Ethernet: From Xerox PARC's Experimental Network to IEEE 802.3

Ethernet is often described as Robert Metcalfe’s 1973 invention and then projected unchanged into modern networks. The actual history is more collaborative and more technically interesting. A Xerox PARC team turned ideas from packet radio into a local network over coaxial cable; Xerox, Digital Equipment Corporation, and Intel then specified a commercial, multi-vendor version; and IEEE developed the related 802.3 standard. Later bridges, twisted-pair links, and full-duplex switches removed the shared collision environment that had defined early Ethernet while retaining a useful service and frame lineage.

The result is not one wire format frozen for fifty years. It is a family that survived because engineers could replace its media, speeds, and access assumptions without forcing applications and higher network layers to start over.

A memo was a milestone, not a solitary act of invention

Metcalfe’s May 1973 memo at Xerox’s Palo Alto Research Center is the conventional birthday of Ethernet. It proposed a local communications network for the Alto personal computers and laser printers being developed there. The name invoked the nineteenth-century “ether,” an imagined medium able to carry electromagnetic waves everywhere, because the new network would act as a shared, passive medium available to attached stations.

But a dated memo should not erase the team. David Boggs co-designed and built crucial parts of the experimental system, and other PARC researchers contributed hardware, software, testing, and requirements. Metcalfe and Boggs jointly authored the definitive 1976 paper, Ethernet: Distributed Packet Switching for Local Computer Networks. Its reported experimental network connected more than one hundred stations over about a kilometer of coaxial cable and operated at 2.94 megabits per second. The original PARC system was therefore not the later ten-megabit commercial Ethernet.

The design also had an acknowledged intellectual predecessor. Metcalfe had studied the University of Hawaii’s ALOHAnet, where geographically separated radio stations transmitted over a shared channel and retried after collisions. Ethernet adapted random access to a cable on which a station could first listen for activity and could detect interference while transmitting. Innovation lay in that adaptation, its implementation, and its integration into a working local-computing environment—not in inventing packets or shared channels from nothing.

What CSMA/CD actually did

Classic shared Ethernet used carrier-sense multiple access with collision detection, abbreviated CSMA/CD. “Carrier sense” meant a station listened before transmitting. “Multiple access” meant many stations used the same medium. Listening reduced collisions but could not prevent them: signals take time to propagate, so two distant stations could both find the cable idle and begin almost simultaneously.

When transmitters detected a collision, they stopped, sent a short signal so every participant recognized the event, and waited before trying again. The wait followed a binary exponential backoff procedure whose range grew after repeated collisions. This avoided having the same stations immediately collide in lockstep, while leaving the network decentralized. A collision in this design was an anticipated contention event, not evidence that a cable or adapter had necessarily failed.

Timing imposed physical constraints. A transmitter had to remain active long enough to detect a collision originating at the farthest legal point in the network. Minimum frame size, propagation delay, repeater limits, and maximum network extent were therefore related. Marketing summaries that reduce CSMA/CD to “everyone talks, then retries” omit the engineering that made collision detection reliable.

Early coaxial installations formed a bus: damage, bad termination, or a poorly connected tap could disrupt a large segment. Later hubs made cabling physically star-shaped, but a hub repeated incoming bits to its other ports. All attached devices still shared bandwidth and a collision domain. Changing the box in the wiring closet did not yet change the underlying contention model.

From Xerox protocol to multi-vendor specification

Xerox could have kept Ethernet tied to its own systems. Instead, Xerox joined DEC and Intel in producing the DIX Ethernet specification, named after the three companies. Version 1 appeared in 1980, with a ten-megabit signaling rate; a revised Ethernet Version 2 followed in 1982. A public specification encouraged independent network interfaces and gave buyers more confidence that equipment from different vendors could interoperate.

IEEE’s 802 project developed a standards framework for local networks, and IEEE 802.3 was approved in 1983. It was closely related to DIX Ethernet but was not originally byte-for-byte identical in every field interpretation. Ethernet II used a two-byte field after the source address as an EtherType identifying the higher-layer protocol. IEEE 802.3 initially interpreted that position as a payload length and relied on additional logical-link-control fields. Values and encapsulation conventions eventually allowed receivers to distinguish common forms, and Ethernet II framing became dominant for protocols such as IP.

That nuance explains why using “Ethernet” and “IEEE 802.3” interchangeably is usually practical today but historically imprecise. Standardization was a process of convergence, amendments, implementation practice, and market selection, not a ceremonial renaming of an unchanged PARC protocol.

Twisted pair made installation easier

Coaxial buses were economical but awkward to expand and diagnose. Ethernet over unshielded twisted-pair telephone-style cabling shifted each station to a point-to-point run terminating in a central hub or switch. IEEE standardized 10BASE-T in 1990. The physical star made it easier to isolate a failed cable to one station and fit networking into structured building cabling.

The notation summarized important properties: “10” indicated ten megabits per second, “BASE” meant baseband signaling, and “T” identified twisted pair. Later physical layers changed encoding, cable requirements, lane counts, and reach. Fiber variants served longer distances and electrically noisy environments. This evolution was possible because applications did not need to know whether a frame crossed thick coax, copper pairs, or optical fiber.

Switching ended normal collisions

A bridge learned which source addresses appeared on each port and forwarded a frame only where needed. A multiport bridge became commercially known as an Ethernet switch. By dividing one shared segment into separate links, switching reduced contention and allowed multiple conversations to proceed concurrently.

The decisive change was full duplex on point-to-point links. One endpoint could transmit while the other transmitted on a separate path, so no other station competed for the medium. There were no normal collisions to detect and CSMA/CD was disabled. Flow control, buffering, congestion, packet loss, and physical errors remained real problems, but they were different problems from half-duplex collision management.

This is why the claim that Ethernet “still uses collisions” is false for ordinary modern switched networks. CSMA/CD remains in the standards history and in support for legacy half-duplex operation, not in the normal traffic path of a full-duplex switch port. It is equally misleading to call Wi-Fi wireless Ethernet: IEEE 802.11 shares some link-layer concepts but uses a different medium-access system, including collision avoidance because a radio generally cannot detect a collision the same way while transmitting.

Faster Ethernet preserved a boundary, not every detail

Ethernet advanced from ten megabits to one hundred megabits, one gigabit, ten gigabits, and much higher rates used in data centers and carrier networks. Achieving those rates required radically different electronics and signaling. Some versions transmit over multiple copper pairs; some use parallel or serial optical lanes; link training and forward-error correction appear in faster families. An original coax transceiver could not participate in a contemporary high-speed link.

Continuity instead exists at an architectural boundary: recognizable addresses and frames, a broadly stable service to higher layers, and bridges that can forward between different physical implementations. Even there, the ecosystem accumulated virtual LAN tags, link aggregation, priority fields, larger-frame conventions, and many specialized amendments. “Backward compatible” should not be misread as “any device can be plugged into any later cable.”

Ethernet’s longevity therefore came from controlled change. The PARC team demonstrated decentralized packet delivery on a shared local medium. DIX made a ten-megabit multi-vendor specification. IEEE supplied an extensible standards process. Twisted pair improved deployment, and switching plus full duplex retired collisions from normal operation. What endured was not the original coax or a permanent contention algorithm, but a sufficiently stable interface around which the physical network could be repeatedly rebuilt. Related: No, Bill Gates Never Said ‘640K Ought to Be Enough for Anyone’ · No, Al Gore Never Said He ‘Invented the Internet’ — Here’s the Actual Quote