What Is MTBF? Mean Time Between Failures for Industrial Computers

MTBF, or mean time between failures, is the average time a repairable system runs between one failure and the next, worked out across a population of identical systems. It is quoted in hours, and the higher the figure, the less often you should expect failures across a fleet.

For an industrial computer the number matters because a failure is rarely a quick swap. A computer on a production line, in a trackside cabinet or at a remote pumping station can mean a stopped process, lost data and an engineer sent to site, and those costs are often far higher than the price of the computer. Buyers use MTBF to estimate how often they will have to deal with that.

It is also one of the most misread figures on a datasheet. A high MTBF does not mean a single system will last that long, and two figures can only be compared when they were worked out the same way and at the same temperature. This guide explains the method behind the number and the design choices that move it, including why fanless computers tend to predict well.

Every MTBF calculation uses a small set of formulas. They all assume a constant failure rate, the flat middle of a product’s life, and they convert between the figures you will see on datasheets and the ones you need for planning.

The worked values follow the fleet example below. The sections that follow explain where each figure comes from, how predictions differ from field data, and how to use the numbers when you specify a computer.

At its simplest, MTBF is total operating time divided by the number of failures in that time. As a worked example, if 50 systems each run for 2,000 hours, that is 100,000 operating hours; if three of them fail in that period, the observed MTBF is about 33,000 hours.

The same figure can be turned round into a failure rate, written λ (lambda): λ = 1 ÷ MTBF, in failures per hour. Component makers often quote failure rates in FIT, failures per billion hours, so a FIT value converts to MTBF as 1,000,000,000 ÷ FIT. Failure rates are useful because they add: the failure rate of a computer is roughly the sum of the failure rates of everything in it, which is why every extra card, drive or connector pulls the system MTBF down a little.

From the failure rate you can estimate reliability, the chance that a system runs for a given time t without failing: R(t) = e^−t/MTBF. That formula assumes a constant failure rate, which describes the flat middle of the well-known bathtub curve. Early-life failures from manufacturing faults sit before it, and wear-out failures from ageing parts sit after it, so MTBF says nothing useful about either end.

To calculate MTBF from your own records, work through four steps.

1. Set the boundary. Decide what counts as the system, for example the computer alone or the computer with its power supply and cabling, and what counts as a failure.
2. Add up the operating hours. Multiply the number of systems by the hours each one actually ran in the period, using the real duty cycle rather than calendar time.
3. Count the relevant failures. Leave out transit damage, misuse and no-fault-found returns, and say that you have.
4. Divide. MTBF = total operating hours ÷ number of failures.

A prediction runs the same arithmetic from the parts list upwards: add the failure rates of the parts, then take the reciprocal of the total. The figures below are round values chosen to show the method, not data for any product.

In this example the fan alone takes the predicted MTBF from around 333,000 hours down to 200,000, which is why fanless designs tend to predict well.

A fleet example

Take 200 fanless industrial PCs running around the clock for a year: 200 × 8,760 hours gives 1,752,000 system-hours. If four of them fail in that year, the observed MTBF is 1,752,000 ÷ 4 = 438,000 hours. Turned round, that predicts an annualised failure rate of 1 − e^{−8,760/438,000}, about 2.0%, or roughly four failures a year across the same fleet.

Two cautions apply. Four failures is a small sample, so the true MTBF could easily be half or double that figure; reliability engineers put confidence limits on results like this, usually with the chi-squared distribution. And the result only holds while the failure rate is steady: if most of those failures came in the first few weeks, the fleet is showing early-life problems rather than its long-term rate.

FIT and MTBF describe the same thing from opposite ends. FIT, or failures in time, is a failure rate: the expected number of failures per billion (10⁹) device-hours, so a higher FIT means a less reliable part. MTBF is the average time between those failures, so a higher MTBF means a more reliable one. With a constant failure rate, each is the reciprocal of the other:

• MTBF in hours = 1,000,000,000 ÷ FIT
• FIT = 1,000,000,000 ÷ MTBF in hours

Semiconductor makers quote FIT because their rates are small: a chip at 10 FIT has an MTBF of 100 million hours, a figure that says little on its own. System builders quote MTBF because it reads naturally for a complete computer. FIT has one practical advantage, shown in the example above: rates add directly, so the FIT values of every part sum to the rate for the system.

The last column, the annualised failure rate (AFR), is often the most useful for planning spares: AFR = 1 − e^{−8,760/MTBF} for a system running all year. Like MTBF, a FIT value is only comparable when the method and temperature match. Semiconductor FIT figures usually come from accelerated life tests, calculated to JEDEC JESD85 and quoted at a stated junction temperature.

Most MTBF figures on a datasheet are predictions, made before enough systems have been in service to measure anything. The engineer builds a list of every component on the board, looks up a base failure rate for each part type in a reliability handbook, then adjusts it for temperature, electrical stress, quality level and the operating environment. Adding the adjusted rates gives λ for the assembly, and MTBF follows as 1 ÷ λ.

Two handbooks dominate:

• MIL-HDBK-217F. The US military handbook for reliability prediction of electronic equipment. Its last revision, Notice 2, dates from 1995, so its component data is old, and it tends to give conservative figures for modern parts. It defines environment categories such as ground benign, ground fixed and ground mobile, and the category chosen changes the result considerably.
• Telcordia SR-332. Developed for telecoms equipment, originally as Bellcore TR-332, and revised more recently. It allows laboratory and field data to be combined with the parts count, and for the same design it usually predicts a higher MTBF than MIL-HDBK-217F.

The practical point is that the same hardware can carry very different figures depending on the method, the reference temperature, the environment category and what the number covers. Our own product data shows the kind of detail to look for: Diamond Systems quotes its EPS-8130-XT PC/104 Ethernet switch at 624,170 hours at 20 °C, and gives separate figures for the main board and daughter board of its EPS-24026-104 switch, both at 20 °C. A stated temperature and a clear boundary make a figure useful; a bare number does not.

Reliability prediction standards compared

MIL-HDBK-217F and Telcordia SR-332 are the methods most often quoted on datasheets, but they are not the only ones. Knowing which was used tells you how to read the figure.

How temperature changes a prediction

Most prediction models apply a temperature factor based on the Arrhenius relationship, which describes how the rate of a chemical or physical failure mechanism rises with temperature. The size of the effect depends on the activation energy of each mechanism, so the figures below use a single, commonly assumed value to show the shape of the curve rather than a figure for any product.

With an activation energy of 0.7 eV, a value often assumed for semiconductor mechanisms, the failure rate relative to 25 °C is about 3.7× at 40 °C, about 12× at 55 °C, about 36× at 70 °C and close to 100× at 85 °C.

A real computer mixes parts with different sensitivities, so its overall figure moves less steeply than a single mechanism, but the direction is the same. That is why two MTBF figures for the same hardware, one at 25 °C and one at 40 °C, can differ several times over, and why the stated temperature matters as much as the number.

Field MTBF is measured rather than predicted: the operating hours of systems in service divided by the confirmed failures among them. It reflects real installations, real power supplies and real handling, so it is the more honest figure, but it takes a large population and years of service to mean much.

It is also sensitive to how it is counted. Operating hours depend on duty cycle, and a fleet running around the clock builds hours three times faster than one running a single shift. Returns that turn out to be wiring faults, damage in transit or no fault found may or may not be counted as failures. When a supplier quotes field data, ask how many systems it covers, how hours were estimated and what counted as a failure.

What is a good MTBF for an industrial computer?

There is no single target, because the figure depends on what it covers and how it was worked out. As a guide, predicted MTBF figures for fanless industrial boards and network equipment in our catalogue mostly fall between about 240,000 and 640,000 hours at 20–25 °C, and some simple I/O cards and modules are quoted well above a million hours. The table in the products section below lists published figures with their conditions.

Compare like with like: the same method, the same temperature and the same boundary. A figure for a board alone will always be higher than one for a complete computer with its storage, power input and cabling, and a higher MTBF is better only when the two figures were produced the same way.

The EPS-8130-XT figure of 624,170 hours works out at about 71 years of continuous running. Nobody expects that switch to run for 71 years, and the figure does not claim it will. It describes the random failure rate during the useful life of a population, at the stated temperature.

Read that way, it is useful. For a fleet of 100 of those switches running around the clock at 20 °C, the expected number of failures in a year is 100 × 8,760 ÷ 624,170, or about 1.4. Put another way, R(1 year) = e^{−8,760/624,170}, about 0.986, so each switch has roughly a 98.6% chance of a failure-free year under those conditions.

Service life is set by other things: the parts that wear out and the time the product stays available. Electrolytic capacitors dry out with age and heat, and a widely used rule of thumb says their life roughly halves for every 10 °C rise. Fan bearings wear, SSDs have a finite write endurance and real-time clock batteries run down. A warranty is a commercial promise, separate from both MTBF and service life.

The bathtub curve

Failure rates over a product’s life are often drawn as a bathtub curve. Early life has a falling failure rate, as manufacturing defects show up and are removed. Useful life is the long, flat floor, where failures are random and the rate is roughly constant. Wear-out is the rising edge at the end, as fans, capacitors, batteries and flash memory reach the end of their lives. MTBF and the formulas in this guide describe only the flat floor. They say nothing about how soon wear-out begins, which is why a figure of hundreds of thousands of hours sits comfortably beside a service life of 10 years or so.

Reliability over time

Under the constant failure rate assumption, the chance that a single system runs for a given time without a failure follows R(t) = e^−t/MTBF. The table applies it to continuous running, 8,760 hours a year.

Read the later columns with care: by year 10 many products are into wear-out, where the real failure rate rises above the constant rate the table assumes. Our FAQ on how long computers last covers service life, and SSD endurance, a common wear-out limit, is explained in SSD TBW.

Three related terms are often used loosely. They measure different things.

In practice many datasheets quote MTBF for parts that are strictly described by MTTF, and the two are treated in much the same way. MTTR is different, because it depends on your support arrangements rather than the hardware. Availability, the share of time a system is working, is MTBF ÷ (MTBF + MTTR), so cutting repair time improves it just as raising MTBF does. Spare systems on site, a prepared system image and remote management all shorten MTTR.

A fourth term, MTTD (mean time to detect), measures how long a failure goes unnoticed. On a remote site it can be the largest part of the outage, which is why monitoring belongs in the reliability plan alongside the hardware. Our guide to AI predictive maintenance shows how monitoring data is used to catch faults before they stop a machine.

Availability is the share of time a system is working, and for a single repairable system it is A = MTBF ÷ (MTBF + MTTR). It is often quoted as a number of nines.

Back to the fleet example. With an MTBF of 438,000 hours and an MTTR of 24 hours, the time it takes to get an engineer to site and swap the computer, each system’s availability is about 99.995%. Across 200 systems that is still about four failures and 96 hours of lost service a year. Keep a configured spare on site with a prepared system image and MTTR might fall to two hours, cutting the lost service to around eight hours without changing the hardware at all.

That is the practical lesson: for the people running a site, cutting repair time is often cheaper than raising MTBF. Each avoided truck roll saves time and travel, a watchdog timer restarts a computer that has stopped responding, and out-of-band management lets an engineer power-cycle, reset or reimage a system remotely.

When a system is built from several parts, how they are connected changes the overall MTBF. In series, where any one failure stops the system, failure rates add, so the system MTBF is lower than that of its weakest part. In parallel, where a spare takes over, the system keeps running after the first failure and its MTBF rises.

The last row shows why redundancy with prompt repair is so effective, and also why its figures need care. The calculation assumes the two parts fail independently, but in practice common causes, such as a shared input supply, the same overheated cabinet or a single controller, dominate. Redundancy is still the right answer where uptime matters: dual power inputs, mirrored storage with RAID, network redundancy with ring or dual-homed switches, and a hold-up supply that rides through short power cuts, which our guide to supercapacitor vs battery UPS compares.

The design choices that improve MTBF are the same ones that make a computer suited to a cabinet, a machine or a vehicle.

• No fans. Fans are among the few moving parts in a computer, and their bearings wear, faster still in heat and dust. A fanless chassis moves heat through heat pipes and conduction to a finned case instead, which removes that wear part altogether. It is one of the advantages of fanless computers that matters most on remote sites.
• Solid-state storage. An SSD has no spinning platters or moving heads. Industrial SSDs add wide temperature ratings and power loss protection, so a sudden power cut is less likely to corrupt data.
• Industrial temperature-rated parts. Components rated across the industrial temperature range have more margin when the installation runs warm, and prediction models reward that margin.
• Derating. Running parts well inside their voltage, current, power and temperature ratings lowers their failure rate. For the processor this means choosing a processor TDP comfortably inside what the chassis can dissipate, rather than at its limit.
• Wide-temperature, wide-voltage design. Solid polymer capacitors in place of liquid electrolytics, protected wide-range DC inputs and fewer internal cables and connectors all reduce the number of things that can fail.

Heat runs through most of these points. Prediction handbooks apply a temperature factor to almost every component, so a system that keeps its internals cool predicts a higher MTBF as well as performing better. Our guides to thermal throttling and active and passive cooling cover how that heat is managed.

MTBF is one input to a decision, not the decision itself. Work through four questions.

• Environment. What is the hottest the air around the computer will get, and is there dust, damp or vibration? Ask for an MTBF figure at a temperature close to yours, and check the environment category behind it.
• Performance. Choose the processor that does the job with thermal headroom. A system running at its thermal limit all day ages faster than one running well within it.
• Expansion. Every add-in card, drive and module adds its own failure rate. Fit what the application needs and leave out what it does not.
• Lifecycle. Plan spares, a recovery image and how quickly you can restore service, because MTTR matters as much as MTBF to the people running the site.

Would a failure on this site mean sending someone out? If so, reliability deserves as much weight as performance in the specification.

We specify fanless industrial computers for exactly these installations, matching the processor, storage and temperature rating to the site, and we supply each system set up and tested as standard. Our guide to installing a fanless industrial PC covers the mounting and airflow details that keep it running cool once it is in place.

These figures come from the manufacturers’ own data for products in our catalogue. They show how much the stated conditions vary, and why a figure is only useful with its method, temperature and boundary attached.

The simplest parts carry the highest figures, because a small card has few components to add up. A complete computer includes a board like those above plus memory, storage, a power input stage and connectors, so its system MTBF is lower than any one of them. We configure, image and test every computer we supply, and we can request the manufacturer’s MTBF report, including its method and temperature, for a specific build.