{
  "id": "static-random-access-memory",
  "title": "Static Random-Access Memory",
  "lang": "en",
  "source_lang": "en",
  "summary": "Static random-access memory stores data in bistable electronic circuits, providing fast access without refresh while power is maintained.",
  "category": "computer-science",
  "subcategory": "hardware",
  "type": "concept",
  "aliases": [
    "SRAM",
    "static RAM",
    "static read/write memory"
  ],
  "generated_at": "2026-11-03",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/static-random-access-memory/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/static-random-access-memory/",
      "json": "https://aiwiki.page/zh/static-random-access-memory.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "Static random-access memory (SRAM) is a form of semiconductor memory that stores each bit in a circuit with two stable electrical states. Unlike dynamic random-access memory (DRAM), it does not require periodic refresh to preserve its contents during normal powered operation. “Static” describes this retention mechanism, not permanence: ordinary SRAM is volatile and loses its guaranteed stored state when its power supply is removed.\n\nSRAM combines low access latency, high access bandwidth, and compatibility with logic manufacturing processes. These properties make it fundamental to processor caches, on-chip buffers, and many small working memories. Its principal disadvantage is area: a conventional cell uses six transistors, whereas a conventional DRAM cell uses one transistor and one capacitor. Consequently, SRAM generally provides less capacity per unit of silicon and costs more per stored bit.\n\nSRAM encompasses both embedded memory blocks inside processors and other chips and standalone memory devices with standardized electrical interfaces. Its behavior depends not only on the storage cell but also on address decoding, sensing, write circuitry, power distribution, and the surrounding system. Designs therefore balance speed, density, energy, stability, and reliability rather than optimizing any one characteristic in isolation."
    },
    {
      "heading": "Storage cells and bistable operation",
      "text": "The standard modern SRAM cell is the six-transistor, or 6T, cell, usually implemented in complementary metal–oxide–semiconductor technology. Four transistors form two cross-coupled inverters, while two additional access transistors connect the internal storage nodes to a pair of complementary bitlines. Each inverter contains a pull-up and a pull-down metal–oxide–semiconductor field-effect transistor. The output of each inverter drives the input of the other, creating positive feedback. In either stable state, one storage node is near the supply voltage and the other is near ground.\n\nThe two access transistors are controlled by a wordline. When this wordline is inactive, the cell is isolated from the bitlines, and the cross-coupled inverters maintain their state. An ideal CMOS latch draws no continuous current through a direct supply-to-ground path in either stable state. Real devices nevertheless consume leakage current, and the surrounding array circuits may draw additional standby current. Retention therefore requires continuing electrical power even when no accesses occur.\n\nFor a conventional differential read, the bitlines are first precharged, commonly to the supply voltage, and equalized. Activating the wordline connects the storage nodes to the bitlines. The cell’s low-valued side slightly discharges its associated bitline through the access and pull-down transistors. A sense amplifier detects the resulting voltage difference before the bitline has to undergo a full voltage swing. This reduces access time and switching energy. Unlike a conventional DRAM read, the operation does not inherently destroy the stored information, although an inadequately designed or operated cell can suffer read disturbance.\n\nWriting requires the external circuitry to overcome the cell’s existing feedback. Write drivers impose opposite values on the two bitlines, and activation of the wordline allows these voltages to force the internal nodes into the desired state. After the cell switches and the wordline becomes inactive, positive feedback preserves the new value.\n\nThese operations impose competing transistor-strength requirements. During a read, the pull-down device must keep the low storage node from rising enough to flip the cell. During a write, the access path must be able to defeat the relevant pull-up device. Designers adjust transistor dimensions, threshold voltages, supply conditions, and timing to obtain an acceptable balance. The central engineering problem is therefore not merely building a latch, but building a compact latch that can be read safely and overwritten reliably across manufacturing variation and operating conditions."
    },
    {
      "heading": "Arrays, decoding, and peripheral circuitry",
      "text": "An SRAM organizes many storage cells into a rectangular array, with wordlines typically running across rows and bitlines along columns. An address is divided into fields that select a row and, where necessary, a subset of columns. Row decoders assert the appropriate wordline, while column-selection circuitry connects selected bitlines to sensing or writing circuits. The externally visible word width need not match the physical number of cells in a row: a row can contain several independently selectable words.\n\nFor a memory containing \\(2^a\\) addressable words of \\(w\\) bits each, the nominal data capacity is\n\n\\[C = 2^a w \\text{ bits}.\\]\n\nThus, an organization described as \\(64\\text{K} \\times 32\\) contains 65,536 words of 32 bits, or 2,097,152 data bits. Physical storage can exceed the advertised capacity because the implementation may include parity bits, error-correction bits, spare rows, and spare columns. Conversely, array capacity does not directly reveal total chip area, since decoding and other peripheral circuits also occupy silicon.\n\nLong wordlines and bitlines introduce resistance and capacitance. Connecting more cells to a bitline increases its electrical load, making it harder for a small cell current to develop a detectable voltage difference quickly. Large memories are therefore divided into banks and smaller subarrays. Local wordlines, local sensing, hierarchical decoding, and intermediate buffers shorten critical electrical paths. This partitioning improves timing but adds circuitry, routing, and sometimes access-selection overhead.\n\nA memory compiler generates embedded SRAM macros for a particular manufacturing process and a chosen combination of depth, width, ports, and performance options. Its output normally includes a physical layout, functional models, timing information, and power characterization. The macro is treated as a specialized component within an integrated circuit, rather than synthesized entirely from ordinary logic cells. Highly repetitive, tightly constrained bitcell layouts can achieve much higher density than an equivalent storage structure assembled from general-purpose flip-flops.\n\nPeripheral design can determine the practical limits of the array. Sense amplifiers have offset caused by device mismatch; if enabled before a sufficient bitline difference develops, they can resolve the wrong value. Write drivers must provide adequate strength without imposing excessive load. Internal timing circuits coordinate precharge, wordline activation, sensing, and restoration of the bitlines. Some designs use replica cells or replica bitlines to track process, voltage, and temperature effects. Such circuits approximate the behavior of the array and help generate appropriately timed internal control signals, although they cannot eliminate all differences between a representative path and the weakest actual cell."
    },
    {
      "heading": "Interfaces, ports, and access timing",
      "text": "Standalone SRAMs are commonly classified as asynchronous or synchronous according to their external interface. An asynchronous SRAM does not require a clock to initiate every transaction. Address inputs, chip enable, output enable, and write enable determine its behavior. After an address and the relevant controls become valid, read data appears within the specified access time. During a write, the device imposes timing requirements on the address, input data, and control pulses.\n\nThese requirements distinguish several quantities that are sometimes loosely grouped under “speed.” Address access time describes the delay from a valid address to valid output data under stated conditions. Cycle time specifies the minimum spacing between accesses. Output-disable time describes how quickly the device releases a shared data bus. Write pulse width, setup time, and hold time determine whether a write completes correctly. A device can have an access time shorter than its minimum cycle time, so one timing number does not fully characterize its behavior.\n\nSynchronous SRAM accepts addresses and control signals in relation to a clock edge. Input registers and output pipelines can simplify system timing and sustain a high transaction rate. Burst devices generate a sequence of related addresses internally, reducing address-interface overhead. Pipelining improves throughput but introduces latency measured in clock cycles: a memory may accept a new request each cycle even though the response to each request appears several cycles later. These are different properties, and both matter when integrating memory with a central processing unit or an accelerator.\n\nPort count is another independent design dimension. A single-port memory typically performs one selected read or write operation at a time. A simple dual-port memory commonly provides one read port and one write port, whereas a true dual-port memory permits more general operation through two addressable ports. Terminology varies among products, making the detailed port specification more authoritative than the label alone. Independently accessible ports usually require extra cell connections, routing, and peripheral circuits, increasing area and complexity.\n\nConcurrent operations also introduce semantic questions. Two ports may read the same address without conflict, but simultaneous writes to that address require defined arbitration or may be prohibited. If one port reads while another writes the same location, the output may be the old value, the new value, or an unspecified value, depending on the implementation. Dual-clock memories additionally require careful treatment of timing relationships. A dual-port array does not by itself solve synchronization problems between independent clock domains; systems often add pointers, handshake logic, or arbitration to establish a reliable communication protocol."
    },
    {
      "heading": "Latency, bandwidth, and power consumption",
      "text": "SRAM’s speed advantage over conventional DRAM arises from its storage and access organization rather than from the word “static” alone. A small bistable cell can drive a local bitline difference rapidly, and SRAM does not require the sequence of charge sensing, destructive-read restoration, and periodic refresh associated with DRAM arrays. Nevertheless, a large SRAM can have substantial delay because of address decoding, interconnect, banking, sensing, and output distribution. A small local SRAM and a large shared cache should not be assumed to have comparable latency.\n\nBandwidth depends on word width, port count, bank organization, and clock rate. If a memory transfers \\(w\\) bits per transaction at a rate of \\(f\\) transactions per second, its ideal data bandwidth is\n\n\\[B = wf.\\]\n\nFor example, a 128-bit interface transferring one word per cycle at 1 GHz has an ideal bandwidth of 128 gigabits per second, or 16 gigabytes per second. This figure excludes arbitration delays, bank conflicts, inactive cycles, and protocol overhead. Multiple banks can increase aggregate bandwidth when requests are distributed across them, but repeated accesses to one busy bank can leave other banks idle.\n\nA useful first-order model for switching power is\n\n\\[P_{\\mathrm{dynamic}} \\approx \\alpha C_{\\mathrm{sw}} V_{\\mathrm{DD}}^2 f,\\]\n\nwhere \\(\\alpha\\) is an activity factor, \\(C_{\\mathrm{sw}}\\) is the effective switched capacitance, \\(V_{\\mathrm{DD}}\\) is the supply voltage, and \\(f\\) is the operating frequency. Bitlines often contribute heavily because many cells and metal segments load them. Reducing the voltage swing, selecting fewer active columns, and dividing an array into smaller subarrays can lower switching energy. The formula is an approximation: sense amplifiers, internal timing, short-circuit currents, and different voltage domains complicate detailed accounting.\n\nStandby consumption follows different constraints. Even without switching, subthreshold and gate leakage flow through real transistors. A large SRAM can therefore consume significant power simply to retain its contents. High-threshold devices can reduce leakage but usually weaken drive current. Lowering the supply voltage reduces power but also reduces the margins available for reading, writing, and retention. The lowest safe retention voltage can differ substantially from the lowest voltage that supports reliable high-speed access.\n\nPower management distinguishes retention from shutdown. A retention mode maintains a reduced supply sufficient to preserve data, usually with access disabled. Power gating disconnects an inactive block more fully and generally discards ordinary SRAM contents. Reactivation then requires initialization, refilling, or restoration from another memory. Cache systems can invalidate discarded lines, while tightly coupled working memories may need explicit software or hardware management. Energy savings must therefore be weighed against wake-up delay, state-reconstruction cost, and expected idle duration."
    },
    {
      "heading": "Cell variants and comparison with other memories",
      "text": "The 6T cell dominates many SRAM designs because it balances density, speed, and compatibility with standard CMOS fabrication. It is not the only possible static cell. Earlier designs included resistive-load cells and other transistor arrangements, while specialized modern cells add devices to improve read stability, support additional ports, or operate at reduced voltage. Transistor count describes an important structural property but does not, by itself, determine practical area or performance.\n\nAn eight-transistor, or 8T, design often separates the read path from the internal storage nodes. A conventional 6T cell uses the same access connections for reading and writing, exposing the stored state to read disturbance. A decoupled read port can reduce this interaction and improve low-voltage read robustness. The cost includes additional transistors and interconnect; some designs also require special measures to control leakage on single-ended read bitlines. Ten-transistor and other expanded cells explore related trade-offs, but their precise circuits and benefits vary.\n\nMultiport SRAM is especially demanding because each independently addressable port can require additional wordlines, bitlines, and access devices. The resulting wiring burden grows quickly. A processor register file, for example, may need to supply several operands and accept several results each cycle. Such structures can use custom multiport SRAM-like cells or other circuit arrangements. They often sacrifice density for access bandwidth and are not necessarily implemented using the same bitcell as the processor’s larger caches.\n\nThe central comparison with DRAM concerns density and retention mechanism. A conventional DRAM cell stores charge in a capacitor accessed through one transistor; leakage requires refresh. A conventional SRAM cell maintains a bistable circuit through feedback. DRAM’s compact storage cell makes it more economical for large main memories, although its manufacturing and access circuitry are specialized. SRAM is often easier to integrate with high-performance logic, but using a standard logic process does not make each stored bit inexpensive.\n\nFlash memory differs more fundamentally because it retains information without continuous power. Its programming and erasing mechanisms, access characteristics, and endurance constraints differ from those of SRAM. SRAM permits frequent ordinary reads and writes without a flash-style erase operation. Products described as “nonvolatile SRAM” may combine an SRAM working array with a nonvolatile backup mechanism, while battery-backed SRAM maintains the supply externally. Neither arrangement changes the volatility of the underlying ordinary SRAM cell. Similarly, pseudo-static RAM presents an SRAM-like interface but uses an internally refreshed dynamic array; interface behavior alone does not establish the physical storage technology."
    },
    {
      "heading": "Roles in computer and electronic systems",
      "text": "SRAM occupies the fast, relatively small-capacity levels of the memory hierarchy. In processors, it commonly stores cache data and the associated tags that identify which memory addresses are present. The fastest cache levels are usually close to execution units and optimized for short access paths, while larger shared levels prioritize capacity and aggregate throughput. Cache latency also includes tag comparison, selection, and control, so the delay of a cache lookup is not identical to the delay of its data array.\n\nCache organization makes extensive use of banking, replicated structures, and multiple independently accessed arrays. A set-associative cache may read several candidate entries and select one after comparing tags. This improves flexibility in placing data but can consume more energy than accessing a single predetermined location. Cache control additionally handles valid bits, replacement information, and consistency state. These supporting fields can occupy significant storage even though published cache capacities commonly emphasize data capacity.\n\nIn microcontrollers and other embedded systems, SRAM often serves as directly addressable working memory for stacks, variables, and buffers. Some processors provide tightly coupled memory or software-managed scratchpads rather than relying only on caches. A scratchpad exposes storage whose placement is explicitly controlled, avoiding cache-miss behavior for the data stored there. Its overall timing can still depend on bus arbitration, port contention, and concurrent transfers. Deterministic access therefore follows from the full system design, not simply from choosing SRAM.\n\nCommunication hardware uses SRAM for packet buffers, queue descriptors, lookup results, and control state. A router may need to absorb simultaneous arrivals, maintain many queues, and deliver packets at the output links’ rates. Some functions require random access to small records, while others benefit from sequential buffering. SRAM is attractive where access rate and predictable latency outweigh capacity cost; larger buffering needs may instead use DRAM or a combination of technologies.\n\nA graphics processing unit uses extensive on-chip storage for registers, caches, and local data exchange. Machine-learning accelerators likewise use SRAM for weights, activations, partial sums, and temporary buffers. Retaining frequently reused operands near arithmetic units reduces transfers to external memory, which can otherwise dominate execution energy and limit performance. The usefulness of any particular SRAM capacity depends on workload dimensions, data precision, and scheduling. Arrays supporting neural-network inference may be organized around tiled operations so that data loaded once can participate in many arithmetic steps before replacement."
    },
    {
      "heading": "Stability, defects, and data integrity",
      "text": "SRAM reliability includes both permanent defects and transient failures. Permanent faults can arise from fabrication defects, damaged interconnect, or devices outside acceptable electrical limits. Transient errors can result from electrical noise, supply disturbances, timing violations, or energetic particles depositing charge near sensitive circuit nodes. These mechanisms are distinct: a cell that fails every read has different diagnostic and repair implications from a normally functioning cell that occasionally changes state.\n\nA central cell-level measure is static noise margin. For a conventional cross-coupled inverter pair, it can be visualized using the “butterfly” formed by plotting one inverter’s transfer characteristic against the mirrored characteristic of the other. The side length of the largest square fitting within the relevant lobes provides a standard measure of tolerable static disturbance. Hold and read conditions produce different margins because activating the access transistors changes the circuit. Write ability requires separate characterization, and static metrics alone do not capture every failure in short, dynamically timed accesses.\n\nManufacturing variation matters because an array must operate correctly across an enormous population of cells. Random differences in device thresholds and dimensions can weaken a small fraction of cells even when a typical cell performs well. Large memories therefore cannot be designed solely around average behavior. Characterization covers process variation, voltage, and temperature, together with timing and local supply effects. Reduced-voltage operation is particularly challenging because available signal swings and transistor overdrive become smaller.\n\nProduction testing commonly uses built-in self-test circuits and sequences designed to expose stuck bits, transition failures, addressing defects, and interactions between cells. Spare rows or columns can replace defective regions through programmed repair information. During normal operation, parity can detect selected error patterns, while an error-correcting code can also reconstruct data within its correction capability. A common single-error-correcting, double-error-detecting organization corrects one erroneous bit per protected codeword and detects two. Larger or correlated error patterns require stronger protection or additional system measures.\n\nPhysical organization affects protection effectiveness. If nearby cells belong to different codewords, a localized disturbance is less likely to produce several errors within one word; this technique is commonly called interleaving. Periodic checking and correction, or scrubbing, can prevent isolated errors from accumulating, but it serves a different purpose from DRAM refresh. Security uses introduce another dimension: startup variations in SRAM can support a physical unclonable function, while short-lived data remanence under some power-loss conditions means immediate erasure cannot always be assumed. Secure systems consequently distinguish logical invalidation, active overwriting, and physical loss of stored information."
    },
    {
      "heading": "Historical development and scaling challenges",
      "text": "SRAM emerged from the broader development of transistorized bistable circuits and monolithic semiconductor memory during the 1960s. Integrated memory gradually displaced earlier technologies where size, speed, and manufacturing economics favored semiconductor implementation. Early static memories used several circuit families, including bipolar designs and metal–oxide–semiconductor designs. The later dominance of CMOS should not be projected backward onto all early SRAM products.\n\nIntel’s first commercial product, the 3101 introduced in 1969, was a 64-bit bipolar static memory organized as 16 words of four bits. The company also introduced the 1101 in 1969, a 256-bit MOS static memory. These capacities were tiny by later standards, but the products illustrate the transition toward commercially manufactured semiconductor arrays. Subsequent improvements in photolithography, circuit design, and production yield expanded capacity and lowered cost per bit.\n\nAs microprocessors became faster, the delay of off-chip memory access became increasingly important. Systems used external SRAM caches to reduce the frequency of slower main-memory accesses, while growing transistor budgets later allowed substantial cache capacity to move onto processor dies. Embedded SRAM became one of the largest consumers of chip area in many high-performance designs. Its design objectives increasingly diverged by function: a small first-level cache emphasized latency, while a larger last-level cache emphasized capacity, power, and shared bandwidth.\n\nScaling creates difficulties because SRAM cells must preserve carefully balanced electrical behavior while occupying minimal area. Smaller geometries can increase sensitivity to device variation and reduce available operating margins. Bitcell layout is also constrained by manufacturing rules, local interconnect, and the need to share power rails and diffusion regions efficiently. The nominal technology-node designation is therefore not enough to predict SRAM density. A logic process can improve transistor performance without delivering a proportional reduction in bitcell area.\n\nAdvanced transistor structures, including fin-based and gate-all-around devices, change the design space but do not remove the underlying trade-offs. Read-assist techniques may reduce wordline voltage to limit disturbance, while write-assist techniques may temporarily lower the cell supply, boost the wordline, or apply a below-ground bitline voltage. Each technique introduces timing, energy, reliability, or circuit-complexity costs. At the architectural level, designers also investigate three-dimensional integration and chiplet arrangements to place more memory near computation. Research on in-memory computing explores performing selected operations within or near SRAM arrays, including bitwise logic and accumulation, but practical implementations must address precision, data movement, circuit variation, and compatibility with ordinary memory accesses."
    }
  ],
  "links": [
    "semiconductor",
    "bit",
    "cpu-cache",
    "cmos",
    "mosfet",
    "integrated-circuit",
    "central-processing-unit",
    "synchronization-computing",
    "flash-memory",
    "memory-hierarchy",
    "router",
    "graphics-processing-unit",
    "neural-network-inference",
    "error-correcting-code",
    "photolithography",
    "microprocessor",
    "chiplet"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}