Flash memory is a form of nonvolatile semiconductor memory that retains information without continuous electrical power and can be electrically erased and reprogrammed. Its defining organizational feature is that erasure normally affects a group of cells rather than an individually addressed byte. Flash combines compactness, mechanical robustness, and relatively low storage cost, making it a principal storage technology in smartphones, removable memory cards, embedded electronics, and solid-state drives (SSDs).
A flash cell stores information by maintaining electrical charge in an insulated region associated with a transistor. The charge changes the transistor’s electrical behavior, allowing a sensing circuit to distinguish stored states. NOR and NAND flash organize these cells differently: NOR favors direct access to code, whereas NAND favors dense storage of large quantities of data. Modern NAND devices commonly store several bits per cell and arrange cells in vertical stacks.
Flash is not an indefinitely rewritable substitute for working memory. Erasing and programming cause physical wear, stored charge gradually changes, and updates must respect page and block boundaries. Its practical behavior therefore depends on a combination of device physics, array architecture, error correction, and controller software. These layers explain both its exceptional usefulness and its limitations.
Origins and technological development
Flash developed from earlier forms of electrically programmable read-only memory. In 1967, Dawon Kahng and Simon Sze described the floating-gate transistor as a nonvolatile storage element. In this structure, an electrically isolated conductor can retain charge that alters a transistor’s threshold voltage. The principle made it possible to store information without the continuous power required by ordinary volatile semiconductor memory.
Early erasable programmable read-only memory (EPROM) devices demonstrated the usefulness of floating-gate storage but were inconvenient to update. Many packaged EPROMs had a quartz window through which ultraviolet light erased the chip. Updating a product could require removing the memory from its socket, erasing it in separate equipment, and reprogramming it. Electrically erasable programmable read-only memory (EEPROM) subsequently permitted electrical erasure and rewriting, commonly with fine-grained access. Its additional cell structures and selection requirements, however, imposed costs in density and chip area.
Fujio Masuoka and colleagues at Toshiba developed flash memory as an approach to electrically erasable storage with a more compact organization. Masuoka’s group presented a flash EEPROM technology at the 1984 IEEE International Electron Devices Meeting. The name “flash” referred to the rapid, collective erasure of memory cells. It did not mean that every operation occurred instantaneously or that all flash memories erased an entire chip at once. Later devices adopted different block and sector arrangements.
Two array architectures became commercially decisive. Intel introduced commercial NOR flash in 1988, while Toshiba commercialized NAND flash in 1989. NOR proved well suited to storing executable programs and embedded firmware. NAND reduced the area needed per stored bit by arranging cells in serial strings, eventually making large-capacity solid-state storage economically practical. These architectures developed into complementary product families rather than successive generations in which one simply displaced the other.
During the 1990s and 2000s, flash expanded through digital photography, portable music players, USB drives, and mobile phones. SSDs later brought NAND into mainstream computers and enterprise storage. Increasing density depended first on smaller planar cells and multilevel storage, then increasingly on three-dimensional integration. Samsung began commercial production of its V-NAND technology in 2013, an important milestone in the transition to vertically stacked NAND. Across this history, progress came not only from memory-cell fabrication but also from controllers, interfaces, and increasingly powerful error correction.
Charge storage and the operation of a memory cell
A conventional flash cell is a modified metal–oxide–semiconductor field-effect transistor. In an ordinary transistor, voltage applied to a gate controls conduction through a semiconductor channel. In a flash cell, stored charge shifts the threshold voltage, , at which substantial channel conduction begins. Reading the cell involves applying selected voltages and measuring whether, or how strongly, it conducts. The result identifies a range of threshold voltages associated with a stored digital state.
Two major charge-storage structures are used. A floating-gate transistor stores charge on an electrically isolated conducting gate, usually surrounded by insulating material. A charge-trap memory cell instead stores charge in localized traps within an insulating layer, commonly silicon nitride. Neither structure is a miniature battery connected to an external circuit. Its information depends on charge remaining sufficiently isolated to preserve distinguishable electrical states.
Programming and erasing move charge across an insulating barrier under strong electric fields. One important mechanism is Fowler–Nordheim tunneling, a quantum-mechanical process in which carriers cross a barrier that would block them in a purely classical description. Some flash designs also use hot-carrier injection, in which energetic carriers enter the storage region. The mechanisms and voltage polarities differ among device families; no single programming description applies to every NOR and NAND implementation.
Programming is normally controlled through a sequence of voltage pulses followed by verification reads. Incremental step-pulse programming increases the applied stress gradually until the cell falls within its intended threshold-voltage range. This avoids relying on a single pulse to produce exactly the required charge. Erasing moves cells toward a common erased state and generally acts on a block sharing the relevant electrical connections. On-chip charge pumps generate programming and erase voltages above the device’s normal external supply.
In a common NAND convention, an erased bit reads as logical 1 and programming moves it toward logical 0. This convention does not imply that multibit cells contain independently writable physical switches: several logical bits are encoded jointly by one cell’s threshold range. Reversing an arbitrary programmed pattern usually requires erasure first.
The insulating layers are excellent but imperfect. Repeated electrical stress creates defects, and charge can escape or redistribute over time. Consequently, a cell’s state is better understood as a measured electrical quantity with tolerances than as an immutable binary object. Device design must balance the ease of moving charge during updates against the difficulty of losing it during storage.
NOR and NAND array architectures
NOR and NAND take their names from the resemblance of their cell interconnections to the corresponding logic gates. The distinction concerns array organization, not the representation of application data. Both can store programs, images, or numerical records, but their wiring creates substantially different access patterns, density, and management requirements.
In NOR flash, cells have relatively direct connections to bit lines, permitting efficient random access to individual memory locations. Parallel NOR devices can expose an addressable memory interface, while serial NOR uses a narrower command-based interface. A processor may execute instructions directly from suitable NOR devices, a capability called execute in place. Serial interfaces can also support memory-mapped access through a controller, although their latency and bandwidth differ from those of a parallel memory bus.
This organization makes NOR valuable for boot code, control software, and systems in which predictable instruction access matters more than the lowest possible cost per bit. Its disadvantages include comparatively large cell-array area and generally lower storage density than NAND. NOR devices also have erase sectors, so direct reading should not be confused with unrestricted byte-by-byte rewriting. A processor can fetch a byte without first reading an entire NAND-style page, but changing that byte may still require sector-level operations.
In NAND flash, many memory cells are connected in series as a string. Selection transistors control access to each string, while shared word lines address cells at corresponding positions. During a read, the selected cell is tested while unselected cells receive pass voltages intended to keep them conductive. Sharing connections reduces overhead per cell, enabling much denser arrays. The same serial organization introduces dependencies among cells and makes direct, arbitrary byte access impractical.
NAND is typically read and programmed in pages and erased in blocks containing many pages. A page may contain several kibibytes or more of user data, together with additional storage for error-correction information and management data. Sizes vary by generation and product. A page read first transfers sensed data into internal registers; an interface then transfers the requested data to the host or controller.
The distinction also affects responsibility for reliability. Raw NAND commonly exposes bad-block handling and error-correction requirements to an external controller. Managed products incorporate those functions internally. NOR and NAND therefore cannot be compared solely by nominal capacity or read speed: a meaningful comparison includes access granularity, execution requirements, erase organization, controller complexity, and the conditions under which data integrity is specified.
Single-level and multilevel storage
A flash cell need not represent only two states. If its threshold-voltage range can be divided into sufficiently distinguishable regions, one cell can encode several bits. For stored bits per cell, the required number of states is
Single-level cell (SLC) storage uses two states and stores one bit per cell. Two-bit storage requires four states, triple-level cell (TLC) requires eight, and quad-level cell (QLC) requires sixteen. Although “multilevel cell” can describe any arrangement with more than one bit per cell, commercial usage often reserves MLC for two-bit cells.
The density advantage is substantial because additional states increase capacity without requiring a proportional increase in the number of physical cells. The trade-off is smaller separation between valid states. Manufacturing variation, electrical interference, charge loss, and accumulated wear all broaden or shift threshold-voltage distributions. As the number of states increases, the sensing and programming systems must distinguish distributions with narrower margins.
Reading a multilevel cell may require several comparison voltages. Logical bits are assigned to threshold ranges using an encoding chosen to limit the consequences of likely errors. Controllers can retry a difficult read using adjusted reference voltages, while advanced decoders can use information about how close a measurement lies to a decision boundary. Programming also becomes more elaborate, often using multiple passes to control the effects of neighboring cells and the sequence in which logical pages are written.
As a broad technological tendency, SLC supports higher endurance and simpler, faster programming than denser modes made with comparable technology. TLC and QLC reduce cost per stored bit but require stronger management. Exact endurance is not determined by the label alone: process generation, cell structure, operating temperature, qualification criteria, and controller behavior all matter. Figures such as thousands of program/erase cycles for many TLC products or hundreds to low thousands for some QLC products are illustrative ranges, not universal specifications.
Many SSDs operate part of a multibit NAND array temporarily in a one-bit mode, often called pseudo-SLC. Incoming writes can be accepted into this faster region and later reorganized into denser storage. This is an operating policy, not necessarily a separate population of physically dedicated SLC chips. The available cache may depend on free capacity and controller configuration. Once it fills, sustained write speed can fall toward the rate at which the device programs its underlying TLC or QLC storage.
Three-dimensional NAND and fabrication
For much of NAND’s early development, manufacturers increased density by shrinking cells arranged across a silicon surface. This planar approach benefited from advances in photolithography, but smaller cells stored fewer electrons and experienced stronger coupling to neighboring cells. Insulating layers, voltage margins, and fabrication variability became increasingly difficult to scale together. Three-dimensional NAND changed the direction of integration by placing memory cells at multiple heights.
In a common vertical NAND structure, a channel passes through a stack of layers, and gate structures at different heights define successive cells along the channel. The resulting string contains memory transistors arranged vertically rather than solely across the wafer surface. Many commercial designs use charge-trap storage, although three-dimensional NAND has also been implemented with floating-gate structures. “3D NAND” identifies the spatial organization, not one exclusive storage material or manufacturing sequence.
Fabrication requires depositing many layers and etching deep, narrow holes with closely controlled profiles. The holes must remain suitable for channel and memory-film formation over their full depth. Small variations in diameter, thickness, or material properties can produce different electrical behavior at different positions in a string. Manufacturers may build stacks in multiple decks and connect them, rather than etching an entire tall structure in one uninterrupted operation.
Layer count is an important but incomplete measure of progress. A product advertised with a given number of layers may distinguish active memory layers from other structural layers. Capacity and cost also depend on lateral cell dimensions, bits per cell, string layout, peripheral circuitry, manufacturing yield, and the number of dies placed in a package. Two products with the same advertised layer count can therefore differ substantially in useful density and performance.
Peripheral circuits provide sensing, decoding, voltage generation, and data transfer. Placing some of this circuitry beneath the memory array can reduce die area. Other manufacturing approaches fabricate array and logic structures separately and join them through wafer bonding. These techniques address the difficulty of optimizing dense memory structures and high-performance control circuitry within one process flow.
Vertical integration relaxes some pressures associated with planar shrinking, but it introduces others: very high-aspect-ratio etching, longer strings, complex interconnections, mechanical stress, and variation across the stack. Increasing layer count does not automatically make each operation faster or each cell more reliable. The economic advantage emerges only when the additional stored bits outweigh the costs of more elaborate processing and acceptable-yield fabrication.
Controllers, address translation, and write management
A managed flash device presents a simpler storage interface than its underlying NAND arrays. The host usually reads and writes logical blocks without selecting physical pages or erase blocks. A controller maintains the mapping between these logical addresses and physical locations through a flash translation layer (FTL). This separation allows the device to replace worn blocks, distribute writes, and reorganize data while preserving the host’s apparent address space.
NAND cannot generally overwrite an arbitrary page in place. When the host updates a logical block, the controller writes the new version into an available physical page and changes the mapping. The previous page becomes invalid, but its block cannot be erased until any still-valid pages have been preserved elsewhere. This out-of-place update process is fundamental to how flash storage differs from a simple array of independently rewritable sectors.
Storage garbage collection recovers space by selecting blocks, copying their valid contents, and erasing them. A block containing mostly invalid pages is relatively cheap to reclaim; one containing mostly valid pages requires much more copying. These internal writes contribute to write amplification:
The precise measurement depends on what internal activity is counted. In ordinary operation, copying, metadata updates, and other management work can cause the NAND to receive substantially more data than the host originally submitted.
Wear leveling prevents frequently updated logical addresses from exhausting a small set of physical blocks. Dynamic wear leveling chooses suitable blocks for new writes. Static wear leveling can also relocate long-lived data so that lightly worn blocks are not permanently occupied while other blocks absorb nearly all updates. This improves wear distribution but itself introduces copying. Controllers must balance immediate performance, free-space availability, and longer-term endurance.
Over-provisioning reserves physical capacity beyond the logical capacity exposed to the host. The reserve supports replacement blocks, garbage collection, and more flexible allocation. TRIM commands, or analogous deallocation commands, tell the device that certain logical ranges no longer contain useful host data. They can reduce unnecessary preservation of obsolete pages, but they do not necessarily trigger immediate erasure.
The mapping tables and other metadata are themselves critical data. Controllers may cache them in SRAM or external DRAM and maintain recoverable records in NAND. Enterprise devices may use capacitors to provide enough energy to complete selected operations after external power fails. The durability of an acknowledged write depends on these protections and on how the controller implements the host’s persistence commands.
Endurance, retention, and error correction
Flash reliability involves several distinct properties. Endurance describes the ability to withstand program/erase cycling. Retention concerns how long data remain recoverable under specified conditions. Read disturb and program disturb describe changes caused by operations on nearby or related cells. These mechanisms interact, but a device can have good retention under one condition and poor endurance under another; no single lifetime number captures every aspect of reliability.
Repeated programming and erasing damage insulating materials and create charge traps. This changes how cells accept, retain, and release charge, widening the distributions used to distinguish states. Retention loss can arise from leakage and charge redistribution, and its rate depends strongly on temperature and prior wear. Consequently, a retention specification is meaningful only together with the conditions under which it applies, including the cell’s cycling history.
Reading is not completely passive at the array level. In NAND, unselected cells receive pass voltages during access to another cell in the string. Repeated exposure can shift their thresholds, producing read disturb. Programming can also influence neighboring cells through electrical coupling or other interactions. Controllers track errors and may relocate data before the accumulated disturbance exceeds the available correction margin.
An error-correcting code adds redundant information so that corrupted data can be detected and, within limits, reconstructed. NAND systems have used Bose–Chaudhuri–Hocquenghem codes and increasingly low-density parity-check codes. Modern decoding can combine multiple reads at different reference voltages to obtain soft information about uncertain bits. A successful logical read may therefore involve substantial internal work even when the host sees only a completed block request.
Flash devices distinguish raw errors from errors remaining after correction. Bad blocks can exist when a NAND chip leaves the factory, and additional blocks can fail during service. Managed devices identify and retire unsuitable blocks and replace their capacity from reserved space. Enterprise storage may add protection across dies or channels, providing recovery from some failures larger than a decoder’s individual codeword. Such protection is different from an independent backup.
Drive-level endurance is often expressed as terabytes written (TBW) or drive writes per day (DWPD). For a rated host-write allowance , usable capacity , and service interval of days,
with consistent units. These ratings describe specified workloads and warranty conditions, not a precise failure threshold for every device. Host write patterns, write amplification, temperature, and unpowered storage intervals all influence the relationship between laboratory qualification and actual service behavior.
Interfaces and performance characteristics
Flash performance is determined at several levels: the cell array, the NAND interface, the controller, the host connection, and the software issuing requests. A page read requires sensing before its contents can be transferred. Programming typically takes longer than reading, and block erasure is generally slower still. Exact timings vary by architecture and operating mode, so host-interface bandwidth alone cannot predict the latency of a particular storage request.
Controllers obtain throughput by operating multiple channels, packages, dies, and planes in parallel. A plane is an internal subdivision that may support operations alongside other planes, subject to device restrictions. By overlapping independent work, a drive can sustain high aggregate transfer rates even when a single NAND operation is relatively slow. This is a storage-specific application of parallel computing, with scheduling constrained by shared buses, block placement, and dependencies between commands.
Raw NAND interfaces, including those standardized through ONFI and Toggle-mode families, connect memory chips to controllers. Managed flash products expose higher-level interfaces. Embedded MultiMediaCard (eMMC) integrates NAND and a controller in a standardized package, while Universal Flash Storage (UFS) provides a serial interface and command architecture widely used in mobile devices. Removable SD cards and USB drives add their own protocols and packaging requirements.
SSDs commonly connect through SATA or through PCI Express using Non-Volatile Memory Express (NVMe). SATA preserves compatibility with an interface ecosystem developed around earlier disk storage. NVMe is designed for nonvolatile storage and supports multiple submission and completion queues, helping reduce protocol overhead and expose device concurrency. NVMe is not a type of flash cell: it is a host communication architecture, and the same NAND technology can appear behind different interfaces.
Benchmark results depend heavily on workload and device state. Sequential transfers can exploit large contiguous requests, whereas small random requests incur different mapping and scheduling costs. Queue depth affects available parallelism. A newly erased or lightly used drive may perform differently from one undergoing sustained overwrites with little free space. Pseudo-SLC cache exhaustion, background garbage collection, and thermal throttling can all change observed throughput.
Latency distributions can be more revealing than average speed. Most requests may complete quickly while a minority wait for reclamation, retries, or internal maintenance. Applications such as a database can be sensitive to these long-tail delays, especially when operations depend on durable log writes. A meaningful performance description therefore includes request size, read/write mixture, queue depth, measurement duration, device fill state, and persistence requirements rather than a single advertised maximum transfer rate.
Roles in computing and embedded systems
Flash occupies a distinctive position in the memory hierarchy. It is denser and less expensive per stored bit than most working-memory technologies, retains data without power, and supports faster random access than rotating magnetic disks. It nevertheless has much higher access latency and more restrictive update behavior than processor caches or DRAM. Mainstream systems therefore use it primarily for persistent storage rather than as a direct replacement for ordinary working memory.
NOR flash remains important in embedded systems that need reliable access to relatively small amounts of executable code. Microcontrollers may incorporate flash on the same chip, enabling electrically updateable control programs. External serial NOR commonly stores boot firmware, configuration data, and code for networking devices, consumer electronics, and industrial equipment. Integrating embedded flash with advanced logic manufacturing can be difficult because the memory requires additional materials and processing steps.
NAND dominates applications in which capacity and cost per bit are central. Digital cameras record image files to memory cards; phones store applications and media in managed flash; and computers use SSDs for operating systems and user data. The absence of moving heads and rotating platters improves resistance to mechanical shock and removes seek delays. These advantages do not make all flash devices equivalent: a low-cost removable drive and an enterprise SSD can differ greatly in controllers, endurance, power-loss behavior, and error reporting.
In data centers, flash supports databases, virtual-machine storage, analytics, and distributed storage services. Low access latency can reduce time spent waiting for small random reads, while large arrays of drives provide substantial aggregate bandwidth. Flash may hold complete datasets or serve as a faster tier in front of magnetic storage. Capacity planning must account for workload endurance, redundancy, spare capacity, and the bandwidth needed to rebuild data after failure.
Software can either hide or explicitly accommodate flash behavior. A conventional file system usually relies on the managed device’s block interface. Filesystems designed for raw flash, including JFFS2 and UBIFS in Linux, incorporate knowledge of erase blocks, out-of-place updates, and bad-block handling. Other approaches expose zones or more placement control to the host, allowing software to organize writes around device constraints.
Flash also has environmental and operational implications. Removing mechanical components can reduce some forms of operating power, but controller activity, high interface bandwidth, and dense arrays still consume substantial energy and generate heat. Its nonvolatility eliminates the need to keep the storage medium continuously powered merely to retain data, while manufacturing cost and finite write endurance remain important constraints on its role in long-term storage systems.
Data persistence, sanitization, and security
The separation between logical addresses and physical pages has important consequences for deletion. When a operating system removes a file, it usually updates filesystem records rather than immediately destroying every physical copy of its contents. Even overwriting the same logical address may place new data elsewhere because the controller uses out-of-place writes. Obsolete contents can remain in invalid pages, spare regions, or blocks awaiting erasure.
Deallocation commands improve storage management but are not equivalent to verified sanitization. A device may return zeros for deallocated logical addresses while old charge patterns still exist in inaccessible physical locations. Conversely, background garbage collection may eventually erase those locations without further host action. The timing and physical consequences depend on the product’s implementation, and the host normally cannot inspect every reserved or retired block through ordinary read commands.
Storage standards define sanitization mechanisms intended to operate beyond simple logical overwriting. Depending on the device, these may include block erasure or cryptographic erasure. Encryption can make all stored user data dependent on an internal media-encryption key; destroying the necessary key material can then render the encrypted contents inaccessible. This depends on complete encryption coverage, correct key handling, and an implementation that actually performs the claimed operation.
Power loss creates a different persistence problem. A write may pass through host caches, controller buffers, mapping updates, and NAND programming before reaching a recoverable state. Flush operations and force-unit-access semantics communicate persistence requirements, but correctness depends on the entire implementation. Filesystem journaling and database transaction protocols address logical consistency; they do not by themselves compensate for hardware that falsely reports durable completion. Power-loss protection reduces some risks by preserving enough energy to complete specified internal actions.
Flash contents can also be examined outside the normal device interface. Chip-off analysis removes or accesses memory components directly, but recovering meaningful files can require reconstructing interleaving, error correction, scrambling, address translation, and encryption. The physical presence of residual data does not guarantee practical recovery, especially when encryption keys are unavailable. Equally, an apparently empty logical volume does not establish that all underlying cells have been erased.
Firmware is another security boundary because the controller determines address mapping, error handling, and access to internal commands. Vulnerabilities in a storage controller can affect confidentiality, integrity, or availability independently of the physical memory cells. Secure update mechanisms, authenticated firmware, and well-defined device commands are therefore relevant to flash-based storage security, alongside the physical questions of charge retention and erasure.