Dead End: Token Ring

Abstract

IBM’s Token Ring was, by several technical measures, a superior network technology to Ethernet. It used a deterministic access protocol that guaranteed every device a turn to transmit, eliminated collisions that could degrade Ethernet performance under load, and at 16 Mbps outpaced Ethernet’s 10 Mbps. IBM pushed Token Ring as the IEEE 802.5 standard and made it the default networking technology for its own products through the late 1980s and early 1990s. Token Ring lost anyway, to cheaper cable, simpler equipment, a patent holder who taxed every adapter sold, and the arrival of switched Ethernet, which removed the shared medium that Token Ring’s whole argument depended on.

The Token Ring Architecture

Token Ring was developed by IBM in the late 1970s and early 1980s. The controlling patent belonged to someone else: Olof Söderblom, a Swedish engineer, held US Patent 4,293,948, “Data transmission system,” filed on 29 October 1974 and granted on 6 October 1981, which described terminals serially interconnected in a closed loop and passing the right to transmit around it. Every vendor that shipped a token-passing adapter paid him, and that bill ended up in the price of the hardware.

The core idea was a token-passing protocol: a special control frame called a token circulated continuously around a ring of computers. A device that wanted to transmit had to first capture the token, then transmit its data, then release the token for the next device.

This architecture had important properties. It was deterministic: the maximum time any device would wait to transmit could be calculated. Under Ethernet’s CSMA/CD protocol, devices listened for a clear channel and transmitted opportunistically; if two transmitted simultaneously, there was a collision and both had to retry. Under high network load, collisions became frequent and Ethernet throughput degraded. Token Ring never had collisions. Every device knew its position in the ring and knew the token would arrive.

It was also fair: the token circulated through every device in sequence. No device could monopolize the network, and every device had a bounded wait time. This made Token Ring predictable in a way that mattered for time-sensitive applications, factory automation, process control, environments where a missed deadline had real consequences.

Error detection was built into the protocol. If a device failed to release the token after transmitting, an Active Monitor device on the ring detected the lost token and generated a new one. The ring could recover from single node failures without manual intervention.

IBM launched its Token Ring product on 15 October 1985, running at 4 Mbps. In 1988 the 802.5 working group standardized 16 Mbps Token Ring, which added Early Token Release: a device could release the token immediately after transmitting rather than waiting for its own data to travel all the way back around the ring, which raised throughput under load.

The Hub Problem

Token Ring’s physical implementation required a Multistation Access Unit (MAU), IBM’s term for what Ethernet called a hub. The MAU was the central point through which all devices connected. Each device ran a cable to the MAU; the MAU internally connected the cables in a ring sequence.

MAUs and adapters both cost several times their Ethernet equivalents, and the gap widened year by year rather than closing. Ethernet’s own prices had started high: a connection cost $3,000 to $4,000 per node in 1979 and 1980, which DEC and Xerox set out to push to $500 by getting the silicon second-sourced. By 1985 an Ethernet adapter card was about $600, and by 1991 more than a hundred vendors were selling 10BASE-T cards. Nothing comparable happened on the other side of the market.

The cable requirements compounded the cost. IBM’s original Token Ring specification required shielded twisted pair (STP) cable, heavier, more expensive, and harder to install than the unshielded twisted pair (UTP) that Ethernet’s 10BASE-T standard (finalized 1990) used. A Token Ring installation in a large office building required specialized cable installers, specific STP connectors, and careful attention to cable run length. Ethernet UTP installations used the same RJ-45 connectors that telephone systems used, in cable that building contractors could install without specialized networking training.

The IBM PC adapter for Token Ring (sold as the IBM Token Ring Network Adapter) carried IBM’s premium pricing. Organizations buying IBM PS/2 computers and IBM-branded networking were paying IBM prices throughout. This was IBM’s strategy: Token Ring was profitable hardware. Ethernet was a commodity where margins had collapsed.

Urs von Burg and Martin Kenney, comparing the two standards’ supplier communities, name four reasons the premium held. Token Ring’s access method was more complex, so its chips were harder and dearer to build. IBM licensed only Texas Instruments as a second chip source, and TI’s chips performed worse and were priced so high that third parties could not undercut IBM with them. IBM itself preferred a profitable business to a large one and never priced forward to buy share. And Madge, its main rival, competed on performance rather than price, which held the premium in place from the other direction.

The Söderblom Royalty

Underneath the price of every Token Ring adapter sat a cost Ethernet never had to carry. Söderblom’s patent covered the token-passing scheme itself, and he charged the vendors and chip makers who implemented it. By 1990 he had collected some $20 million in royalties, and the fee was one reason Token Ring adapters stayed dearer than Ethernet ones no matter how many were sold. Scale could not work the price down past a per-unit levy.

Robert Madge, whose company was then the largest Token Ring vendor after IBM, refused to pay and went to court. On 24 June 1990 a British court ruled the patent invalid, finding that it did not actually describe token passing or cover local area networks. The victory was symbolic in Britain, where the patent had already expired in 1988. The American patent stood until 1998, which covered almost the whole commercial life of the technology.

Ethernet had no equivalent toll. Digital, Intel, and Xerox had published the specification and licensed it on terms designed to spread it, and the IEEE standard that followed was open to anyone. Two technologies competing on price, one of them paying an inventor on every unit shipped, do not converge.

The IEEE 802.5 Political Battle

During the early 1980s, both Token Ring and Ethernet were competing to be adopted as the IEEE networking standard. The IEEE 802 committee that covered local area networks created working groups for different technologies: 802.3 for Ethernet (CSMA/CD), 802.4 for Token Bus (a token-passing protocol used in factory automation), and 802.5 for Token Ring.

IBM lobbied aggressively for Token Ring adoption. The company had significant influence within IEEE, large corporate customers who trusted IBM’s networking recommendations, and a history of setting de facto standards (SNA, EBCDIC, 3270 terminals) within its customer base. IBM’s position was that Token Ring’s deterministic behavior made it technically superior, and that organizations with serious networking needs should adopt the superior standard.

Ethernet’s advocates, led by Digital Equipment Corporation, Intel, and Xerox (the original Ethernet developers), argued that Ethernet’s lower cost and simpler implementation would accelerate adoption. The IEEE adopted both 802.3 (Ethernet) and 802.5 (Token Ring) as standards. The market then determined which would dominate.

IBM’s mainframe and minicomputer customers, who trusted IBM’s networking recommendations and bought IBM hardware, adopted Token Ring. Organizations without a strong IBM relationship, organizations in price-sensitive environments, and organizations that found Token Ring’s installation complexity onerous adopted Ethernet. The installed base split roughly along IBM customer alignment through the late 1980s.

10BASE-T and the Turning Point

Token Ring’s decline started in 1990, when the IEEE ratified 10BASE-T, Ethernet over unshielded twisted pair cable. Before 10BASE-T, Ethernet ran on coaxial cable (10BASE-5, “thick Ethernet,” and 10BASE-2, “thin Ethernet” or “cheapernet”). Coaxial cable required connecting devices in a bus topology, and a break anywhere in the cable segment could bring down the entire network. It was also harder to install in buildings with existing telephone wiring.

10BASE-T brought Ethernet to the same star topology Token Ring used (each device connected by its own cable to a central hub) while using the cheaper, thinner unshielded twisted pair cable that was already installed in most office buildings for telephone use. An organization that had existing UTP telephone cable could add Ethernet by installing hubs and NICs, often reusing cable runs they already had.

10BASE-T turned the adapter into a commodity. Over a hundred vendors were shipping 10BASE-T cards by 1991, competing on price against each other; hub revenue at SynOptics and Cabletron went from $6.1 million and $9.5 million in 1987 to $248.3 million and $180.5 million in 1991. Token Ring’s cards came from a handful of vendors, and their prices moved accordingly.

The Switched Ethernet Solution

Token Ring’s primary technical argument (deterministic access that prevented performance degradation under high load) was addressed by switched Ethernet technology that emerged in the early 1990s.

Ethernet switches replaced shared hubs. In a shared Ethernet hub, all devices on the segment competed for the same bandwidth. A collision on any device affected all devices on the segment. In an Ethernet switch, each port was an independent collision domain, the switch examined the destination MAC address of incoming frames and forwarded them only to the appropriate port. Two devices connected to a switch could transmit simultaneously without interference.

Switched Ethernet effectively eliminated the shared-medium collision problem that had been Token Ring’s strongest argument. With each device on its own switched port, the network behaved like dedicated point-to-point connections between each device and the switch. Token Ring’s deterministic access advantage over shared Ethernet became irrelevant because switched Ethernet no longer shared access.

100BASE-TX (Fast Ethernet), ratified as IEEE 802.3u in 1995, brought Ethernet to 100 Mbps, ten times Token Ring’s 16 Mbps maximum. Token Ring vendors responded with High Speed Token Ring (HSTR) specifications for 100 Mbps and 1 Gbps Token Ring, but these were never widely deployed.

IBM’s Strategic Retreat

IBM’s response to Fast Ethernet was cautious and late. The company continued manufacturing and selling Token Ring products through the 1990s, maintaining its existing customer base while the broader market shifted. IBM’s networking revenue from Token Ring was significant (the equipment carried higher margins than Ethernet) and IBM was reluctant to acknowledge that the transition was happening.

By the late 1990s Token Ring was a maintenance market rather than a growth one. New installations were predominantly Ethernet, organizations with Token Ring were keeping existing networks running rather than extending them, and new buildings were wired for Ethernet. What remained was fought over by three vendors: IDC put the 1999 Token Ring split at IBM 49 percent, Madge 24 percent, and Olicom 22 percent.

That market then folded in on itself. In September 1999 Olicom, in financial trouble, sold its entire Token Ring division to Madge Networks for $15 million in cash plus payments tied to three years of Madge’s Token Ring revenue. IBM wound down its own Token Ring production in 2004 and handed the entire business, products and service, to Madge, leaving one vendor in a category with nothing left to fight over. Madge Networks N.V. had already filed for bankruptcy protection in 2003, and the Token Ring operation ended up at Ringdale after a management buyout.

The final standard arrived after the argument was over: Gigabit Token Ring (IEEE 802.5z) was published on 4 May 2001 and never saw meaningful deployment.

The existing Token Ring installed base continued operating for years after IBM’s exit. Organizations with large Token Ring deployments (particularly those in IBM-heavy environments, financial services, and factory automation) had no immediate reason to replace working infrastructure. Migration to Ethernet required replacing NICs in every computer, replacing MAUs with Ethernet switches, and potentially recabling if STP installations were being replaced with UTP. These migrations happened gradually through the 2000s as equipment was replaced during normal refresh cycles.

Token Ring’s Legacy Markets

Factory automation and process control (the environments where Token Ring’s deterministic behavior was most valuable) largely adopted PROFIBUS and later Industrial Ethernet variants rather than standard IEEE 802.5 Token Ring. The industrial networking market developed its own specifications (PROFIBUS, DeviceNet, ControlNet, later EtherNet/IP) that provided determinism guarantees on top of standard physical layers. IBM’s Token Ring, designed for office networking, was not optimized for the industrial environments where its technical properties would have mattered most.

Dead End

Token Ring’s failure repeats through technology history: when two technologies compete and one is cheaper, the cheaper one tends to win even when the more expensive one is measurably better.

Token Ring’s deterministic access was genuinely valuable in specific scenarios, high-utilization networks, time-sensitive applications, environments where maximum latency mattered more than average latency. But most office networks were not high-utilization, and most office applications were not time-sensitive. Email, file transfers, and shared printers worked adequately on shared Ethernet at 10 Mbps; the collision-driven performance degradation at high load was a theoretical problem that few real installations reached.

The economics of Ethernet were driven by scale. The PC NIC market was enormous, every PC sold needed one, and competition among manufacturers drove prices down fast. Token Ring’s smaller installed base could not reach the same economies, and the Söderblom royalty put a floor under its component costs that volume could not push through. The larger technology and the smaller one pay very different prices for the same silicon, whatever the merits.

Ethernet also benefited from timing: when the transition from coaxial to twisted-pair happened in 1990, Ethernet got there first and at lower cost. If Token Ring had offered a credible UTP solution at competitive prices in 1990, the market might have gone differently. Instead, IBM’s proprietary pricing and STP cable requirements gave Ethernet an opening, and Ethernet’s scale advantages closed the door.