EVOPEDIA

← all systems · see it in The System

Computer memory

The component that stores information, such as data and programs, for immediate use in the computer: instructions fetched by the processor and the data those instructions fetch and store, held in memory cells (today MOS semiconductor cells on integrated circuits, volatile DRAM and SRAM and non-volatile flash and ROM), grouped into words reached by binary addresses, arranged in a hierarchy from registers and caches through main memory to mass storage, and managed by the operating system through virtual memory and protection.

Properties published scienceour proposal

Sourced fields, and our own readings.

aliasesComputer memory (the store the rule lives in)
born atearly 1940s (ENIAC's tube store; delay lines 1942-44); the first random-access store 1946-48; contested (C-2884)
born precisionabout five years
boundaryInside: the storage cells (DRAM capacitors and access transistors, SRAM flip-flops, flash floating gates; historically mercury delay lines, Williams-tube charge spots, magnetic cores), the arrays of rows and columns with their wordlines, bitlines, sense amplifiers and refresh, the chips and modules that group cells into addressable words, and the hierarchy of levels from processor registers and caches to main memory and mass storage. Outside: the processor that issues addresses and consumes words, the operating system that maps and protects memory, the programs whose instructions and data it holds, the power supply, the heat sink and air, the particles that pass through it, and the people who build, program and attack it. On the line: the memory controller and memory channel, the memory management unit and page tables (virtual memory), and error-correcting codes.
nested levelBuilt from MOS transistors and capacitors on integrated circuits (earlier mercury columns, cathode-ray tubes and ferrite rings), organised into cells, arrays, chips and modules, and those into a hierarchy; computer memory is a part of computer hardware and the store (M) of a computer. Programs, the stored program and software run on it (booked from their side as runsOn).
scaletechnological

Measures (38) published science

characteristic clock: DRAM refresh interval64 or less (each row); one row every 7.8 us for 8,192 rows staggered ms
register access0.25 ns (1 cycle at 4.0 GHz); 18,432 bits per core
L1 data cache32 KiB; 1 ns (4 cycles); up to 64 GB/s
L2 cache1 MB; 3.5 ns (14 cycles)
L3 cache16-32 MB shared by 8 cores; 11.75 ns (47 cycles)
main memory (DDR5)64 GiB; about 60 GB/s; 82.5 ns
solid-state drive2 TB; 2000 MB/s; 0.2 ms
hard disk drive18 TB; 500 MB/s; 4.16 ms
spun-down disks (nearline)petabytes; 25 s
tape libraryexabytes; 160 MB/s; minutes
Atlas one-level store16,384 words core plus 98,304 words drum
delay-line capacity rulecapacity = recirculation time / time per bit; early systems a few thousand bits, recirculation in microseconds
EDSAC store256 35-bit words in 16 delay lines of 560 bits (later 512 words)
UNIVAC I average access timeabout 222 us
speed of sound in mercury1450 (air 343.2) m/s
mercury line temperature40 C
core memory density and costabout 32 kilobits per cubic foot by the late 1960s; about $1 to about 1 cent per bit bits/ft3; $/bit
electrons per flash cellabout 100 (1x-nm MLC) electrons
flash endurance150,000 (older SLC) to 3,000 (1x-nm MLC) P/E cycles
Landauer bound per erased bitkB T ln 2, about 3e-21 at 300 K J
measured erasure heat, long cycles0.72 kB T (bound 0.693 kB T)
share of system energy spent moving data40% scientific, 35% mobile, 62% consumer
energy of a memory access against an additionabout 115 times ratio
DRAM share of system powermore than 40% in recent CPU or GPU systems; off-chip hierarchy 40-50% in 2003 IBM servers
memory capacity per core trendfalls about 30% every two years (cores double every two years, DRAM capacity every three) %/2 yr
DRAM latency change over 20 yearsreduced by only 30% while capacity grew about 128x and bandwidth about 20x
refresh reduction with retention bins74.6% fewer refreshes; DRAM power -16.1%; performance +8.6% (8 cores, 32 GB)
RowHammer thresholdas few as 139K accesses; up to one in 1.7K cells susceptible
modules vulnerable to RowHammer110 of 129 (2008-2014 modules; all from 2012-2013)
DRAM error rate in the field25,000-70,000 per billion device hours per Mbit; over 8% of DIMMs per year FIT/Mbit
soft-error rate with altitudeneutron flux about 14 per cm2 per hour at New York sea level; x2.2 per 1000 m; mountains an order of magnitude higher; aircraft over 300x per cm2 h
alpha emission limit for packagingunder 0.001 (a shoe sole 0.1-10) counts per hour per cm2
DRAM remanence when cooledunder 1% of bits decayed after 10 min unpowered at about -50 C
DRAM remanence in liquid nitrogen0.17% after 60 min at about -196 C
durability of conventional storage mediaat most 50 years in optimal conditions
DNA storage theoretical densitypetabytes per gram (usually unreachable) bytes/g
DRAM price per bit, 2017+47% (largest jump in 30 years); early 2026 increases some above 200%
address spaceN address bits reach 2^N words

Open contentions (4) open contention

Where sources or readings disagree. Both stay on the map.

Partners and parts published science

Built from / builds (15)

Sits inside / holds (3)

Trades with (valence) (17)

expressed by (17)

floor filled by (34)

Kinds (9)

wired (4)

Notes published science

Layers

The ways it takes in its world, where they meet, and how it acts back, each from L1 (where it meets the world) up.

computer memory as computer engineers describe it: charge, fields and signals at the cells up to the stored program held for immediate use (reader's view)

  1. L1: The physical world at the memory: commands, addresses and data words arriving over the memory channel, the supply voltage, the controller's refresh timing, the heat around the chips, and the particles (cosmic-ray neutrons, alpha particles from packaging) that pass through the cells
  2. L2: Storage cells, one bit each: charge or no charge on a DRAM capacitor behind an access transistor, a six-transistor SRAM flip-flop, electrons on a flash floating gate; historically a recirculating pulse in mercury, a charge spot on a CRT, a magnetized ferrite core
  3. L3: Arrays and their circuits: rows (wordlines) and columns (bitlines), sense amplifiers that detect, latch and write back a whole row, refresh that rewrites leaking cells, and the codes that detect and correct flipped bits
  4. L4: Addressable words: cells grouped into words of fixed length, each reached by a binary address; chips and modules (DIMMs) on a memory channel, run by a memory controller
  5. L5: The memory hierarchy: registers, caches (SRAM), main memory (DRAM), mass storage (SSD, disk), nearline and offline storage; each level smaller and faster than the next; locality keeps what is in use near the processor; hardware moves data between memory and caches
  6. L6: Memory management: virtual memory (pages, page tables, the MMU's address translation, page faults and replacement), protection of each program's region, working sets; done by the operating system with hardware help
  7. L7: The stored program and its data held for immediate use: the instructions the processor fetches and the data those instructions fetch and store, kept available, buffered and protected

the boardroom (shared floors)

  1. L4: Addressable words, chips, modules, caches' arrays and the memory controller: where requests, writes, refresh, error checks, wear counts, page-ins and the access history meet as addressed words
  2. L5: The memory hierarchy: registers, caches, main memory, mass storage; locality and the hardware's movement of data between levels
  3. L6: Memory management: virtual memory and paging, the MMU's translation, protection, replacement, error handling (an uncorrectable error shuts the machine down), the flash translation layer's wear leveling
  4. L7: The stored program and its data held for immediate use

way in: Access history (the reference stream, read for locality)

  1. L1: The stream of references a running program makes to addresses
  2. L2: Recently used page table entries and lines kept in small fast stores (TLB, caches)
  3. L3: Replacement order: which page or line is least recently used

way in: Disturbance from neighbouring cells (RowHammer)

  1. L1: A neighbouring row's wordline toggled again and again (hammered)
  2. L2: Vulnerable cells lose charge faster than the 64 ms refresh can restore
  3. L3: Bits flipped in physically adjacent rows at predictable places

way in: Pages in from mass storage

  1. L1: Pages held in a swap file or the program's file on mass storage
  2. L2: A reference to a page not in real memory raises a page fault
  3. L3: The page copied into a free frame and its table entry updated

way in: Particles (cosmic-ray neutrons, alpha particles)

  1. L1: Energetic neutrons (about 95% of soft-error particles at the surface) and alpha particles from radioactive traces in packaging
  2. L2: Charge deposited near a cell changes its state
  3. L3: The upset stored until read; error-correcting codes detect and correct one or more bits

way in: Read requests (commands and addresses from the memory controller)

  1. L1: Commands and addresses issued by the memory controller across the memory channel
  2. L2: The addressed row opened (activated): its wordline switches the row's access transistors onto the bitlines
  3. L3: The row's charges sensed, amplified and latched by the sense amplifiers; a column address selects the word

way in: Refresh timing (the controller's refresh commands)

  1. L1: Refresh commands from the controller, every row within 64 ms (in a burst, or one row every 7.8 us for 8,192 rows)
  2. L2: The row read out
  3. L3: Charge restored by the sense amplifiers' write-back

way in: Supply power

  1. L1: The supply voltage
  2. L2: Volatile cells hold their state only while powered (and refreshed); on power loss DRAM contents decay over seconds, non-volatile cells keep theirs

way in: Temperature

  1. L1: The temperature of the chips and their surroundings
  2. L2: Cell charge decays faster when warm, slower when cold (under 1% in 10 minutes unpowered at about -50 C); in the field its effect on error rates is small

way in: Wear (program/erase cycles on flash)

  1. L1: Program and erase operations, electrons injected into and ejected from the floating gate through the tunnel oxide
  2. L2: Charge trapped in the oxide with each cycle; the oxide's insulation degrades and the cell leaks
  3. L3: The raw bit error rate rises with P/E cycles until ECC can no longer correct

way in: Writes (data words to store)

  1. L1: Data words driven onto the bus to be stored
  2. L2: The column's sense amplifier forced high or low, charging or discharging the cell capacitor (DRAM); electrons injected onto a floating gate (flash)
  3. L3: The whole open row written back, the new bit among them

way out: Charge restored to the cells (write-back and refresh)

  1. L1: Charge back on the capacitors
  2. L3: Sense amplifiers drive the bitlines to full rails
  3. L4: The controller's refresh command, or a read of the row

way out: Errors reported (corrected, or uncorrectable and fatal)

  1. L1: A crash or a shut-down, or corrupted data used
  2. L4: Correctable and uncorrectable errors counted at the controller and logged
  3. L6: An uncorrectable error judged serious enough to stop the machine

way out: Heat

  1. L1: Heat; at least kT ln 2 for each bit erased
  2. L4: Every access, transfer and refresh draws power; DRAM draws it even when idle

way out: Pages written out to mass storage

  1. L1: Modified frames written back to auxiliary storage
  2. L6: A page frame chosen for reclaim by the replacement algorithm

way out: Remanence (contents left readable after power-off)

  1. L1: A full memory image readable by anyone with physical access (cold boot)
  2. L2: Cells keep most of their charge for thousands of refresh intervals unpowered, longer when cold

way out: Words read out to the processor

  1. L1: The word sent across the memory channel to the controller and into a cache or register
  2. L4: The column's latched bit connected to the data bus
  3. L7: An instruction or datum asked for

Sources (27) published science

Manchester Baby (Wikipedia, rev 1375332732, 2026-09-17, via Wikimedia Enterprise)
Computer memory (Wikipedia, WME snapshot)
Dynamic random-access memory (Wikipedia, WME) (behind its links)
Delay-line memory (Wikipedia, rev 1365416787, 2026-07-22, via Wikimedia Enterprise) (behind its links)
Wikipedia: Soft error (behind its links)
Denning PJ. The locality principle. Communications of the ACM 48(7) (2005) (behind its links)
Cai Y, Ghose S, Haratsch EF, Luo Y, Mutlu O. Error characterization, mitigation, and recovery in flash-memory-based solid-state drives. Proceedings of the IEEE (September 2017); arXiv:1706.08642 (behind its links)
Mutlu O, Kim JS. RowHammer: a retrospective. arXiv:1904.09724 (2019) (behind its links)
Memory hierarchy (Wikipedia, WME snapshot) (behind its links)
Virtual memory (Wikipedia, WME) (behind its links)
Magnetic-core memory (Wikipedia, WME) (behind its links)
Bérut, Petrosyan & Ciliberto, Information and thermodynamics: Experimental verification of Landauer's erasure principle, arXiv:1503.06537 (J. Stat. Mech. 2015) (behind its links)
Mutlu O, Ghose S, Gomez-Luna J, Ausavarungnirun R. Processing data where it makes sense: enabling in-memory computation. arXiv:1903.03988 (2019) (behind its links)
Liu J, Jaiyen B, Veras R, Mutlu O. RAIDR: retention-aware intelligent DRAM refresh. ISCA 2012 (behind its links)
Kim, Daly, Kim, Fallin, Lee, Lee, Wilkerson, Lai and Mutlu, Flipping Bits in Memory Without Accessing Them: An Experimental Study of DRAM Disturbance Errors, ISCA 2014 (doi:10.1145/2678373.2665726) (behind its links)
Schroeder B, Pinheiro E, Weber W-D. DRAM errors in the wild: a large-scale field study. SIGMETRICS/Performance '09 (2009) (behind its links)
Halderman JA, Schoen SD, Heninger N et al. Lest we remember: cold boot attacks on encryption keys. USENIX Security 2008 (behind its links)
Buko T, Tuczko N, Ishikawa T. DNA data storage. BioTech (Basel) (2023). PMC10296570 (behind its links)
Flash memory (Wikipedia, WME) (behind its links)
von Neumann, First Draft of a Report on the EDVAC (Moore School, 30 June 1945); sections 1-4 as excerpted in Aspray, Papers of John von Neumann on Computing and Computer Theory (MIT Press 1987) (behind its links)
Computer hardware (Wikipedia, via Wikimedia Enterprise snapshot) (behind its links)
Williams tube (Wikipedia, rev 1369816839, 2026-08-17, via Wikimedia Enterprise) (behind its links)
Call stack (Wikipedia, via Wikimedia Enterprise snapshot 2026-10-01) (behind its links)
Von Neumann architecture (Wikipedia, rev 1381757639, 2026-10-24, via Wikimedia Enterprise) (behind its links)
Jouppi et al., In-Datacenter Performance Analysis of a Tensor Processing Unit (arXiv:1704.04760; ISCA 2017) (behind its links)
Sorting algorithm (Wikipedia, via Wikimedia Enterprise snapshot 2026-11-01) (behind its links)
Asorey & Mayo-García, Calculation of the High-Energy Neutron Flux for Anticipating Errors and Recovery Techniques in Exascale Supercomputer Centres, arXiv:2212.07770 (2022) (behind its links)