Electronics Guide

Every computer ever built has needed somewhere to put the data it is not using at this instant. That requirement has never changed. Almost nothing about the way it has been met has stayed the same. In seventy-five years the physical carrier of a stored bit has been a pressure wave in liquid mercury, a patch of charge on the face of a cathode-ray tube, a magnetized ring of ferrite threaded by hand, a track on a spinning cylinder, a circle on a stack of aluminum platters, a stripe on a ribbon of coated plastic, and finally a few dozen electrons trapped on an isolated gate. The bit survived; every one of its houses was demolished.

This genealogy follows that succession. The interesting question in any lineage is not which device was largest or fastest but which constraint forced the change, and the constraints here arrived in a definite order. Capacity came first, because the earliest electronic stores held a few hundred words at ruinous cost. Latency came second, once capacity was cheap enough that waiting for it became the dominant expense. Reliability came third, when mechanisms grew precise enough that a speck of dust destroyed them. Power and physical form came fourth, when computers left the machine room and had to run on a battery in a pocket. Each transition described below answers one of those four questions, and each answer created the conditions for the next.

A second pattern runs through the whole family and contradicts the usual language of technological replacement. Storage tiers are almost never deleted. They are displaced upward: a new technology occupies the fast and expensive position, and the incumbent slides down into the bulk position. Magnetic tape, first attached to a computer in 1951, is still sold in new generations and still sits beneath disk. The companion articles on vacuum tube to transistor to IC and mechanical to electronic to digital follow the processing side of the same machines; this one follows what those machines remembered.

Reading Storage as a Lineage

A storage technology is defined by six numbers, and none has ever been best at all of them: cost per bit, access latency, sustained bandwidth, retention through a power failure, areal or volumetric density, and energy per stored bit at rest. Gains along one axis are routinely bought by concessions along another, which is why several technologies coexist rather than one winning outright.

That trade structure produces the memory hierarchy, and the hierarchy is the most durable inherited trait in this family. It appeared in recognizable form in the mid-1950s, when fast random-access working store separated cleanly from slow bulk store, and it has been rebuilt at every generation since with different components in the same slots. The names have all changed. The shape has not.

The Volatile Ancestors: Delay Lines and Electrostatic Tubes

The first electronic stores were built because the alternatives were worse. Punched cards held data permanently but could not be read at anything near the speed of vacuum-tube arithmetic, and flip-flops built from tubes cost several tubes per bit, which put a store of even a thousand words beyond any budget. The two technologies that broke the impasse in the late 1940s both worked by refusing to hold data still.

J. Presper Eckert developed the mercury delay line at the Moore School of Electrical Engineering at the University of Pennsylvania, drawing on wartime radar work in which a delayed echo was used to cancel stationary clutter. A quartz transducer at one end of a tube of mercury turned an electrical pulse train into an acoustic wave, a second transducer at the far end turned it back, and an amplifier reshaped the pulses and injected them at the input again. The data existed as sound in transit. Mercury was chosen because its acoustic impedance is close to that of quartz, which minimized reflections, and because sound travels through it at roughly 1,450 meters per second, slow enough that a tube of practical length held a useful number of bits. The tanks were held near 40 degrees Celsius, because the speed of sound in mercury varies with temperature and a drifting delay corrupts the bit timing.

EDSAC, operational at Cambridge in 1949, used sixteen tanks holding 256 words of thirty-five bits between them. UNIVAC I, delivered to the Census Bureau in 1951, held one thousand words in mercury with an average access time of about 222 microseconds. That last figure conceals the defining weakness. A delay line is sequential: a word is available only when it arrives, and a program that wants the word which has just passed waits a full circulation. Programs were written around the memory's rhythm rather than around the logic of the problem.

The Williams-Kilburn tube, developed by Freddie Williams and Tom Kilburn at the University of Manchester and covered by British patents filed in December 1946 and October 1947, attacked the same problem from the opposite side. Writing a bit meant firing the electron beam of an ordinary cathode-ray tube at a spot on the phosphor, knocking secondary electrons loose and leaving a charge well whose presence or absence encoded the value. A pickup plate outside the glass detected the disturbance when the beam revisited the spot. Because the beam could be deflected anywhere on the screen in a fixed time, the store was genuinely random access, which was a real advance over the delay line. A typical tube held between 256 and 2,560 bits, and the charge leaked away within a fraction of a second, so the whole screen had to be scanned and rewritten continuously.

The Manchester Baby ran the first stored program on such a tube on June 21, 1948, and the design spread to the Ferranti Mark 1, the Institute for Advanced Study machine, the UNIVAC 1103, and the IBM 701. It also acquired a reputation for temperament, since tubes drifted with age and most installations needed hand tuning by an engineer who knew that particular machine, as the account under computer industry origins records.

Both branches terminated, and for the same reason. Neither had a path to higher density, because the bit occupied a volume or an area fixed by physics that did not shrink, and both lost everything within milliseconds of the refresh stopping. What they contributed was not hardware but a principle. Storage as a maintained process, rewritten faster than it decays, is exactly how dynamic memory works, and the destructive read followed by a restoring write that core memory made routine is the same idea in another form. The ancestors of modern dynamic memory technologies are conceptual rather than physical.

The Magnetic Drum: Latency as a Function of Rotation

Magnetic recording entered computing from a different direction, and it entered early. Gustav Tauschek patented a magnetic drum store in Austria in 1932, long before any machine needed one. The idea was resurrected after the war by Engineering Research Associates, which built drums for the United States Navy from 1946 and made one the internal memory of the machine sold commercially as the ERA 1101.

A drum is a rotating cylinder coated with magnetic material, with a row of fixed read-write heads along its length, one per track. Because every head already sits over its track, there is no seek: the only wait is for the required sector to rotate past. That single architectural fact defined the technology. Latency was purely rotational, bounded by the rotation period, and entirely predictable. The IBM 650, announced in 1953 and delivered from 1954, made the drum its main memory, holding one thousand, two thousand, or four thousand signed ten-digit words depending on model, spinning at 12,500 revolutions per minute for an average access of roughly 2.5 milliseconds.

Predictable latency invited a programming style with no modern equivalent. Because the 650 fetched an instruction only when it came around, a programmer who placed the next instruction at the address arriving just as the current one finished ran several times faster than one who laid the program out sequentially. That technique, called optimum programming, is the first appearance of a theme that never goes away: arranging data to suit the physical timing of the medium, as disk scheduling and flash page mapping still do.

What killed the drum was geometry. A drum records only on the curved outer surface of a cylinder, and every extra track needs another head and another channel of electronics, while the disk records on the flat faces of a stack, where area grows with the square of the radius and one moving head serves every surface. Drums nonetheless persisted into the 1970s in real-time and military systems that valued their bounded access time and tolerance of vibration, which is the usual fate of a displaced technology: retreat into the niche defined by its one remaining advantage.

Core Memory and the Separation of Tiers

The decisive event of the mid-1950s was not a new bulk store but the arrival of a working store good enough to make bulk storage a separate problem. Jay Forrester's group at the Massachusetts Institute of Technology developed magnetic core memory for the Whirlwind computer and put it into service in 1953, replacing electrostatic storage that had been the machine's chief source of downtime. A core is a small ring of ferrite with a rectangular hysteresis loop, magnetized clockwise or counterclockwise. Coincident-current selection drives half the switching current down one wire of a grid and half along the other, so only the core at the intersection sees enough field to flip, and a sense wire threaded through the plane reports whether a flip occurred.

Reading was destructive, since sensing forced the core to a known state, which is why core specifications quote a short access time and a longer full cycle time that includes the restoring write. Manufacturing was brutally labor-intensive, each ring threaded with several wires by hand under magnification. The economics were nevertheless favorable, because core was fast, unaffected by a power failure, and reliable enough that memory faults stopped dominating the maintenance log. IBM shipped it in the 704 from 1955 and offered it as a retrofit for the Williams tubes of the 701.

Core's contribution to this genealogy is structural rather than technological. Before core, the boundary between working store and bulk store was blurred: the 650's drum was its main memory, and UNIVAC's delay lines were both the fast store and, in practice, the only store. After core, a machine had a small, fast, uniformly accessible working memory and a large, slow, mechanically served bulk memory, and the two were designed by different people to different criteria. Every later architecture inherits that split. Core held the upper position for about two decades, until semiconductor memory, beginning with the Intel 1103 dynamic RAM in 1970, undercut it on cost and density. Notice what the successor took from each ancestor: from core, the role and the addressing structure; from the Williams tube, by way of the same physics of leaking charge, the requirement to be refreshed.

RAMAC: The Moving Head and the Disk Stack

IBM's San Jose laboratory, established under Reynold Johnson in 1952 with a brief to work on storage, announced the 305 RAMAC in September 1956. The acronym stood for Random Access Method of Accounting and Control, and the name identified the point of the machine. Its 350 disk storage unit held fifty aluminum disks, twenty-four inches in diameter, coated with iron oxide and stacked on a common spindle. One hundred recording surfaces, each carrying one hundred tracks of five hundred characters, gave five million characters in total, roughly 3.75 megabytes in the six-bit encoding of the day. Average access time was about 600 milliseconds, and the system rented for approximately 3,200 dollars a month.

By modern measures those numbers are unimpressive, and the areal density, usually cited at about two thousand bits per square inch, is roughly a billion times below what a current drive achieves. The importance of the 350 lies entirely in its access mechanism. A single arm assembly carrying a pair of heads moved vertically to select a disk, then drove horizontally into the gap between two platters, floating the heads on a cushion of compressed air. One mechanism therefore served one hundred surfaces.

That is the branch point of the entire lineage, and it is worth stating precisely why. On a drum, or any fixed-head device, capacity and head count rise together, so electronics cost scales with storage. On a moving-head disk stack, capacity is a function of recording area while cost is dominated by one mechanism, one actuator, and one channel. Adding a platter adds surface almost for free. Every subsequent decade of the disk industry consists of exploiting that decoupling: more platters, more tracks per inch, and more bits per inch, all served by the same single actuator.

The consequences above the storage were as large as those within it. Tape and card systems forced batch processing, in which transactions were accumulated, sorted into the order of the master file, and applied in a single pass. A disk file made it possible to find and update one record on demand, which is what inventory control, airline reservation, and account inquiry require, and which was the technical precondition for the interactive computing traced under batch to time-sharing to personal to cloud.

From Removable Packs to the Sealed Winchester

The fifteen years after RAMAC refined the mechanism rather than reconsidering it. The IBM 1301 of 1961 replaced externally supplied compressed air with a self-acting air bearing, in which the rotation of the disk drags a film of air under a shaped slider and lifts the head. Removing the air supply removed a subsystem and, more importantly, let the flying height fall, because the bearing could now be engineered rather than merely supplied.

The IBM 1311 of 1962 introduced the removable disk pack, a cartridge of six disks holding two million characters that an operator could lift out and put on a shelf. This restored to disk the property tape had always had, that the medium is separable from the mechanism, so capacity could be bought by the pack rather than by the drive. Packs dominated commercial installations for a decade, and the practice of keeping one pack per application, mounted on demand, shaped the operating systems of the period.

Removability carried a defect that grew more serious as flying heights fell from thousandths of an inch toward millionths. An open drive breathes room air. The clearance under a head became far smaller than a smoke particle or a fingerprint ridge, and anything arriving under a head caused a crash that destroyed the head, the surface, and frequently every other surface in the pack. Filtration helped; the fundamental exposure did not go away.

IBM's answer, developed at San Jose under Kenneth Haughton and announced in 1973 as the 3340, sealed the heads and disks together in a filtered assembly that was never opened in the field. The project's code name, Winchester, came from an early configuration of two thirty-megabyte modules, which suggested the .30-30 Winchester rifle cartridge; the shipping data modules held thirty-five and seventy megabytes. Sealing bought three things at once. Contamination control allowed a lower and more consistent flying height, which directly raises areal density. The head and the disk became a matched pair rather than parts built to a specification that any vendor's pack had to meet. And the heads could be made lighter and loaded more gently, landing on a lubricated zone when the disk stopped, an arrangement known as contact start-stop.

The 3340 still permitted removability, since the sealed data module including its heads was the removable unit, but that compromise did not last. The sealed head-disk assembly became a permanent, non-serviceable component, and the disk drive stopped being an instrument that read interchangeable media. The constraint driving this transition was reliability rather than capacity or latency, and the price paid was removability, which the lineage handed permanently to tape and, in small quantities, to the floppy disk.

Magnetic Tape: The Tier Beneath

Tape reached computers before disk did and has outlived every prediction of its demise. The UNIVAC I used the UNISERVO drive from 1951, recording on nickel-plated bronze tape, and IBM introduced the 726 with the 701 in 1952, recording on half-inch plastic tape at 100 to 200 bits per inch. Open reels gave way to the enclosed cartridge with the IBM 3480 in 1984, which removed the threading operation and made robotic libraries practical.

The modern line descends from the Linear Tape-Open format, which Hewlett-Packard, IBM, and Seagate defined in the late 1990s as an explicitly multi-vendor alternative to the proprietary cartridge formats then competing. LTO-1 shipped in 2000 with one hundred gigabytes of native capacity. LTO-9 arrived in 2021 at eighteen terabytes native, and LTO-10 followed in 2025 at thirty terabytes native with a transfer rate of about 400 megabytes per second. The consortium publishes a roadmap several generations ahead, which is itself a competitive instrument, since a customer buying a robotic library commits to a decade of media purchases and wants evidence that the format will still exist.

Tape survives because its weaknesses do not matter in the position it occupies. Access is sequential and a mount plus seek takes tens of seconds, which disqualifies it from any interactive role and is irrelevant for a restore that was scheduled anyway. Its advantages are three: the medium is separate from the mechanism, so a growing archive buys cartridges rather than drives; a cartridge on a shelf consumes no power, which matters enormously at petabyte scale; and a cartridge that is not in a drive cannot be reached over a network, which has given tape a second career as protection against ransomware under the name air gap.

The genealogical reading is that tape did not lose to disk. It was displaced downward into the archive tier, where the property that defines that tier, cost per stored bit including the cost of keeping it powered, is one it still wins.

The Floppy Disk: Storage as a Distribution Medium

The floppy disk began as the answer to a problem that was not about capacity at all. IBM's System/370 processors were microcoded, and the microcode had to be loaded at power-on from something cheaper and more easily updated than the read-only stores then in use. The answer, shipped in 1971, was an eight-inch flexible disk of coated Mylar in a paper envelope, read-only, holding about eighty kilobytes. A read-write version followed in 1973 with the 3740 data entry system, holding roughly 243 kilobytes and intended to replace the punched card in data preparation.

The format left IBM quickly. Shugart Associates produced the SA400 five-and-a-quarter-inch drive in 1976, marketed at 110 kilobytes, in response to a request for something smaller and cheaper than the eight-inch unit, and it arrived exactly when the microcomputer did. Sony's three-and-a-half-inch design, with a rigid shell and a sliding metal shutter that protected the medium from handling, reached double-sided 720-kilobyte capacity in 1984 and the familiar 1.44-megabyte high-density format from 1986. Sony, the last major manufacturer, ended production in March 2011.

What the floppy contributed was not capacity, in which it was always poor, but interchange. The personal computer created a need no previous class of machine had: software written by one company had to reach hundreds of thousands of machines owned by people with no technical staff, and data had to move between machines that shared no network. A medium that could be duplicated in bulk at a factory, mailed, and read by any drive of the same format solved that problem, and ubiquity of format mattered far more than capacity.

The successors in that role show how a function can survive the technology that performed it. Optical media took it first, since a compact disc held four hundred times as much. The USB flash drive took the interchange half in the early 2000s, and the network took the rest at effectively zero marginal cost, which is why the medium vanished rather than shrinking into a niche. The lineal descendant of the floppy disk is not a disk at all; it is the download. The consumer end of this story continues under digital storage and media.

Read Heads and the Areal Density Regimes

Between 1956 and the present, the areal density of magnetic disk recording rose by a factor of roughly one billion. No single invention accounts for that. What accounts for it is a sequence of four sensing technologies, each lifting a ceiling the previous one had reached, and it is a rare case of a component being replaced outright three times while the system around it kept its shape.

Inductive Heads

The original head was a small electromagnet: a ferrite core with a gap and a winding, which wrote by driving current through the coil and read by detecting the voltage a passing magnetic transition induced in the same coil. The reciprocity is elegant and the limitation is fundamental. Induced voltage is proportional to the rate of change of flux, so it falls as bits get smaller and as the medium moves more slowly past the head. IBM's 3370 of 1979 introduced thin-film heads, whose pole pieces and coil are deposited and patterned photolithographically on a wafer, and that step changed the industry's manufacturing base: disk drive companies acquired semiconductor process lines and began to treat heads as chips.

Magnetoresistive Heads

The escape from the velocity problem was to stop using induction for reading. A magnetoresistive element changes its electrical resistance in the presence of a magnetic field, so the read signal depends on the field itself rather than on the rate at which it changes, and it does not weaken as bits shrink or the disk slows. IBM used magnetoresistive readers in the 3480 tape drive in 1984 and brought them to disk in 1991, in a one-gigabyte three-and-a-half-inch drive developed under the name Corsair. The architectural consequence outlasted the sensor: because the reader was no longer the writer, the two could be optimized independently, and every head since has carried an inductive writer and a separate resistive reader on one slider.

Giant and Tunneling Magnetoresistance

Albert Fert and Peter Grünberg independently discovered giant magnetoresistance in 1988, in multilayer stacks of magnetic and non-magnetic films whose resistance changes by a much larger fraction than any ordinary magnetoresistive material, and shared the 2007 Nobel Prize in Physics for it. Engineered into a spin valve, in which one magnetic layer is pinned and the other rotates with the external field, the effect produced heads sensitive enough for the densities of the late 1990s, and drives using them reached the market in 1997. Under a decade from laboratory discovery to shipping product is unusually fast, and it happened because the disk industry had an acute and quantified need for exactly that sensitivity.

Tunneling magnetoresistance replaced it in the following decade, separating the two magnetic layers with a thin insulating barrier, commonly magnesium oxide, and passing the sense current through that barrier by quantum tunneling for a larger resistance change still. The same physics underlies magnetoresistive random-access memory, so the read head in a hard disk and a leading candidate for a universal non-volatile memory are close relatives, and the films involved belong to the wider subject of magnetic materials.

Perpendicular Recording and the Thermal Wall

For half a century, disks recorded longitudinally, with the magnetization of each bit lying in the plane of the medium and pointing along the track. As bits shrink, adjacent longitudinal bits present opposing poles to one another and the demagnetizing field grows, so the medium fights the very state it is asked to hold. Shun-ichi Iwasaki at Tohoku University argued from the mid-1970s that orienting the magnetization perpendicular to the surface reverses that relationship, since neighboring perpendicular bits assist rather than oppose each other, and that the geometry also permits a soft magnetic underlayer beneath the recording film to complete the write flux path and effectively double the field available at the medium.

The argument was accepted long before it was adopted, because a transition required new media, new heads, and new channel electronics at once. What forced the change was the superparamagnetic limit. A recorded bit consists of many small magnetic grains, and its stability against thermal agitation depends on the product of grain volume and magnetic anisotropy. Shrink the grains for density and the stored energy per grain falls toward the thermal energy at operating temperature, at which point bits reverse spontaneously and data decay on the shelf. Raising anisotropy restores stability but demands a stronger write field, while using fewer, larger grains per bit degrades the signal-to-noise ratio. Those three requirements cannot all be met at once, a deadlock known in the field as the magnetic recording trilemma.

Perpendicular recording bought roughly a decade of continued scaling by improving the write field and the medium geometry together. Toshiba shipped the first commercial perpendicular drive, a 1.8-inch unit, in 2005; the rest of the industry followed within about two years, and Hitachi reached one terabyte in a 3.5-inch drive in 2007. Shingled magnetic recording then extracted further density by overlapping adjacent tracks like roof shingles, so that writing one track partially erases its neighbor and whole zones must be rewritten in order. That trade converts the disk into an append-oriented device and pushes the complexity into firmware or into the host filesystem, which is precisely the bargain flash makes, and it is a useful reminder that the same constraint tends to produce the same architecture in unrelated technologies.

The current attack on the trilemma is to make the medium easier to write only at the instant of writing. Heat-assisted magnetic recording integrates a laser and a near-field optical transducer into the head, heating a spot a few tens of nanometers across for a fraction of a nanosecond so that a very high anisotropy medium, typically an iron-platinum alloy, becomes writable and then cools back into thermal stability. Seagate states that its heat-assisted platform is shipping in volume to large cloud customers at capacities up to forty-four terabytes, with an areal density near 2.5 terabits per square inch; those are the manufacturer's figures. Western Digital has pursued energy assistance by microwave excitation instead. Both confirm the pattern established by the read heads: the limit sits at the interface between head and medium, and each generation moves it by changing what the head can do rather than what a bit is.

Flash and the Solid-State Transition

Every technology described so far stores data as a physical configuration that some mechanism must visit. Flash memory ended that. Its ancestry runs through semiconductor devices rather than through recording, which is why the arrival of solid-state storage is not a refinement of the disk line but the entry of a different family into the same ecological position.

The Floating Gate

Dawon Kahng and Simon Sze proposed the floating-gate MOSFET at Bell Laboratories in 1967. An additional gate, completely surrounded by insulator, sits between the control gate and the channel of an ordinary field-effect transistor. Charge placed on that isolated gate shifts the transistor's threshold voltage, and because the gate is isolated the shift persists for years without power. Retention comes from the height of the oxide barrier; endurance is limited because the tunneling used to move charge across that barrier gradually damages it. Two commercial generations followed before flash: the erasable programmable read-only memory, sold by Intel from 1971 as the 1702 and erased by ultraviolet light through a quartz window, and the electrically erasable version, which needed extra transistors per cell for byte-level erasure and so had a large and costly cell. Both were code stores, not mass storage.

Masuoka's Bargain

Fujio Masuoka's group at Toshiba, working from about 1980, made the trade that turned the floating gate into a storage medium. If erasure is performed on a large block rather than byte by byte, the per-cell erase circuitry disappears and the cell shrinks dramatically. The name came from a colleague, Shoji Ariizumi, who observed that erasing a whole block at once resembled a camera flash. The team presented the NOR configuration at the International Electron Devices Meeting in 1984 and the NAND configuration at the same conference in 1987; Intel introduced the first commercial NOR chip in 1988, and Toshiba brought NAND to market shortly after.

The two configurations are the branch point of the family. NOR cells connect in parallel to the bit line, so any byte can be read directly and quickly and a processor can execute code straight from the chip. NAND cells sit in series strings, which removes most of the contacts and makes the cell far smaller, at the cost of access that is page-oriented rather than random. NOR became the firmware store of embedded systems and NAND became the mass storage medium, and the descendants of each line have stayed in those roles ever since, a contrast developed under non-volatile storage.

More Bits in the Same Cell

Because the stored quantity is an analog charge, the threshold window can be divided into more than two regions. NEC demonstrated a chip storing two bits per cell in 1998, three bits per cell reached mass production around 2010, and four-bit cells appeared in products over the following decade. Each additional bit multiplies capacity by a smaller factor than the last while dividing the voltage margin between adjacent levels, so error rates rise steeply and endurance falls. The response was to make the controller work harder, with low-density parity-check decoding replacing simpler codes and read-retry strategies stepping the reference voltage through trial positions when the first read fails.

The Third Dimension

Planar scaling then stopped for a reason with no analogue in logic. As the cell shrinks, the number of electrons separating one stored level from the next falls toward a few tens, and at that point the loss of a handful of electrons is a bit error. The industry stopped shrinking and started stacking. Toshiba proposed a stacked architecture in 2007, Samsung commercialized a twenty-four-layer part in 2013, and modern devices stack hundreds of layers, often by bonding two stacks together, with control circuitry moved underneath the array. Density now advances by layer count rather than by lithographic shrink, the same reorientation from two dimensions to three that has appeared in packaging and in logic.

What Solid State Changed About the Problem

Replacing a mechanism with a semiconductor deleted one set of problems and created a different one, and the exchange was not symmetrical. What disappeared was everything that depended on motion. There is no seek, so random access costs what sequential access costs, and the elevator scheduling and contiguous-allocation strategies built into decades of operating systems became largely irrelevant. There is no rotational latency, no spin-up delay, no head crash, and no sensitivity to shock or orientation while operating. Access latency fell from milliseconds to tens of microseconds, a change of two orders of magnitude that no refinement of a mechanism could have produced.

What appeared in its place is a mismatch among the three operations NAND supports. The device reads a page, programs a page, and erases only a much larger block, and a programmed page cannot be reprogrammed until its whole block is erased. The medium therefore cannot be overwritten in place, which is exactly what every filesystem written for disk assumes. Bridging that gap is the work of the flash translation layer, which maps logical addresses to physical pages, writes updates to fresh pages while marking the old ones invalid, and reclaims blocks by copying out the pages still in use.

Two quantities govern the result. Write amplification is the ratio of physical writes performed by the device to logical writes requested by the host, and it rises as the device fills, because reclaiming a block relocates more valid data. Endurance is the number of program-erase cycles a block survives, and the commonly quoted orders of magnitude run from tens of thousands for single-level cells to a few thousand at two bits per cell, roughly a thousand at three, and a few hundred at four. Wear leveling spreads erasures evenly, over-provisioning reserves hidden capacity to keep reclamation efficient, and the TRIM command lets the host say which logical blocks the filesystem has abandoned. These mechanisms are examined under flash controllers.

The interface had to change as well. Storage protocols built for disks assumed a device that served one request at a time after a mechanical delay, and their command overhead was invisible against a ten-millisecond seek but dominant against a fifty-microsecond flash read. NVM Express, first published in 2011, replaced that model with many deep command queues carried directly over PCI Express, matching both the parallelism of an array of NAND dies and the parallelism of a multicore host.

The deepest change is easy to overlook. A hard disk is a passive medium plus a servo system, and the intelligence sits in the host. A solid-state drive contains a multicore controller, its own dynamic memory, error-correction engines, and what amounts to a log-structured filesystem in firmware, all running independently of the host. Storage became a computer, with its own scheduling, its own failure modes, and its own opaque decisions about where data actually live, and much of the difficulty of modern storage systems engineering follows from that opacity.

Branches That Ended and Why

A genealogy that records only survivors explains nothing, because it cannot say what distinguished them. Three terminated branches are instructive here, and they failed for three different reasons.

Magnetic bubble memory was the most heavily funded. Andrew Bobeck's group at Bell Laboratories developed it from 1967, exploiting small cylindrical magnetic domains that can be moved through a thin garnet film by a rotating in-plane field, so that data circulate through a shift register with no moving parts at all. Texas Instruments sold the first commercial device in 1977, and Intel followed with a one-megabit part, the 7110, in 1979. The properties bubbles offered were exactly the ones flash would later sell: non-volatility, silence, immunity to shock, and no mechanism to wear out. They lost because they sat between two tiers whose incumbents improved faster. Semiconductor memory above them fell in price and rose in speed, disk below them fell in price per bit faster still, and bubble memory, dearer per bit than disk and far slower than memory, had no defensible position. Most manufacturers withdrew by 1981.

The magnetic drum failed for a structural reason rather than a competitive one. Its capacity scaled with the surface of a cylinder and its cost with the number of tracks, because every track needed a head, while the disk stack scaled with area and amortized one mechanism across every surface. It is the cleanest case here of a branch terminated by geometry.

Holographic storage failed repeatedly, over five decades, for a reason that has become familiar in emerging storage. The physics is genuinely attractive, since data are recorded through the volume of a medium rather than on its surface and a whole page of bits is read in parallel in one optical exposure. What defeated it each time was everything around the physics: photosensitive media that shrank or degraded, alignment tolerances that demanded laboratory conditions, and the fact that magnetic and solid-state alternatives kept improving during the years each holographic program needed to reach a product. That history is examined under holographic storage, and the same pattern is visible in current work on DNA data storage.

The common lesson is that extinction in storage rarely means the technology stopped working. Bubble memory worked. Drums worked, and some ran for decades after they were obsolete. Extinction means failing to improve as fast as the neighboring tiers, and the most dangerous position is the middle of the hierarchy, where a technology is attacked from above on latency and from below on cost. That argument generalizes well beyond storage and is developed under failed technologies and obsolescence.

Displacement Upward: Why No Tier Disappears

Set the transitions side by side and the recurring shape becomes plain. Core memory did not delete the drum; it took the fast working-store position and left the drum as intermediate bulk. Disk did delete the drum, because it offered the same random access on a fundamentally better scaling law, but it did not delete tape, which slid down into archive. Flash did not delete disk; it took the working-store and transactional positions and pushed disk toward capacity and archival tiers, which in turn pushed tape down into deep archive. At each step the incumbent lost the tier it had occupied and kept a lower one.

The economics explain why this is normal. Latency and cost per bit improve in different currencies, and a technology that wins decisively on one rarely wins on the other at the same moment. Flash reads in tens of microseconds and disk in milliseconds, but disk still stores a bit for a fraction of the price, and tape for a fraction of that while consuming nothing at rest. As long as the cost ratio between two tiers exceeds the value of the latency difference for some class of data, both tiers pay for themselves.

The exceptions are informative because they are so few. A tier vanishes when its distinguishing property becomes free somewhere else. The drum's property was random access below the price of core, and the disk supplied it more cheaply. The floppy's property was interchange, and the network supplied it at effectively zero marginal cost. In both cases the technology disappeared quickly and completely rather than retreating into a niche, because no residual advantage remained to retreat into.

The current frontier is the gap between dynamic memory and flash, roughly three orders of magnitude wide in latency and similarly wide in cost per bit, into which several technologies have tried to insert themselves. Persistent memory built on phase-change materials reached commercial products and was then withdrawn; magnetoresistive and resistive memories occupy specialized positions; and standards that allow memory to be attached over a switched link rather than a dedicated bus are reopening the question of where the boundary between memory and storage should fall. The candidates are surveyed under advanced memory and storage. The genealogical prediction is modest but usually correct: whatever occupies that tier in fifteen years is already working in a laboratory today, and it will displace its neighbors rather than replace them.

Summary

Digital mass storage begins with two volatile technologies adopted because nothing better existed. Mercury delay lines stored data as sound in transit and delivered words only in sequence; Williams-Kilburn tubes stored charge on a phosphor and delivered genuine random access at the cost of continuous refresh and constant maintenance. What they passed on was the principle of storage as a maintained process, inherited today by dynamic memory. The magnetic drum then put a fixed head over every track and made latency purely rotational, and magnetic core memory, in service in Whirlwind from 1953, separated fast random-access working store from slow bulk store and held that upper position for two decades, establishing the hierarchy every later system has rebuilt with different parts.

The IBM 350 of 1956 introduced the moving head and the disk stack, decoupling capacity from head count and making random access to bulk data economical for the first time. Removable packs followed in 1962, and the sealed Winchester assembly of 1973 traded removability for contamination control, lower flying height, and matched head and media, the architecture every hard disk has used since. Four generations of read head, from inductive through magnetoresistive and giant magnetoresistive to tunneling sensors, each released a new density regime, and perpendicular recording followed by energy-assisted writing has kept the medium ahead of the superparamagnetic limit. Tape survived beneath all of it, from the IBM 726 of 1952 to LTO-10 at thirty terabytes a cartridge in 2025, and the floppy disk occupied a role that was never about capacity and was eventually absorbed by the network.

Flash arrived from the semiconductor side rather than the recording side. The floating gate of Kahng and Sze in 1967 became a mass storage medium when Masuoka's group at Toshiba traded byte-level erasure for block erasure, presenting NOR in 1984 and NAND in 1987, and the line has since advanced through multi-level cells and three-dimensional stacking. Solid state removed the seek, the rotation, and the mechanism, and added erase granularity, wear, write amplification, and a controller complex enough to be called a computer. Bubble memory, the drum, and repeated holographic programs mark the terminated branches, each defeated not by failing to work but by improving more slowly than its neighbors. The tiers themselves persist, displaced downward rather than deleted, which is the most reliable prediction this lineage supports.

Related Topics