while SRAM provides faster access without refresh at the expense of larger cell sizes. Non-volatile , particularly NAND variants, achieves terabyte-scale capacities through 3D stacking but faces scaling challenges below 10 nm, spurring research into emerging technologies like resistive RAM (RRAM) and phase-change memory (PCM) for improved endurance and speed. As of 2025, the global market for these technologies exceeds $190 billion annually, driven by demands in AI, data centers, and consumer gadgets.

Overview

Definition and Characteristics

Semiconductor memory refers to a form of electronic data storage that utilizes semiconductor devices, such as transistors and diodes, to store bits of information as electrical charges, currents, or other physical states within integrated circuits. This enables the representation of —0s and 1s—through the presence or absence of charge or a specific electrical state in individual memory cells, allowing for efficient digital information handling without mechanical components. Key characteristics of semiconductor memory include high storage , which permits billions of bits to be packed into compact chips; low power consumption; and fast access times, typically on the order of nanoseconds for read and write operations. It also demonstrates excellent , with fabrication advances enabling exponential increases in capacity over time, such as density improvements exceeding 2,000,000 times in recent decades. In distinction from magnetic or methods, which depend on and mechanical processes, semiconductor memory relies on for rapid, reliable data access and retention. Compared to earlier storage technologies like vacuum tubes or , semiconductor memory offers substantial advantages in size, being far more compact with cell areas as small as around 1300 square nanometers; speed, providing access times 10 to 100 times faster; and reliability, eliminating mechanical wear and vacuum-related failures through all-electronic operation. Semiconductor memories may exhibit volatile behavior, losing stored data without power, or non-volatile behavior, retaining data indefinitely, based on their design principles.

Role in Electronics

Semiconductor memory serves as the primary storage for instructions and data in central processing units (CPUs), forming the foundational memory unit essential to the that underpins modern computing systems. This architecture relies on , implemented through semiconductor technologies like DRAM and SRAM, to enable the shared storage and sequential fetching of programs and data, facilitating efficient computation. In computer systems, semiconductor memory occupies a critical position in the , positioned between high-speed CPU registers and slower secondary storage devices such as hard drives or SSDs, to optimize the balance between access speed, capacity, and cost. Caches, built from SRAM, provide rapid access for frequently used data near the processor, while main memory using DRAM offers larger capacities for active workloads, bridging the performance gap with bulk storage. This layered structure mitigates the von Neumann bottleneck by minimizing data movement latency. The integration of semiconductor memory into integrated circuits (ICs) has profoundly driven device , enabling the development of compact, power-efficient portable such as smartphones and wearables. By embedding memory directly onto chips, scaling transistor densities has reduced overall system size while enhancing performance and battery life in mobile applications. Semiconductor memory has been instrumental in sustaining , with DRAM densities doubling approximately every two years through continuous process scaling, thereby fueling exponential growth in computing capabilities. Economically, the global memory integrated circuits market is projected to reach $187.50 billion in revenue by , underscoring its pivotal role in the semiconductor industry's expansion and technological advancement. The semiconductor memory industry is known for its cyclical nature, featuring alternating periods of boom and bust driven by mismatches between supply and demand. Key factors include supply overabundance, which can lead to significant price declines when fabrication capacity expansions coincide; geopolitical tensions, such as U.S. tariffs and export controls on advanced chips; intensified competition among dominant players like Samsung and SK Hynix, who influence production shifts and market pricing; and macroeconomic fluctuations that exacerbate inventory cycles through effects like the bullwhip phenomenon, where small demand changes amplify upstream order variations.

Basic Principles

Semiconductor Physics

Semiconductor materials form the foundation of memory devices due to their tunable electrical properties, which lie between those of conductors and insulators. (Si) is the most widely used elemental , characterized by a crystal structure, an indirect of approximately 1.12 eV at , and high abundance, making it ideal for large-scale integrated circuits. (Ge), another elemental with a similar structure, has a smaller of about 0.66 eV, enabling higher carrier mobilities but suffering from higher leakage currents at elevated temperatures. Compound semiconductors like (GaAs) offer a direct of 1.42 eV, superior (around 8500 cm²/V·s compared to silicon's 1400 cm²/V·s), and are employed in high-speed applications, though their higher cost limits widespread use in memory. Doping introduces controlled impurities into the semiconductor lattice to alter its electrical conductivity, creating n-type or p-type materials essential for memory device functionality. In n-type doping, pentavalent impurities such as phosphorus (P) or arsenic (As) are added to silicon, donating an extra electron to the conduction band while the impurity ion remains positively charged, increasing electron concentration. Conversely, p-type doping incorporates trivalent impurities like boron (B) or aluminum (Al), which accept an electron from the valence band, creating mobile holes as the primary charge carriers and leaving a negatively charged impurity ion. The dopant concentration typically ranges from 10¹⁴ to 10¹⁸ atoms/cm³, far below the host lattice density of about 10²² atoms/cm³, ensuring minimal disruption to the crystal structure while significantly enhancing conductivity. Charge carrier dynamics in semiconductors govern the transport and storage mechanisms in memory cells, involving electrons in the conduction band and holes in the valence band separated by the band gap energy (E_g). Electrons, negatively charged quasiparticles, contribute to conduction when excited across the band gap by thermal energy or electric fields, while holes behave as positively charged vacancies that facilitate current flow in the opposite direction. Conductivity (σ) arises from the drift of these carriers under an applied electric field (E), expressed as σ = q (n μ_n + p μ_p), where q is the elementary charge, n and p are electron and hole concentrations, and μ_n and μ_p are their respective mobilities; in intrinsic semiconductors, n = p ≈ 10¹⁰ cm⁻³ at room temperature for silicon, but doping shifts this balance dramatically. Thermal generation of carriers follows the Fermi-Dirac distribution, which describes the probability of occupancy for energy states; for non-degenerate semiconductors where the Fermi level (E_f) is several kT below the conduction band edge (E_c), the electron concentration simplifies to the Boltzmann approximation: n=Ncexp(EcEfkT)n = N_c \exp\left( -\frac{E_c - E_f}{kT} \right) where N_c is the effective density of states in the conduction band (≈ 2.8 × 10¹⁹ cm⁻³ for silicon at 300 K), k is Boltzmann's constant, and T is temperature in Kelvin—this equation highlights how thermal energy (kT ≈ 0.026 eV at room temperature) enables carrier excitation across the band gap. PN junctions, formed by joining n-type and p-type regions, underpin transistor-based elements through their rectifying behavior and depletion regions. Upon junction formation, electrons from the n-side diffuse to the p-side and recombine with holes, creating a space-charge region depleted of free carriers and an built-in (≈ 0.7 V for ) that opposes further , with the depletion width (W) scaling as W ∝ √(1/N_a + 1/N_d), where N_a and N_d are acceptor and donor concentrations. Under forward bias (positive voltage on p-side), the barrier reduces, allowing majority carriers to inject across the junction and increase current exponentially (I ∝ exp(qV/kT)). In reverse bias, the depletion widens, suppressing majority carrier flow and limiting current to thermally generated minority carriers, forming the basis for isolation and switching in memory transistors.

Memory Cell Operation

Semiconductor memory cells typically consist of basic components such as transistors, capacitors, and diodes, which enable the storage and manipulation of . Transistors, particularly , function as switches to control access to the storage element, while store charge representing the data state. Diodes, often used in configurations, provide unidirectional conduction to define fixed data patterns at intersections of row and column lines. A representative example is the one-transistor-one-capacitor (1T-1C) structure, where a single connects a to bit lines, allowing compact storage of a bit as charge presence or absence on the capacitor. Read and write operations in these cells involve applying specific voltages to manipulate and sense charge levels. During a write operation, a voltage is applied to the 's gate to turn it on, charging or discharging the through the bit line to represent a logic '1' or '0'; in non-volatile cells, higher voltages enable programming by injecting charge onto a floating gate or erasing by removing it via tunneling effects. Reading entails activating the transistor to allow charge from the to flow onto the bit line, where sense amplifiers detect small voltage differences (typically around 200-300 mV) to determine the stored value, often destructively discharging the in charge-based designs. Charge-based cells require periodic refresh cycles to counteract leakage, where the sensed data is rewritten to restore the original charge level, typically every 64 milliseconds to prevent . Access to individual cells within an array is achieved through addressing mechanisms involving row and column decoders, word lines, and bit lines. Row decoders select a specific word line to activate the transistors in an entire row, while column decoders enable the corresponding bit lines to connect to sense amplifiers or data inputs for the targeted columns. Word lines run horizontally to gate the transistors, and bit lines run vertically to carry charge or signals, forming a grid that allows efficient random access to any cell by decoding the address into row and column signals. The fundamental storage in capacitor-based cells relies on the equation for charge Q=C×VQ = C \times V, where QQ is the stored charge, CC is the , and VV is the voltage across the , determining the logic state. Retention time in such cells is limited by leakage current, approximated as Ileak=Q/tretentionI_{\text{leak}} = Q / t_{\text{retention}}, where charge gradually dissipates through subthreshold leakage or junction currents, necessitating refresh to maintain .

Types

Volatile Memory

Volatile semiconductor memory refers to electronic data storage devices fabricated using semiconductor materials that retain information only while powered, losing all stored data upon power removal. This class of memory maintains data through active electrical states, such as stored charge in capacitors or stable current loops in transistor circuits, which dissipate without continuous supply voltage. Unlike non-volatile alternatives, volatile memory prioritizes speed and density for temporary data handling in computing systems. The primary subtype is (DRAM), which stores each bit as a charge in a paired with a single access per cell. Due to inevitable charge leakage through the 's , DRAM requires periodic refreshing—typically every 64 milliseconds—to recharge cells and prevent data loss, a managed by dedicated circuitry that reads and rewrites row data across the array. This dynamic nature enables high integration density but introduces overhead from refresh operations, consuming additional power and bandwidth. In contrast, static random-access memory (SRAM) employs a bistable flip-flop circuit, commonly implemented with six transistors (6T cell) forming two cross-coupled inverters connected to bit lines via access transistors, to hold data indefinitely without refresh as long as power is supplied. The stable feedback loop in the inverters ensures data retention through voltage levels rather than charge, allowing faster read/write operations but at the cost of larger cell size—typically 20-30 times that of DRAM cells—due to the multiple transistors required. Contemporary DRAM achieves densities up to 32 Gb per chip as of 2025, driven by advancements in process nodes and stacking techniques, with typical times around 10-20 ns reflecting the time to activate a row, sense the charge, and output data. SRAM, optimized for low latency, offers access times as low as 1 ns, making it suitable for high-speed applications, though its 6T cell structure limits on-chip densities to the megabit range per array. These subtypes exhibit key trade-offs: DRAM excels in cost-effective high-density storage for bulk main in systems like servers and PCs, where refresh overhead is tolerable for capacity gains, while SRAM provides superior speed and reliability without refresh for smaller, power-hungry caches and registers in processors. SRAM's higher static power from always-on transistors and larger footprint restrict its use to performance-critical areas, whereas DRAM's lower per-bit power suits expansive data buffers despite slower access.

Non-Volatile Memory

Non-volatile semiconductor memory retains stored data without an external , distinguishing it from volatile types by employing mechanisms that preserve through physical states such as trapped charges or properties. In charge-based non-volatile memories, data is stored by injecting electrons onto a floating gate—an isolated conductive layer within a structure—or into charge-trapping layers, where the trapped charges modulate the transistor's to represent binary states. These charges are introduced via , where high-energy electrons from the channel are accelerated into the gate under strong electric fields, or through Fowler-Nordheim tunneling, in which electrons quantum-mechanically tunnel through a thin insulating layer under high voltage. Erasure reverses this by tunneling electrons away, often to a substrate or control gate, enabling data persistence for years or decades due to the insulating barriers that prevent charge leakage. Among the primary subtypes, Mask ROM is factory-programmed during fabrication, where specific connections in the memory array are permanently defined using custom photomasks to hardwire the desired data pattern, making it ideal for high-volume, unchanging content like but non-reprogrammable in the field. EPROM extends this by allowing user programming via , with erasure achieved by exposing the chip to light through a window, which provides sufficient energy for charges to escape the floating gate en masse, typically requiring 10-20 minutes of exposure. EEPROM advances reprogrammability with electrical erasure and programming, using tunneling for both operations to selectively modify bytes or words without external light, though at the cost of more complex circuitry involving two transistors per cell for finer control. Flash memory, a block-oriented EEPROM variant, dominates modern storage; NOR Flash supports random byte access and is suited for code execution due to its page-level programming and sector-level erasure, while NAND Flash prioritizes density with block-level erasure (typically 128 KB to 2 MB units) and page-level programming (2-16 KB), enabling efficient sequential data handling in . Advances in NAND include 3D stacking, where memory cells are vertically layered up to 300 or more tiers using or channel-hole , achieving areal densities exceeding 10 Gb/mm² and capacities over 1 Tb per die by 2025. Performance characteristics of Flash memory highlight its trade-offs: typical endurance is around 10⁵ program/erase cycles per block due to oxide degradation from repeated charge injection and tunneling, beyond which retention and reliability diminish, though error correction mitigates this in practice. Read access times are approximately 10-50 μs, enabling rapid compared to mechanical alternatives, while write operations take 100-500 μs per page owing to the high voltages (10-20 V) needed for injection or tunneling. These attributes provide significant advantages in density over traditional hard disk drives (HDDs), with 3D NAND offering up to 10 times the bits per unit volume through vertical scaling, alongside shock resistance and lower power use, though HDDs retain edges in cost per bit for archival storage. Hybrid non-volatile variants expand beyond pure charge storage; FRAM (ferroelectric RAM) uses a ferroelectric integrated with a , where data is encoded in the material's remnant polarization states—remanent electric orientations that switch under applied voltage via motion, offering non-destructive reads, unlimited (>10¹² cycles), and SRAM-like speeds (10-100 ns) with densities up to 16 Mb in embedded applications as of 2025. (magnetoresistive RAM) employs magnetic tunnel junctions (MTJs), consisting of two ferromagnetic layers separated by a thin insulator (e.g., MgO), where the free layer's aligns parallel or antiparallel to a fixed layer, producing a tunable resistance via tunnel (up to 200%) for state detection; writing occurs through spin-transfer , injecting spin-polarized currents to flip the free layer, achieving high (>10¹⁰ cycles) and sub-10 ns access while scaling to advanced nodes such as 7 nm and below as of 2025.

History

Early Developments

Before the advent of semiconductor memory, early computers relied on bulkier and less efficient storage technologies such as and . , which used propagating through mercury-filled tubes or nickel wire to store data as serial bits, was employed in machines like the introduced in 1951 by , marking a significant step from vacuum tube-based systems like ENIVAC. By the mid-1950s, —consisting of small ferrite rings threaded with wires to represent binary states—emerged as the dominant form, offering and non-volatility; it powered systems like the MIT in 1953 and became standard in commercial computers throughout the decade, though it required manual assembly and occupied considerable space. The transition to semiconductor memory began in the mid-1960s with bipolar transistor-based designs, replacing discrete components with s for greater density and speed. In 1965, introduced the first memory in its System/360 Model 95 mainframe, a 16-bit () using (SLT) modules with bipolar transistors, which reduced memory size from cubic inches of core planes to compact chips while enabling faster access times around 2 microseconds. Experiments with thin-film magnetic memory, involving evaporated ferromagnetic alloys on substrates for higher speeds, were conducted around 1966 by researchers at and Sperry Rand, achieving cycle times as low as 200 nanoseconds in prototypes but facing challenges with yield and stability that limited commercial adoption. These efforts addressed key hurdles, shifting from hand-wired discrete transistors—each memory bit requiring multiple components—to monolithic ICs that integrated dozens of transistors per chip, slashing power consumption and volume by orders of magnitude. A pivotal advancement came in 1969 when released the , the first commercial metal-oxide-semiconductor (MOS) (SRAM) chip, storing 256 bits using silicon-gate PMOS technology and serving as a precursor to modern SRAM designs with access times of about 1000 nanoseconds. Concurrently, in 1967, researcher invented (DRAM), patented in 1968 as U.S. Patent 3,387,286, which employed a single and (1T-1C) per bit to achieve higher than bipolar alternatives; this innovation, prototyped by 1970 and commercially released as 's 1103 chip that year, overcame refresh and leakage issues in MOS capacitors, enabling scalable that displaced core systems in cost-sensitive applications. These early developments laid the foundation for memory's dominance by integrating physics—briefly, leveraging gate voltage to control charge flow—into practical, chip-scale storage.

MOS Technology Evolution

The Metal-Oxide-Semiconductor Field-Effect Transistor (), invented in 1959 by and Mohamed Atalla at Bell Laboratories, laid the foundation for modern semiconductor memory by enabling scalable, high-density integrated circuits through its insulated-gate structure that minimized leakage and improved control over charge flow. This breakthrough shifted memory design from earlier bipolar transistor-based approaches, which were power-intensive, toward more efficient MOS architectures. By 1971, Intel's 4004 demonstrated the first commercial application of to integrated memory, incorporating MOS silicon-gate transistors to achieve a 4-bit processor with embedded ROM for program storage on a single chip. In the 1980s, the industry transitioned from NMOS to Complementary MOS () technology, which paired n-type and p-type transistors to drastically reduce power consumption and heat generation, making it ideal for dense memory arrays in portable and battery-powered devices. This shift enabled the proliferation of CMOS-based (DRAM) and static RAM (SRAM), with production scaling to support early personal computers. Entering the , sub-micron process nodes—dropping below 1 micrometer to around 0.5 micrometers—facilitated exponential density increases, culminating in the commercialization of gigabit-scale DRAM by the late decade, such as Samsung's 1 Gb DRAM in 1996, which relied on advanced MOS scaling to pack billions of transistors. The 2010s introduced FinFET (Fin Field-Effect Transistor) structures, a 3D MOS transistor design that wrapped the gate around a fin-shaped channel to enhance electrostatic control and mitigate short-channel effects at nodes below 20 nm, significantly boosting memory performance and density in both logic-integrated and standalone applications. Key innovations further propelled MOS evolution: in 1987, Fujio Masuoka at Toshiba invented NAND flash memory, a non-volatile MOS-based technology using floating-gate transistors for block-erasable storage, which revolutionized data persistence in consumer electronics. By 2006, Samsung advanced stacking techniques with its Wide I/O Stack Package (WSP), vertically integrating eight 2 Gb NAND flash dies to achieve 16 Gbit capacity in a compact form, paving the way for higher-density modules. Companies like Samsung and Micron continued this trajectory, with Micron contributing to joint developments in scaled MOS processes for NAND and DRAM. Following , MOS node shrinks have sustained transistor density doublings roughly every two years, reaching 3 nm processes by the mid-2020s, where enhanced gate-all-around variants of FinFET maintain scaling benefits for memory applications. These advancements have positioned MOS technologies as the dominant force in the semiconductor memory market by 2025 through their versatility in volatile and non-volatile types, far surpassing legacy bipolar shares.

Applications

Computing and Storage Systems

Semiconductor memory plays a central role in modern architectures, where volatile memories such as SRAM and DRAM provide high-speed, temporary essential for processor performance. In central processing units (CPUs), SRAM is predominantly used for on-chip caches, including L1 and L2 levels, to minimize latency for frequently accessed instructions and . For instance, Intel's Lake desktop processors, such as the Core Ultra 9 285K released in 2024, feature up to 36 MB of L3 cache built with SRAM, enabling rapid retrieval that significantly boosts computational efficiency in multitasking and gaming workloads. Beyond caches, DRAM serves as the primary system memory (RAM) in personal computers, offering larger capacities for running applications and operating systems. High-end consumer PCs in 2025 commonly support up to 128 GB of DRAM, allowing seamless handling of memory-intensive tasks like and virtual machines. In storage hierarchies, non-volatile semiconductor memory, particularly NAND flash, has revolutionized data persistence by powering solid-state drives (SSDs) that serve as boot drives and primary storage in systems. SSDs leveraging NAND flash provide sequential read/write speeds exceeding 7,000 MB/s, offering approximately 10 times the performance of traditional hard disk drives (HDDs) with typical speeds of 150-250 MB/s, which has led to widespread replacement of HDDs in desktops, laptops, and servers for faster boot times and application loading. Hybrid storage systems, such as those formerly employing Intel's Optane technology based on memory, bridged the gap between DRAM's speed and NAND's capacity by acting as a fast cache layer, though production of Optane DIMMs concluded shipments by late 2025. System integration of semiconductor memory relies on specialized components to ensure reliable flow and integrity. Memory controllers, integrated into CPUs or chipsets, manage transfers between memory modules and processors, while high-speed buses like DDR5—established as the industry standard by 2025 with rates up to 8.4 GT/s—facilitate bandwidth-intensive operations in both consumer and server environments. Error-correcting code (ECC) mechanisms, commonly implemented in DRAM for servers, detect and correct single-bit errors to maintain accuracy in mission-critical applications. In performance terms, server systems achieve DRAM bandwidths exceeding 100 GB/s through multi-channel configurations, supporting AI workloads and , while NAND flash arrays in setups enhance redundancy and throughput in centers, reaching effective speeds over 10 GB/s per array.

Embedded and Specialized Uses

Semiconductor memory plays a critical role in embedded systems, where space, power, and reliability constraints demand efficient storage solutions. In microcontrollers like those used in boards, serves as the primary non-volatile storage for , enabling persistent code execution without frequent reprogramming. Similarly, is employed in automotive engine control units (ECUs) to store calibration data and updates, ensuring stable operation in harsh vehicular environments. For temporary data handling, low-power SRAM is integrated into wearables, such as fitness trackers, to provide fast access times and minimal energy consumption during real-time sensor processing. In (IoT) devices and mobile applications, NOR flash is favored for its ability to support direct code execution, which is essential for resource-constrained sensors monitoring environmental conditions. This architecture allows embedded processors to run directly from the memory without loading into RAM, reducing latency and power draw in battery-operated nodes. For higher-capacity needs in smartphones, embedded (eMMC) and (UFS) provide scalable solutions, with UFS 4.0 enabling up to 1TB of storage in compact form factors for multimedia and app data. Specialized environments require memory hardened against extreme conditions, such as in space applications. Radiation-hardened SRAM is widely used in satellites for its robustness against single-event upsets caused by cosmic rays, ensuring reliable and in . In secure applications like smart cards, magnetoresistive (MRAM) offers tamper-resistant storage due to its non-volatile magnetic state retention, which resists physical attacks and better than traditional flash. Industrial settings leverage one-time programmable (OTP) ROM for permanent configuration in appliances, such as washing machines, where it stores immutable to prevent unauthorized modifications and ensure long-term reliability. For oil drilling operations, high-temperature semiconductor memories, including specialized non-volatile variants, operate reliably up to 200°C in downhole tools, capturing real-time geological data without cooling systems. These solutions prioritize endurance in corrosive, high-pressure environments to support extended drilling missions.

Emerging Technologies

Novel Memory Types

Novel memory types in semiconductor technology encompass emerging non-volatile memory devices that leverage beyond-traditional mechanisms to achieve higher performance, density, and scalability compared to established options like NAND flash. These technologies, including resistive RAM (ReRAM), phase-change memory (PCM), and spin-transfer torque MRAM (STT-MRAM), are advancing through commercialization and advanced stages as of 2025, targeting applications in storage-class memory, embedded systems, and . They address limitations in speed, endurance, and power efficiency by utilizing distinct physical phenomena for data storage and retrieval, often integrated with processes for feasibility at advanced nodes. Resistive RAM (ReRAM) functions through the formation and rupture of conductive filaments driven by oxygen vacancy mechanisms in insulating oxides, such as HfO₂, where applied voltage modulates the migration of oxygen ions to create low- or high-resistance states. This enables bipolar resistive switching with a simple metal-insulator-metal structure, offering non-volatility and compatibility with back-end-of-line fabrication. ReRAM devices frequently employ crossbar arrays, where selector elements mitigate sneak currents, allowing for 3D stacking to enhance integration density; such configurations have demonstrated areal densities exceeding 10 Gb/mm² in multi-layer prototypes. Write speeds below 10 ns have been achieved in HfO₂-based cells, supporting ultra-fast operations suitable for cache or neuromorphic applications. Prototypes have scaled to 5 nm nodes, verifying compatibility with leading-edge semiconductor processes. Phase-change memory (PCM) stores data by exploiting reversible amorphous-to-crystalline phase transitions in chalcogenide materials, such as Ge₂Sb₂Te₅, induced by localized heating from electrical pulses to alter resistivity. The amorphous state represents the high-resistance reset condition, while the crystalline phase denotes the low-resistance set state, enabling multi-bit storage per cell. and Micron's technology, a PCM variant developed in the late , utilized a 3D cross-point architecture that achieved approximately 1000 times the write endurance of NAND flash, with over 10¹² cycles per cell versus NAND's typical 10⁴–10⁵; however, its commercial development was discontinued in 2021. This endurance stemmed from the byte-addressable, in-place overwrite capability without block erases, reducing wear. PCM integrates seamlessly with processes, as evidenced by 40 nm embedded implementations and lab-scale devices at 5 nm feature sizes with reset currents under 10 μA. As of 2025, ongoing research and potential commercialization efforts, such as those by Sandisk, continue to advance PCM for similar high-endurance applications. Spin-transfer torque MRAM (STT-MRAM) relies on magnetic junctions (MTJs), where is encoded in the relative orientation of ferromagnetic layers separated by a thin MgO barrier; spin-polarized current switches the free layer's for non-volatile storage. This mechanism provides unlimited and access times, positioning STT-MRAM as a for SRAM in last-level caches. Everspin Technologies commercialized 1 Gb STT-MRAM chips at 28 nm by 2023, featuring DDR4-compatible interfaces for enterprise applications, with ongoing efforts targeting densities up to 16 Gb through advanced MTJ stacking and process shrinks. These devices exhibit high thermal stability and low switching currents, enabling integration at 5 nm nodes in prototypes.

Challenges and Future Directions

One of the primary challenges in advancing semiconductor memory is scaling to sub-3 nm nodes, where quantum tunneling effects become dominant, leading to increased leakage currents that compromise device reliability and elevate power dissipation. As thicknesses approach or fall below 3 nm, direct tunneling of carriers through the insulator results in significant gate leakage, which can exceed acceptable limits for sustained operation in high-performance applications. This issue is exacerbated in memory cells, where thin barriers intended to store charge allow unintended flow, reducing retention times and overall . Additionally, power consumption in data centers poses a growing concern, with DRAM accounting for up to 40% of a server's total energy use, a trend projected to intensify by 2025 amid rising demands from AI and workloads. Reliability in high-density non-volatile memories, particularly NAND flash, is another critical hurdle, as scaling increases raw bit rates due to factors like charge trapping and read/write disturbances. In (MLC) and triple-level cell (TLC) architectures, bit errors can accumulate, potentially exceeding 10^{-3} per bit without mitigation, threatening in storage systems. To address this, error-correcting codes (ECC) such as low-density parity-check (LDPC) codes are employed to detect and correct multiple bit flips per page, while wear-leveling algorithms distribute program/erase cycles evenly across blocks to prevent premature endurance failure, extending device lifetimes beyond 10^5 cycles. Looking ahead, compute-in-memory paradigms are emerging as a key direction to overcome von Neumann bottlenecks in AI applications, with analog resistive RAM (ReRAM) accelerators enabling in-situ matrix-vector multiplications that reduce data movement overhead by orders of magnitude compared to traditional architectures. These systems perform computations directly within the array, achieving efficiencies up to 100x for tasks while maintaining compatibility with deep neural networks. Sustainability efforts focus on novel materials, such as 2D semiconductors like (MoS2), which enable low-power transistors and devices operating at sub-1 V thresholds, potentially cutting use by 50% in next-generation nodes through reduced leakage and faster switching. The pursuit of universal memory—a technology combining SRAM-like speed (sub-nanosecond access), DRAM-like density (gigabits per chip), and non-volatility for instant-on operation—remains a central goal, with candidates like ferroelectric and phase-change memories showing promise in bridging these attributes. Market projections indicate that emerging memory types, including ReRAM and (MRAM), could capture up to 20% of the overall semiconductor memory sector by 2030, driven by their role in edge AI and energy-efficient storage, as the total market expands to over $240 billion.

References

Add your contribution
Related Hubs
Contribute something
User Avatar
No comments yet.