Computers of the first half of the 20th century generally used relays or vacuum tubes as their logic elements. Each of these components has there own methods of failure, but relays and tubes have a much shorter life than the transistors and integrated circuits that later replaced them.

Was there a study comparing the reliability or lifespan of relays and vacuum tubes used in computers? Which kind had a longer mean-time-between-failure or was considered more reliable?

The answer to Why were relays prevalent in early 1940s computers when vacuum tubes were also available? appears to have nothing to do with reliability, the subject of this question.

  • 1
  • The wikipedia article on the Whirlwind I computer outlines the continuing reliability problems with the 5,000 vacuum tubes it contained. The Whirlwind I went operational in 1951, shortly after the heyday of relay based computing. May 17, 2021 at 10:45
  • I think there are two very different questions of reliability here. Human-scale and compute-scale. A vacuum tube that lasts only one hour can still do more logical calculations at ~1 million Hz during its working lifespan, than a relay that lasts a year switching at 1 Hz. Is a computer that can run for an hour doing a billion calculations more reliable than one which can run for a year doing ten million? Depends how you look at it and what you're trying to compute, I suppose.
    – RETRAC
    May 24, 2021 at 18:08

4 Answers 4


There is a case study in the form of the Harwell WITCH, which is a hybrid valve-relay based computer from roughly 1950. One of the design goals was explicitly to minimise the number of "hot cathode" valves in its design, as those were the chief source of reliability problems in the existing late-1940s experimental computers, and Harwell wanted a reliable computer to do real work with.

The memory elements of the WITCH were made of cold-cathode Dekatron valves, each of which could store a decimal digit with decent power efficiency for the time, and as a bonus provided a human-readable indication of their contents. Most of the control logic, including the memory selectors and instruction register, consisted of Post Office type relays which were readily available and had proven reliability. This meant that only a small amount of high-speed logic needed to be implemented using conventional hot-cathode valves. In this case "high speed" was relative, with the main oscillator being derived from the mains AC supply at 50 Hz, and the detailed operation of the machine can be seen with the naked eye.

This yielded a computer that was relatively inexpensive to build (by the standards of the time), was fast enough to replace a small room of human computing clerks equipped with mechanical calculators, made substantially fewer mistakes than said clerks, and did not need to be shut down at the end of the working day. Indeed on one occasion it was given a series of problems to work on over a Christmas and New Year multi-week holiday, and was still faithfully working on them when the operator came back to work afterwards. This reliability made the computer relatively cheap to operate, compared to one that needed to maintain a stock of spare valves and a staff of trained diagnosticians at all times.

Eventually the science of building computers advanced to the point where obtaining one of useful reliability and considerably greater performance than the WITCH was feasible, and the computer was handed off to a university as an educational tool, which is where it acquired the WITCH name. This led to its survival and restoration in the Computer History Museum today. No other computer of its time has survived substantially intact, and in working order with so many of its original parts.

An example of these newer, faster, more reliable computers of around that time would be the LPC-21. Here the main memory element was a spinning magnetic drum, the logic was implemented using a hundred or so hot-cathode valves, and the microcode ROM was implemented using solid-state diodes - probably germanium. Germanium diodes had their own reliability problems, but they were vastly superior to thermionic diodes in all respects. So again, we can see that obtaining a reliable computer may have hinged on keeping the use of hot-cathode valves down to a manageable number.

  • mains AC supply at 50 Hz is wrong for the USA, it is 60 Hz
    – Uwe
    May 17, 2021 at 9:01
  • 1
    @Uwe The Harwell machine was only ever run in the UK, where the mains frequency is 50Hz.
    – Chromatix
    May 17, 2021 at 9:52

I'm going to suggest that this is very much asking to compare "apples to oranges". Vacuum tubes have MTBF values based on their design and operating conditions but they don't experience any "wear and tear" from normal operation such as switching.

Relays, on the other hand, are electromechanical devices which experience mechanical wear and will have a useful lifetime dominated by the number of cycles they can endure. So usage patterns make all the difference. A relay that switches only a few times per hour will last much longer than one that switches a few times per second.

But both vacuum tube based computers as well as relay-based computers were very problematic and experienced significant downtime due to the need to identify and replaced failed elements. Solid-state computers, even the early ones built with discrete transistors were at least an order of magnitude more reliable due to the much longer MTBF value of transistors. Then the introduction of integrated circuits raised the reliability by at least another order of magnitude again due to the much longer MTBF of those devices.

Today's computers very often last for their entire useful life without experiencing a hardware failure despite having several orders of magnitude more active elements in them when compared to early computers.

To my knowledge nobody bothered to study the reliability of relay vs. vacuum tube systems. The difference in speed, vacuum tubes being much faster than relays, meant that nobody in their right mind would continue to use a relay-based computer when a vacuum tube design is so much faster.

Might someone have actually studied this? Possibly, but you'll need to start digging through some very old publications.

  • several orders of magnitude more active elements in them when compared to early computers I suggest (but no proof, haven't checked individual IC MTBFs) that while there was a huge difference vacuum tube/relay -> transistor, that the MTBF of today's machines is largely (not entirely) superior not so much due to ICs being that much more reliable than discrete transistors, but rather that each part is comparable. A modern desktop computer has in the 100s of components, May 16, 2021 at 14:02
  • so the MTBF based on individual component (not transistor or gate) failure will be far superior to any vacuum tube/relay/transistor computer since you needed 1,000s of components to do anything substantive, and often far more than that. May 16, 2021 at 14:02
  • 2
    ENIAC (17,000 tubes) had a clock speed of 100KHz and on average had one tube failure every 2 days. That is about 17 billion clock cycles per machine failure. Modern mechanical relays are only designed with a MTBF of about 1 million switching cycles. On the other hand the MTBF of a relay that is not switching is effectively infinite. Relays in railway switching equipment, operating in uncontrolled harsh environments (not just outdoor temperatures, but also smoke and steam, before the modern era) are expected to have a MTBF of 50 years, though the switching rate is obviously very low.
    – alephzero
    May 16, 2021 at 14:10
  • ... whether you count the "clock speed" of ENIAC as 100Khz (the actual clock) or 5 kHz machine code instructions / second maximum, the number of "switching cycles to machine failure" is still at least 1000 times bigger for tubes than for relays, for the same number of tubes or relays in the complete system. "Lies, damned lies, and statistics" applies to any attempt to compare apples and oranges, of course.
    – alephzero
    May 16, 2021 at 14:18
  • @alephzero: Do relay MTBF stats distinguish between situations where a relay fails to open or close, but a subsequent attempt to activate the relay would clear the fault, versus those where a relay gets permanently stuck? Computers that use a two-of-seven decimal representation could be designed to allow recovery from some failures of the former kind if they include circuitry to ensure that calculation results always have a valid combination of outputs set.
    – supercat
    May 17, 2021 at 17:32

This is a very interesting question. Apparently, at the time relays and tubes were used in computational equipment, there were no studies specifically regarding the lifetime of these components in such equipment. The state of the art was advancing at such a pace that relays and tubes were used for only a brief period of time and then left behind after about 1955 to 1960. However, relays which had been used a brief period longer than tubes in computation, had proven useful and durable in telephonic switching equipment. That use would continue long after computational use ended. And, as should be noted, certain specialty tubes were used in computation but apparently only on one platform developed by the Institute for Advanced Study. In answer to the question, was there a study comparing the reliability or lifespan of relays or vacuum tubes used in computers, the answer seems there was none. Nevertheless, there was extensive experience in the use of such devices and their performance capabilities were compared and documented. The following summary of the then state-of-art is offered...

Reliability in computing devices is of the utmost importance, and therefore a serious problem arises with the use of electromechanical relays. Interference with one or more computing operations may result if a particle of dust happens to rest on the contact faces, since electrical contact is made between two points of microscopic size. Since there is no direct way to detect such a failure either before or after it occurs, the use of checking circuits is desirable. Fortunately, it is relatively easy to add contacts to an electromechanical relay for checking purposes. This was done on the Bell Telephone Laboratories machines, and their ability to resume operation quickly after a failure is one of their outstanding features.

A comparison between the electromagnetic relay and the electron tube reveals that the former has a longer life expectancy in operating hours but will operate fewer times than the latter. Furthermore, the electromechanical relay is subject to failure from vibrations and dust on the contact faces and is apt to require more delicate care than a vacuum tube. The relay usually operates on less power than the filaments of a vacuum tube, but the latter can operate on alternating current taken directly from the power mains through step-down transformers. Most fast-acting relays operate on direct current; hence when the rectification conversion loss is taken into account, their power requirements are comparable to those of electron tubes. The two are nearly the same in cost and size. The greatest difference in characteristics is in the maximum speed of operation; the electron-tube relay is at least 10E4 times faster than the electromechanical relay. Therefore, in situations where computing may be done serially, a single tube can perform the same operations in the same time as a large number of relays operating simultaneously in parallel.

Source: Engineering Research Associates (ERA), Inc., 1950. High-Speed Computing Devices. McGraw-Hill Book Company, New York. pp 313-314.

Reading through the ERA text, one thing becomes abundantly clear. For the state of the art in computing existing in 1949, components used in the construction of computing equipment were not designed and life-tested for such use. In essence, parts such as relays and tubes were off the shelf, already available. Manufacturers, sensing the need for durability in relays, began to manufacture equipment that had higher durability than previously. Nevertheless, tubes such as 6SA7 and 6L7, for instance, were just radio-equipment tubes and were manufactured as they had always been. Although similar tube designations may have unique design characteristics (folded vs twisted filaments, for example), their use in computing equipment was always subject to eventual failure. As noted by ERA, for the ENIAC, tube failures usually occurred within the first 250hrs of operation, or after 5000 hours. The ENIAC operating staff was reportedly not in sympathy with replacing a large number of tubes at once, presumably to preclude failure during equipment use.

The most common source of trouble in using the ENIAC related to trouble with the 18,000 vacuum tubes used in the machine. Although the practice was to test several hundred of the 3-tube amplifiers each week, and replace all three tubes if any amplifier was found to not be operating satisfactorily, the quantity of replaced tubes was considerable. ERA reports that during the first 11 months of operation in 1949, about 2000 tubes per month were replaced each month, of which about half were actually bad.

At this particular time, say 1948 to 1950, there were two approaches to reducing equipment failure during operation. One approach was to make components so reliable that they would outlast the useful life of the computing equipment in which they were being used. At this time there were not statistical tests of lifetime durability for these components (except, presumably, by experience gained), but if the expected statistical life expectancy were sufficiently long, the actual failure would have proved to be sufficiently low during the initial years of machine operation. The sense gained from today's view of this approach was that equipment would be replaced or superseded before components substantially failed. The second approach was to extend the operating life of computing equipment as far as possible by detecting incipient failures before they occurred. As an independent and retrospective view 30 years later (about 1980), this approach was used productively with many mainframe systems.

Also at this early-1950s period, computing equipment failures were expensive and annoying given the value of the operating time of these machines. Nevertheless, other failures caused problems, such as those causing intermittent machine errors, for instance. In relay computers, approximately 85 percent of the failures were intermittent. For the ENIAC, the majority of failures were intermittent. To detect this problem, in the Bell Telephone Laboratories machines, continuous automatic checking, or low-level checking, was done at the level of the individual relay. Detection of a failure would cause the results to be held in their respective relays until the fault was cleared, upon which computation would continue with the results obtained previously. For the ENIAC, checks were made of the solutions or partial solutions to problems, a process termed high-level checking. Operation of the ENIAC was so fast that holding results in stasis was deemed impractical and results were recalculated after a failure.

ERA makes the following comment regarding the ENIAC -

It has been estimated that the ENIAC is actually in useful operation on problems about 50 percent of the time and that, allowing for a complete repetition of each problem, it therefore turns out about one-fourth as many useful answers as could an ideal, perfectly reliable, machine with the same basic mathematical characteristics, which required no maintenance whatever. This is an excellent performance record.

  • 1
    Welcome to Retrocomputing. Good find!
    – DrSheldon
    May 17, 2021 at 18:30

Tubes are inherently way less reliable than relays. Simply as they decay by operating hours, while relays only decay by the product of numbers of switching operation and (non liner) value of current switched. This is especially true with relays used strictly for logical operations, at maximum handling the current of only one (or a few) coils of other relays.

This is (next to) independent of technological advances.

Relays based machines got replaced by tube based for the very same reason horse drawn carts got replaced by inherently less reliable, but faster, steam and later petrol fuelled lorries:

  • Single deliveries got handled in shorter time,
  • thus more deliveries could be made before the next breakdown.

Lifespan (and thus MTBF) is only a relevant metric within the same technology/use case. But the use case of tube vs. relays is one about performance delivered, not durability.If at all, then a

Mean Load Between Failure

might be a more appropriate yardstick.

  • I would expect that the MTBF of tube-type computers could have been improved if one ran each tube in a tester for a certain amount of time before putting it into a machine, and replaced tubes before they accumulated enough hours of usage that failure would become likely. A problem with doing so, however, is that even if tubes that survived 10 hours could be guaranteed never fail before they'd run 100 hours, replacing every tube after every 100 hours of operation would be impractical. Most tubes would last well beyond the point where a few start to fail, so the only way to get useful life...
    – supercat
    May 16, 2021 at 17:33
  • 1
    ...out of tubes is to keep running them well past the point where failures would start to occur.
    – supercat
    May 16, 2021 at 17:33
  • @supercat that's what has been done anyway. Or at least was possible, as one could buy them either with standard burn in, or extended. One problem abut burn in is that it takes away operation hours.
    – Raffzahn
    May 16, 2021 at 17:44
  • I am recalling that many early tube computers were designed such that the stresses on the tube were only about 10% of what would be considered "normal".
    – Hot Licks
    May 16, 2021 at 18:55
  • 1
    @Raffzahn: My point is that there was probably a considerable difference between the length of time one could use a tube before the probability of failure in a given hour would reach 1 in 10,000,000, and the length of time one could use it before the probability of failure in a given hour would reach 1 in 10,000. Replacing vacuum tubes only when they failed would result in many computations being ruined by tube failures, but wouldn't require replacing nearly as many tubes as...
    – supercat
    May 16, 2021 at 19:08

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .