Thermal Design for Always-On Hardware: Your Heatsink Was Sized for a Device That Idles

Why the duty-cycle assumption quietly broke, and what it costs you at DVT. Your heatsink was sized for a device that idles — and it doesn't anymore. Steady-state thermal design, soak testing, and the DFM decisions that lock it in.

IDEA TO DESIGN

Engineering Team

8/9/20268 min read

case temperature vs run time
case temperature vs run time

In May 2026, a physics professor at Utah State University became briefly famous for pointing out that a data center is, thermodynamically, a heater. His analysis of the proposed Stratos campus in Box Elder County put its total thermal load near 16 gigawatts — 9 GW of compute plus 7 to 8 GW of waste heat from a 55%-efficient gas plant — and the number travelled a long way.

It shouldn't have been surprising. Every watt you put into silicon comes back out as heat. There is no third place for it to go.

What is worth your attention isn't the gigawatt figure. It's that the same shift driving those campuses — compute that runs continuously instead of in bursts — has already reached the products on your bench. And most thermal designs shipping today were sized for a load profile the device no longer has.

The assumption nobody wrote down

Open almost any thermal calculation from the last fifteen years and you'll find an unstated premise: the device works, then it rests.

A smart speaker wakes on a keyword and goes quiet. A camera detects motion, encodes thirty seconds, and idles. A handheld instrument takes a reading and sleeps. The heatsink was never sized for the peak — it was sized for the average, on the assumption that thermal mass absorbs the spike and the idle period dumps it back out.

That assumption was correct, and it was load-bearing. It let you use a smaller heatsink, a cheaper enclosure, thinner copper, and no fan.

Always-on inference broke it. When a device holds context, watches a sensor stream, or runs a model continuously, there is no idle period. The thermal mass fills up and stays full. The average converges on the peak.

This is not a niche problem for edge AI boxes. It shows up anywhere a firmware change turned an event-driven product into a polling one — and firmware changes don't trigger a thermal re-review.

A ten-minute test and a two-hour test are different tests

The most expensive mistake we see at NPI is not a bad heatsink. It's a good heatsink validated with a short test.

Small enclosed devices have thermal time constants measured in tens of minutes. A 10-minute bench run on a system with a 40-minute time constant reaches roughly 22% of its final temperature rise. It looks fine. It is fine — for ten minutes.

Run the same board for two hours and you're at 95% of the true steady-state value. On a design with a 55 °C rise above ambient, that's the difference between reading 37 °C and reading 77 °C.

Teams routinely ship on the first number. The failure then arrives in the field, in the second hour of operation, in a customer's warm room, and it looks like an intermittent firmware bug.

If your thermal test is shorter than three time constants, you have not measured your device. You have measured its warm-up.

What continuous load actually does to your BOM

Sustained temperature does not degrade parts the same way cycling does. The failure modes change, and so should the component selection.

Aluminium electrolytic capacitors. The Arrhenius rule of thumb — roughly halved life per 10 °C rise — is well known and routinely ignored at design time. A 105 °C, 5,000-hour part running at 85 °C core temperature has a useful life somewhere around 20,000 hours. That's just over two years of genuinely continuous operation. If your product is sold with a three-year warranty, the capacitor is the warranty.

Thermal interface material. This is the one most people get backwards. Pump-out — where TIM migrates out of the joint — is driven by cycling, by repeated CTE mismatch between die and heatsink. Continuous operation largely removes that mechanism and replaces it with a different one: dry-out. Volatiles and silicone oil bleed off at sustained high temperature, the material hardens, contact resistance climbs, and junction temperature drifts upward over months. A grease that performs beautifully in a cycling application can be the wrong choice for a device that never cools down. Pads, phase-change materials, and low-bleed formulations behave better here.

Lithium cells. Calendar ageing is temperature-driven and irreversible. A cell held near a warm SoC at 45 °C loses capacity substantially faster than the same cell at 25 °C, whether or not it is being cycled. If your battery sits next to your hottest component and your device now runs continuously, your battery is on a different ageing curve than the one in the datasheet. Placement that was acceptable under duty-cycled load may not be.

Solder joints. The nuance here cuts the other way. Continuous operation means fewer thermal cycles, so Coffin-Manson fatigue actually improves. What replaces it is creep at elevated mean temperature. Different mechanism, different qualification test — and a reason not to assume that passing a thermal-cycling profile tells you anything about a device that never cycles.

Silicon. Most SoCs will throttle rather than fail, which sounds like a safety net and isn't. Throttling converts a thermal problem into a performance problem, and performance problems get reported as software bugs. We have seen teams spend six weeks profiling code to explain a latency regression that was a heatsink.

Sealed enclosures: the IP rating tax

Nothing in thermal design conflicts with product requirements more directly than ingress protection.

An IP65 or IP67 rating means no vents. No vents means no convective airflow across the board. Every watt now has to leave by conduction into the enclosure wall and then by natural convection and radiation from the outer surface. Your entire thermal budget is set by the external surface area and the material you chose for cosmetic reasons.

The practical consequences:

  • A plastic enclosure is a thermal insulator. Filled or metallised plastics help at the margins; they do not turn plastic into aluminium.

  • If the enclosure is the heatsink, the internal heat path to it matters as much as the fin design. A 2 mm air gap between a hot component and the wall it is supposed to heat is a wall.

  • Conformal coating and potting change conduction paths, usually for the better and occasionally in ways nobody modelled.

  • The industry response is worth watching: at COMPUTEX 2026, YPlasma demonstrated a 200-micrometre solid-state plasma cooling actuator on a Jetson Orin Nano — no moving parts, sub-1 W draw, no dust ingress path, aimed squarely at sealed enclosures. It's early, but the fact that a category is forming tells you how many designs are stuck against this wall.

thermal image inside enclosure
thermal image inside enclosure

The five decisions that set your thermal ceiling

By the time a thermal problem is measurable, most of these are frozen. In order of how expensive they are to reverse:

  1. Enclosure material and external surface area. Locked at ID sign-off, months before anyone measures a temperature. This single decision caps your dissipation in a fanless design.

  2. Vent geometry — if you get vents at all. Where the aesthetic and the airflow fight. Vent patterns can be designed to serve both, but only if the conversation happens during ID, not after.

  3. Copper weight and layer count. Going from 1 oz to 2 oz outer copper, or adding a dedicated internal plane as a heat spreader, costs a few percent on bare-board price. Retrofitting it after DVT costs a stackup change, re-qualification, and a schedule.

  4. Component placement. Hot parts near connectors, near cells, near the user's hand, or clustered on the same side of the board. Cheap to fix in week two of layout, expensive in week twenty.

  5. Thermal via arrays under power components. Almost free at design time. Genuinely impossible to add later.

None of these are exotic. All of them get decided before anyone has thermal data, which is exactly why they should be decided against a written steady-state target rather than an intuition.

How to test for steady state

A workable protocol that doesn't require a thermal chamber:

Run to true steady state. Minimum two hours, ideally until three consecutive 15-minute readings agree within 0.5 °C. Log continuously; don't spot-check.

Test at realistic ambient, not lab ambient. Safety standards are evaluated at 25 °C. Your customer's utility closet is 35 °C. A device in a vehicle cabin sees more. Every degree of ambient is a degree of case temperature — the rise is what your design controls, not the absolute.

Instrument the enclosure, not just the die. On-die sensors tell you about throttling. Thermocouples on the outer surface tell you about certification and about the user's hand. You need both, and they are not related by a constant offset.

Test in the orientation and the environment the product ships in. Natural convection is orientation-dependent. A board tested flat on a bench and a board mounted vertically inside a wall bracket are two different thermal systems.

Distrust θJA. The junction-to-ambient figure in a datasheet is measured per JEDEC JESD51 on a standardised test board in still air. It is a comparison metric between parts, not a prediction for your design. Treating it as a prediction is one of the most common sources of a thermal budget that looks fine on a spreadsheet and fails on a bench.

The certification trap

Steady-state temperature isn't only a reliability question. It's a pass/fail gate.

IEC 62368-1 Clause 9 sets touch temperature limits by accessible surface material and expected contact duration — and the limits differ by material because heat crosses into skin at a rate governed by thermal effusivity, not surface temperature alone. Metal gets a stricter limit than plastic at the same temperature. GB 4943.1-2022 aligns closely for the China market.

The trap is the timing. Touch temperature is measured at steady state, under normal and abnormal operating conditions. A team that validated thermals with short bench runs arrives at the test lab with a device that has never been run long enough to reach the temperature it will be judged at. That discovery, at that stage, means a tooling change.

What this costs, by stage

Rough multipliers we see on real programmes:

The gap between the first row and the last is entirely a question of when someone wrote down a steady-state target and tested against it.

The short version

The physics has not changed. What changed is the duty cycle, and the duty cycle was doing more work in your thermal design than anyone documented.

If your product runs continuously — or if a firmware update made it run continuously — the honest first step is not a better heatsink. It's a two-hour test at realistic ambient, with the enclosure closed, in the orientation it ships in. Most of the time, that test tells you everything.

Frequently asked questions

What is a thermal time constant, and how do I estimate mine? It's the time a system takes to reach about 63% of its final temperature rise. Estimate it empirically: run the device, log case temperature, and find the time at which the rise reaches 63% of where it eventually settles. For small sealed consumer devices it is commonly 20 to 60 minutes. Three time constants gets you to roughly 95% of steady state.

How long should a thermal soak test run? At minimum two hours, or three thermal time constants, whichever is longer. The stopping criterion should be data-driven: three consecutive 15-minute readings agreeing within 0.5 °C.

Does a fanless design mean I can't run continuous workloads? No, but it means your dissipation ceiling is set by enclosure surface area and material, and that ceiling is fixed at ID sign-off. Fanless continuous designs are entirely achievable — they just need the thermal budget written before the industrial design is frozen, not after.

Why does thermal interface material choice change for always-on products? Cycling drives pump-out; sustained heat drives dry-out and oil bleed. A grease optimised for cycling applications can harden over months of continuous operation, raising contact resistance and junction temperature gradually. Pads, phase-change materials, and low-bleed formulations generally hold up better under continuous load.

Is IP67 compatible with high thermal loads? It's compatible, but expensive in thermal terms. With no vents, all heat leaves by conduction to the enclosure and then convection and radiation from the outer surface. Expect to spend budget on a metal enclosure, a solid internal conduction path to it, or an emerging solid-state cooling solution.

Can I use the θJA value from the datasheet to size my design? Not for prediction. θJA is measured on a JEDEC standard board in still air and exists to compare packages against each other. Real-world junction-to-ambient resistance in your enclosure is typically much worse. Use it to rank parts; measure to size the design.

Does thermal throttling protect my product? It protects the silicon. It does not protect your user experience, your warranty, or your reliability numbers — and it converts a hardware problem into what looks like a software bug, which is the most expensive way to find it.

At what point in development should thermal be reviewed? Before enclosure material and vent geometry are frozen. That is normally weeks before the first board exists, which is why the review has to be based on a written steady-state power budget rather than measurements.