Electronics Guide

Almost every byte that leaves a computer over a wire departs inside an Ethernet frame, and most of what leaves over the air departs in a frame that carries Ethernet's addresses. The frame format was fixed in a specification published in 1980, and the 48-bit addressing inside it has outlived nearly every other element of the design. Yet the medium beneath that frame has changed completely: the thick coaxial cable of the first standard is gone, the collisions the protocol was designed to resolve no longer occur on an ordinary switched link, and the data rate has risen by a factor of eighty thousand. Ethernet is therefore an unusually clear case of descent with modification, in which an inherited outer form survived while nearly every internal mechanism was replaced.

This genealogy follows that line from its conceptual root. The root is not a device but an idea: that a communication network should chop traffic into addressed blocks and interleave those blocks on shared links, rather than reserving a dedicated path for each conversation. The idea appeared in the early 1960s, was proved at national scale by the ARPANET, was adapted to radio by ALOHAnet, and was then compressed into a single building by Ethernet. From there it spread outward again, until packet switching over Ethernet became the default not only in offices but in factories, vehicles, telephone carrier networks, and the data centers that host the cloud.

The companion article on telegraphy to telephony to Internet follows the wide-area branch of this family, in which the descendants are protocols and carrier systems. This article follows the local branch, in which the descendants are cables, transceivers, and switching silicon.

The Ancestor Idea: Blocks, Packets, and Queues

The nineteenth-century telegraph already sent messages in discrete units and forwarded them hop by hop, with clerks receiving a message, storing it, and retransmitting it toward the next office. The telephone network that displaced it took the opposite approach, building a continuous electrical path for the duration of a call and holding that path idle through every pause in the conversation. Circuit switching suited voice and wasted capacity on data, whose traffic is bursty by nature: a terminal session consists of short exchanges separated by long silences while a person thinks or types.

Two Inventions and a Theory

Two researchers arrived at the resolution independently and for opposite reasons. Paul Baran, at the RAND Corporation, began in 1960 to design a communication system that would survive the destruction of many of its nodes. He proposed that traffic be divided into standard message blocks and routed adaptively through a redundant mesh, briefed the scheme to the Air Force in 1961, and published the report series On Distributed Communications in 1964. Donald Davies, at the National Physical Laboratory in Britain, reached a similar design in 1965 from the opposite motive, namely efficient interactive computing, and set it out in a 1966 proposal for a national data network. Davies supplied the word that endured, naming the units packets on a linguist's advice because the term would carry into other languages without distortion.

Leonard Kleinrock is often named beside them, and his contribution was real but different in kind. His doctoral research at the Massachusetts Institute of Technology, published as Communication Nets in 1964, applied queueing theory to store-and-forward message networks and predicted how shared links would behave under statistical load. Whether that work also constituted an independent invention of packet switching is genuinely disputed. Kleinrock has claimed priority; Davies wrote that he could find no evidence Kleinrock had grasped the principles of packet switching, and other ARPANET participants have contested the claim in print. The defensible statement, and the one most historians make, is that Baran and Davies invented packet switching independently while Kleinrock supplied much of the theory by which its behavior was analyzed and measured.

Both proposals shared one commitment: capacity would be allocated statistically rather than reserved. If each block carried its own destination address, blocks belonging to many conversations could interleave on one link, and the link would be busy whenever anyone had something to send. That single decision is the trait every descendant in this lineage inherits.

What the Idea Cost

Statistical sharing is not free. A circuit, once established, delivers a predictable delay; a shared link delivers a delay that depends on what everyone else is doing. Packet networks therefore had to acquire mechanisms that circuit networks never needed, including buffering, congestion control, retransmission, and reordering at the receiver. Much of the subsequent history of data networking consists of rediscovering that these costs are real and inventing new machinery to contain them, from the sliding windows of the 1970s to the priority shapers of contemporary industrial Ethernet.

ARPANET: The Idea Made Manufacturable

The Advanced Research Projects Agency of the United States Department of Defense funded the first large implementation. The design separated the network from the computers attached to it: each site received an Interface Message Processor, a ruggedized Honeywell DDP-516 minicomputer programmed by Bolt Beranek and Newman, which accepted data from the local host, divided it into packets, and forwarded them over 50 kilobit per second leased telephone lines toward the destination. The first node entered service at the University of California, Los Angeles, in 1969, and the first host-to-host message was sent from there to the Stanford Research Institute on October 29 of that year.

Two inheritances from the ARPANET matter for this lineage. The first is the interface processor itself, a dedicated machine whose only job was to move packets. Every network interface controller, bridge, and switch that followed is a descendant of that division of labor, in which packet handling is offloaded from the general-purpose processor to hardware built for the purpose. The second is the discipline of the leased line: the ARPANET rented capacity from the telephone company and could not change the physical layer, so all of its innovation had to happen in software and in the interface processors. Ethernet would take the opposite path, inventing its own physical layer and keeping the software simple.

The ARPANET also established that a packet network could be operated, measured, and debugged, which was not obvious in advance. Kleinrock's group at UCLA served as the network measurement center, and the traffic statistics it gathered confirmed that interactive computing traffic was as bursty as the theory had assumed. Confidence in statistical sharing became empirical rather than speculative.

ALOHAnet: Contention as a Design Principle

The branch that leads directly to Ethernet begins in Hawaii. Norman Abramson and colleagues at the University of Hawaii needed to connect terminals on several islands to a central computer in Honolulu, and leased lines across open water were impractical. ALOHAnet, operational in 1971, instead used ultra high frequency radio channels shared by every terminal. Its governing rule was startlingly simple: a station with a packet to send transmits immediately. If two stations transmit at once the packets are destroyed, and each station, detecting the absence of an acknowledgment, waits a random interval and tries again.

Random access traded efficiency for simplicity. Analysis showed that the pure ALOHA channel saturates at about 18 percent of raw capacity, and that dividing time into slots so that transmissions can begin only at slot boundaries, a refinement introduced by Lawrence Roberts in 1972, doubles that figure to about 37 percent. Those numbers look poor beside a scheduled system, but the scheme required no central controller, no reservation protocol, and no coordination among stations that could not hear one another reliably. For a network whose alternative was no network, that was a decisive advantage.

ALOHAnet contributed the trait that distinguishes this branch from the ARPANET line: the shared broadcast medium with contention-based access. Everything after it is an attempt to keep that simplicity while recovering the wasted capacity.

Ethernet at Xerox PARC

Robert Metcalfe encountered the ALOHA analysis while completing a Harvard doctorate on packet networks and concluded that contention could be made far more efficient if the stations shared a wire instead of the open air. On a wire, a station can listen before transmitting and can also listen while transmitting, which permits two improvements that radio did not allow. Carrier sense stops most collisions before they start, and collision detection aborts the remainder within microseconds instead of wasting an entire packet time.

The 1973 Memorandum

Metcalfe circulated a memorandum describing the scheme at the Xerox Palo Alto Research Center on May 22, 1973, and built the first system with David Boggs; it ran in November of that year. He named it after the luminiferous ether, the hypothetical medium that nineteenth-century physics had supposed filled space, on the reasoning that the coaxial cable was a passive medium through which signals simply propagated. The experimental network ran at 2.94 megabits per second, exactly half the 5.88 megahertz system clock of the Xerox Alto workstation it was built to connect, and it identified stations with addresses only eight bits wide. Metcalfe and Boggs published the design in Communications of the ACM in July 1976 under the title "Ethernet: Distributed Packet Switching for Local Computer Networks," and Xerox, which had filed in March 1975, received United States Patent 4,063,220 on the collision-detecting multipoint system on December 13, 1977, naming Metcalfe, Boggs, Charles Thacker, and Butler Lampson as inventors.

What Ethernet Inherited and What It Changed

Ethernet inherited packet switching from Baran, Davies, and the ARPANET, and inherited contention from ALOHAnet. It changed three things. It replaced the random-access free-for-all with carrier sense multiple access with collision detection, which listens first and aborts on conflict. It added binary exponential backoff, doubling the mean retransmission delay after each successive collision so that the network degrades gracefully rather than collapsing under load. And it made the medium entirely passive, so that a failed station could not take the network down with it, an important property when the alternative topologies of the period relied on every station to relay traffic.

The context mattered as much as the mechanism. Xerox PARC had built the Alto, a personal workstation, and a laser printer, and needed something to connect many of the former to a few of the latter. Ethernet was designed for a world of numerous cheap machines rather than one expensive machine with terminals, and that assumption is why it scaled with the personal computer, a trajectory examined in batch to time-sharing to personal to cloud.

From Prototype to Standard: DIX and IEEE 802

A laboratory network becomes a technology only when more than one company can build it. Metcalfe left Xerox and founded 3Com in 1979, and persuaded Xerox to open the specification jointly with Digital Equipment Corporation and Intel. The resulting DIX consortium published Ethernet Version 1.0 on September 30, 1980, raising the rate to 10 megabits per second and widening station addresses from the experimental network's eight bits to 48, and issued Version 2.0 in November 1982. The 48-bit address space, administered so that each manufacturer receives a block of prefixes, is the single most durable element of the design and remains in use across Ethernet, Wi-Fi, and Bluetooth.

The Institute of Electrical and Electronics Engineers began Project 802 in February 1980 to standardize local area networks. The committee could not agree on a single access method, so it split into working groups, and the split preserved the industry's disagreements rather than resolving them: 802.3 standardized carrier sense multiple access with collision detection, 802.4 standardized token passing on a bus, and 802.5 standardized token passing on a ring. The committee approved the 802.3 specification in December 1982, published it as a draft in 1983, and issued it as a full standard in 1985. That standard defined 10BASE5 over thick coaxial cable, with segments up to 500 meters, transceivers attached by clamp-on taps, and a maximum of 100 stations per segment.

Standardization also introduced the lineage's most persistent scar. DIX used a two-byte field after the addresses to identify the upper-layer protocol, while 802.3 used the same field to carry the frame length. The industry ran both interpretations for years, and the reconciliation, finally written into the standard in 1997, simply declared that values of 1536 and above are protocol identifiers and values of 1500 and below are lengths, leaving the narrow band between the two ranges undefined. The compromise is invisible in operation and unintelligible without the history.

The Rival Branches

Ethernet did not win because it was technically superior on the metrics its competitors emphasized. It won on cost, tolerance of imperfection, and the pace at which its supporting silicon improved.

Token Passing

ARCNET, introduced by Datapoint in 1977, passed a token among stations on coaxial cable at 2.5 megabits per second and was widely deployed in the early 1980s. IBM's Token Ring, shipped in October 1985 at 4 megabits per second and standardized as IEEE 802.5 that same year, later reached 16 megabits per second and offered deterministic access: a station transmits only when it holds the token, so worst-case delay is bounded and throughput does not fall as load rises. Token Bus, standardized as 802.4, was adopted for General Motors' Manufacturing Automation Protocol for the same reason. Determinism is genuinely valuable on a factory floor, and the argument that Ethernet could not guarantee delivery times was correct as far as it went.

The token branches nonetheless declined. Their adapters and wiring concentrators cost several times the Ethernet equivalent, their protocols required a monitor station and token recovery machinery that added complexity, and their speed increments arrived slowly. When switched Ethernet later eliminated contention altogether, the determinism argument lost most of its force.

Fiber and Cells

The Fiber Distributed Data Interface, standardized by the American National Standards Institute committee X3T9.5 in the late 1980s, ran a token on dual counter-rotating fiber rings at 100 megabits per second and became the standard campus backbone of the early 1990s. It was fast, resilient, and expensive, and Fast Ethernet undercut it within a few years. Asynchronous Transfer Mode, promoted from 1991 by the ATM Forum, segmented all traffic into 53-byte cells so that voice, video, and data could share one network with guaranteed quality of service. It succeeded in carrier backbones and failed in the office, where LAN Emulation added complexity, adapters stayed costly, and the cell tax of five header bytes in every fifty-three looked increasingly indefensible as Ethernet rates climbed.

A shorter-lived rival deserves mention because it shows how close the contest was. Hewlett-Packard's 100VG-AnyLAN, standardized as IEEE 802.12 in 1995, offered 100 megabits per second with a demand-priority access method that was deterministic and could carry Token Ring frames. It arrived at the same moment as Fast Ethernet, which was compatible with the installed base of ten-megabit adapters, and it disappeared within a few years. Compatibility beat elegance, as it usually does in this family. The wider pattern of technologies that worked and lost is treated under failed technologies and obsolescence.

From Coaxial Bus to Twisted-Pair Star

The first physical change to Ethernet altered its topology while preserving its logic. Thick coaxial cable was awkward to install and unforgiving to fault: a single break or bad tap disabled an entire segment, and locating the fault meant walking the cable. The thinner 10BASE2 variant of 1985, limited to 185-meter segments and 30 stations, was cheaper and easier to route but shared the failure mode, since a workstation disconnected carelessly from its T-connector opened the bus.

AT&T's StarLAN, standardized as 1BASE5 in 1987, showed that ordinary telephone-grade twisted pair could carry Ethernet, though only at one megabit per second. SynOptics shipped LattisNet at the full ten megabits over twisted pair the same year, ahead of any standard. The IEEE ratified 10BASE-T in 1990, and it changed the character of the technology. Each station now had a dedicated pair of wires to a central hub, so a single failure isolated a single station; the hub could report link status, making faults visible from one location; and the wiring followed the star topology that telephone installers already understood. The structured cabling standard published as TIA/EIA-568 in 1991 formalized the practice, and the wall outlet became a fixture of office construction.

Logically, the hub still repeated every bit to every port, so the shared medium and its collisions persisted. Ethernet had become a star that behaved like a bus. That contradiction set up the next transition.

Bridges to Switches: The End of Contention

Bridges appeared in the 1980s to join separate Ethernet segments while filtering traffic that had no reason to cross, and the spanning tree algorithm that Radia Perlman devised at Digital Equipment Corporation, standardized in 1990, allowed bridged networks to contain redundant links without looping frames endlessly. A bridge with two ports reduced contention; the insight that followed was that a bridge with many ports, one per station, would eliminate it.

Kalpana shipped the first multiport Ethernet switch in 1990, and Cisco Systems acquired the company in 1994. Each port became its own collision domain, and forwarding moved into application-specific integrated circuits fast enough to handle every port at once. Kalpana also proposed the further step of running each link in full duplex, with separate transmit and receive pairs used simultaneously, which the IEEE standardized in 1997 along with the PAUSE frame for flow control. At that moment carrier sense and collision detection became vestigial. A modern full-duplex link has no collisions to detect, and the 10 gigabit standard of 2002 dispensed with the half-duplex mode entirely.

Several inherited traits nonetheless survive from the contention era. The minimum frame size of 64 bytes exists because a transmitting station had to still be transmitting when a collision from the far end of the largest permitted network arrived back at it; a station that had already finished would never learn that its frame was destroyed. Sixty-four bytes is 512 bit times, which at 10 megabits per second is 51.2 microseconds, and it was that slot time which set the permissible diameter of a collision domain rather than the diameter which set the frame size. Gigabit Ethernet had to pad the slot time out to 512 bytes with carrier extension to keep half-duplex operation viable at all. The 64-byte minimum persists in every frame sent today. The preamble that precedes each frame exists to let a receiver's clock recovery circuit lock onto a transmission that begins without warning. Vestigial structures of this kind are the clearest evidence of ancestry in any genealogy, technical or biological. The internal architecture of the devices that replaced the shared wire is examined under switch fabric architecture.

Speed as Inherited Borrowing

Each increase in Ethernet's data rate was achieved less by invention than by adopting a physical layer already proven elsewhere, then wrapping it in the unchanged frame format. The pattern is consistent enough to be a rule of this lineage.

Fast Ethernet and Gigabit Ethernet

Fast Ethernet, standardized as 802.3u in 1995, reached 100 megabits per second over two pairs of Category 5 cable by borrowing the 4B/5B coding and the twisted-pair physical medium dependent sublayer developed for FDDI. The same standard also defined 100BASE-T4, which spread the same rate over four pairs of the older Category 3 cable, but the two-pair variant is the one that survived. Fast Ethernet also introduced auto-negotiation, which extended the link pulses of 10BASE-T into bursts that advertise capabilities, so a new adapter could fall back and interoperate with the installed base. Gigabit Ethernet over fiber, standardized as 802.3z in 1998, borrowed the 8B/10B coding and serializer technology of Fibre Channel and ran it at 1.25 gigabaud. Gigabit over copper, standardized as 802.3ab in 1999, was the genuine innovation of the sequence: it used all four pairs of Category 5 cable in both directions at once, requiring hybrids, echo cancellers, and adaptive equalizers implemented in digital signal processing, with five-level pulse amplitude modulation to keep the symbol rate at 125 megabaud.

Ten Gigabits and Beyond

The 10 gigabit standard of 2002 defined fiber interfaces only and abandoned half duplex. 10GBASE-T followed in 2006, pushing Category 6a cable to 500 megahertz with sixteen-level modulation and low-density parity-check coding, at a power and latency cost that kept it out of some applications for years. Aggregated lanes carried the next steps: 40 and 100 gigabits per second in 2010, 200 and 400 gigabits per second in 2017 using four-level pulse amplitude modulation on each lane, and 800 gigabits per second in 2024, with work toward 1.6 terabits per second continuing in the IEEE 802.3dj task force. Every one of these frames is still recognizable as the frame DIX defined in 1980, carrying the same 48-bit addresses and, by default, the same 1500-byte payload limit.

The persistence of that payload limit illustrates how genealogy constrains engineering. Jumbo frames of roughly 9000 bytes reduce per-frame overhead substantially at high rates, and most switches and adapters support them, but they were never standardized, because a frame larger than the standard would fail on any device that had not been configured for it. An inherited parameter chosen when links ran at ten megabits per second still governs interoperability at eight hundred gigabits.

Ethernet Beyond the Office

Having displaced its local rivals, Ethernet expanded into domains that had their own established networks, generally by the same method: undercut the incumbent on cost, then add the missing capability.

Carrier and Metropolitan Networks

Telephone carriers had built their data services on SONET and SDH transport and on ATM, and Ethernet initially lacked the operations, administration, and maintenance facilities that carriers required. Ethernet in the First Mile added link-layer management in 2004, provider backbone bridging, standardized in 2008, supplied the address scaling that carrier networks needed, and the Metro Ethernet Forum, founded in 2001, defined service types that customers could purchase by bandwidth. An Ethernet handoff is now the ordinary way to deliver business connectivity. ATM has largely left carrier networks, surviving where it survives as legacy circuits emulated by pseudowires over packet transport, while the label-swapping idea it contributed lives on in multiprotocol label switching, which carries Ethernet frames rather than cells. Contemporary carrier equipment is surveyed under network infrastructure.

Data Centers

Large data centers abandoned the hierarchical tree in favor of leaf-and-spine fabrics built from many identical switches, a topology derived from the Clos networks that Charles Clos described for telephone switching in 1953. Much storage traffic, which had run on Fibre Channel, and a growing share of high-performance and machine-learning cluster traffic, which had run on InfiniBand, moved onto the same Ethernet fabric once the data center bridging extensions added priority-based flow control and bandwidth allocation, permitting remote direct memory access over converged Ethernet. Neither incumbent has vanished, and InfiniBand in particular remains widely deployed in large training clusters, but the direction of migration has been consistent. The economic argument was familiar: one high-volume technology, improved continuously, beats several specialized ones.

Factories and Vehicles

Industrial networks had good reasons to distrust Ethernet, since a machine tool cannot wait an indeterminate time for a command. Determinism returned through the IEEE 802.1 audio-video bridging work of the late 2000s, renamed time-sensitive networking in 2012, which added precise time synchronization and time-aware traffic shapers that reserve scheduled windows for critical frames. Power over Ethernet, standardized in 2003 and extended in 2009 and 2018, now delivers up to 90 watts from the source equipment, of which about 71 watts reaches the device after cable loss, and it removed the need for separate power wiring to cameras, access points, and sensors. Automotive Ethernet followed, with single-pair variants standardized from 2015 onward.

One automotive variant closes the circle of this genealogy. The 10BASE-T1S standard of 2019 places multiple nodes on a single shared pair, reviving the multidrop bus that twisted-pair wiring had abolished thirty years earlier, because a vehicle harness cannot afford a dedicated run to every sensor. The shared segment still specifies carrier sense multiple access with collision detection, but an optional sublayer called physical layer collision avoidance grants each node a scheduled transmit opportunity in turn, so that collisions are prevented rather than resolved. The shared medium returned when the economics of wiring made it attractive again, with a modern access method in place of the old one. These developments are treated further under Ethernet and industrial protocols and industrial network infrastructure.

The Wireless Sibling

Wi-Fi is not a descendant of Ethernet but a sibling, descended directly from ALOHAnet. Because a radio transceiver cannot listen while transmitting on the same channel, collision detection is unavailable, and 802.11 therefore uses collision avoidance: stations wait for a clear channel, add a random backoff, and acknowledge every frame. It nonetheless adopted Ethernet's 48-bit addresses and presents an Ethernet-like service to the layers above, so that software written for one runs unchanged on the other. The two branches of the ALOHA family converge at the bridge in every access point. That expansion is traced under wireless networking expansion.

Why This Branch Survived

Four factors explain the outcome, and each generalizes to other lineages in electronics.

The first is that Ethernet's simplicity placed its cost on a steeper improvement curve. Its access method required little state and few timers, so it could be reduced to silicon early and cheaply, and each generation of semiconductor process cut the price of an interface again. Competitors whose protocols demanded more logic could not follow the same curve.

The second is that the frame stayed stable while everything under it changed. Because addressing and framing were held constant, a physical layer borrowed from FDDI, Fibre Channel, or optical transport could be adopted without disturbing any software, and auto-negotiation let new equipment coexist with old on the same wiring. Stability at one layer bought freedom at another.

The third is that the technology was multi-vendor from the beginning. The DIX publication and the subsequent IEEE process meant that no single company controlled Ethernet, which enlarged the market, attracted component suppliers, and made buyers confident of second sources. Token Ring's association with one dominant vendor had the opposite effect.

The fourth is that Ethernet absorbed its critics' requirements instead of disputing them. Determinism, quality of service, guaranteed bandwidth, and long reach were all genuine gaps, and all were eventually filled by extensions rather than by a replacement architecture. A technology that can absorb the strengths of its rivals does not need to be better than them at the outset.

Summary

Packet switching began as an argument about how to allocate capacity, invented independently by Baran and Davies in the first half of the 1960s, furnished with its analytical theory by Kleinrock, and proved at scale by the ARPANET at the end of that decade. ALOHAnet adapted the idea to a shared radio channel and contributed contention-based access. Ethernet moved that contention onto a wire, where carrier sense and collision detection made it efficient, and Xerox, Digital Equipment, and Intel turned the result into a published specification in 1980 that the IEEE standardized as 802.3.

The technology then changed almost completely while appearing not to change at all. The bus became a star, the hub became a switch, the collisions disappeared, half duplex was retired, and the physical layer was replaced repeatedly with technology borrowed from elsewhere. What persisted was the frame: 48-bit addresses, a minimum length set by a constraint that no longer exists, and a payload limit chosen for ten-megabit coaxial cable. Rival branches including ARCNET, Token Ring, FDDI, desktop ATM, and 100VG-AnyLAN each held a technical advantage and each lost to cost, compatibility, and the pace of silicon improvement.

The lineage remains active. Ethernet now reaches from a single twisted pair in an automobile harness to 800-gigabit links between data center buildings, and the statistical sharing principle that Baran and Davies proposed for a network of a few dozen nodes governs the traffic of the entire Internet.

Related Topics