What Is Thermal Throttling? Why Computers Slow Down When They Get Hot

Thermal throttling is a processor reducing its clock speed, and usually its voltage, when its silicon reaches the maximum temperature it is designed to run at. Lower clocks mean less power and less heat, so the temperature stops rising. The alternative would be damage or a shutdown, so throttling is a safety feature rather than a fault.

The trouble is when it happens. A processor throttles when it is working hardest, which is exactly when an industrial application needs its full performance. A vision inspection system drops frames, inference latency rises, and a control loop that met its cycle time on the bench starts to miss it in a warm cabinet.

Desktop PCs and laptops deal with heat by spinning their fans faster. Fanless PCs have no fan to speed up: their cooling is fixed when the chassis is designed. The way to avoid throttling in a fanless system is to match the processor, the chassis and the installation from the start. This guide explains what triggers throttling, how to recognize and measure it on Windows, Linux and Jetson, why SSDs and GPUs throttle too, and how to choose a system that holds its speed.

Several different limits can slow a computer down, and they look alike from the outside: the clock drops and the work takes longer. Telling them apart is the first step to fixing the right thing.

Strictly, only the first two are thermal throttling. Power-limit throttling is a deliberate design choice that keeps a processor inside the heat its cooling can remove, and in a well-specified fanless system it is the normal state under sustained load.

Junction temperature is the temperature of the silicon die itself, measured by digital thermal sensors built into the processor. The datasheet gives a maximum, Tj max, which is often around 100 °C for Intel Core processors; embedded and mobile parts vary, so check the datasheet for the exact part.

When a core reaches that limit, the processor’s thermal control steps the frequency and voltage down until the temperature holds. On Intel systems the same mechanism can be triggered from outside the processor through the PROCHOT signal, for example when the voltage regulators on the board run hot. If temperature keeps rising regardless, a critical thermal trip shuts the system down to prevent damage.

The die temperature depends on a chain. Heat leaves the die through the heat spreader and a thermal interface material, into a heatsink or the chassis, and then into the surrounding air. Each step has a thermal resistance, written θ, and the die settles at the ambient temperature plus the power multiplied by the total resistance of the path.

Take a worked example. A processor running at 28 W through a path of 2 °C per watt sits 56 °C above the air around it. At 25 °C ambient the die is at 81 °C, comfortably below a 100 °C Tj max. In a 45 °C cabinet the same system reaches 101 °C and throttles. Nothing about the computer changed; only the air around it did.

TDP, which Intel now calls Processor Base Power, is the average power in watts that a processor dissipates under a sustained, demanding workload at its base frequency. It is the figure a cooling system is designed to remove, and it is not the most power the processor can draw. Our white paper on how processor TDP and base power work goes into the choice of TDP in more depth.

Intel processors manage power with two main limits:

• PL1. The long-duration power limit, normally set to the TDP. The processor can run at this power indefinitely if the cooling can remove it.
• PL2. The short-duration limit, which Intel calls Maximum Turbo Power. The processor may draw up to PL2 for a time window, often called Tau, before dropping back to PL1.

AMD uses its own names for similar limits, such as PPT, TDC and EDC on Ryzen processors, and NVIDIA Jetson modules use selectable power modes that set a power budget for the whole module. On many embedded processors the limit can also be lowered with configurable TDP, so a part fits a smaller thermal budget at a lower but steady speed. On hybrid Intel processors the scheduler also decides which work runs on performance and efficient cores, which our guide to P-cores and E-cores explains.

Dropping from PL2 to PL1 is power-limit throttling, not thermal throttling, although both lower the clock. Thermal throttling is when the cooling cannot hold the die below Tj max even at PL1. Board makers can set PL1, PL2 and Tau in the BIOS, and in an industrial build they are worth setting deliberately rather than left at whatever the board defaults to.

Throttling is easy to miss because nothing fails. Look for these signs:

• Performance falls after a few minutes. Frame rate, inference rate or cycle time is fine at first, then drops once the chassis warms through.
• Clocks fall while load stays high. Core frequency sinks well below base while utilization stays near 100%.
• Temperatures sit at a ceiling. Package temperature plateaus just under Tj max instead of continuing to rise.
• It is worse on hot days. The same workload runs fine in winter and struggles in summer, or only in one cabinet.
• Fans run flat out. On an actively cooled computer, fans at full speed under modest load suggest clogged fins or filters.

Monitoring tools confirm it. Each reports clocks, power and temperature over time, and most can show why the processor is limited.

A short benchmark finishes inside the turbo window and reports the best case. To find out whether a system throttles, test the way it will be used.

1. Run a sustained load. Use the real application, or a stress tool such as stress-ng on Linux, and keep it running until the chassis reaches a steady temperature. For a fanless enclosure that can take well over half an hour.
2. Log clocks, power and temperature. turbostat on Linux, HWiNFO on Windows and tegrastats on Jetson show core frequency, package power and temperature over time.
3. Check the throttle counters. On Intel systems Linux exposes core and package throttle counts under /sys/devices/system/cpu/cpu0/thermal_throttle/. A rising count means the processor has hit its thermal limit.
4. Allow for the real ambient. Test in a climatic chamber where you can, or account for the difference between the test room and the hottest day at the installation.
5. Compare with the requirement. The question is whether the sustained speed still meets the frame rate, cycle time or latency the application needs.

NVMe SSDs

Yes. An NVMe SSD reports a composite temperature, and the NVMe specification defines a warning and a critical threshold set by the drive maker, commonly somewhere in the 70s or 80s °C. As the controller approaches them, the drive slows its transfers to cool down, so a long write or a recording stream suddenly falls to a fraction of its rated speed. PCIe Gen4 and Gen5 drives are the most affected, because their controllers draw more power; our guide to PCIe generations covers that heat.

In a sealed fanless chassis there is no airflow across the drive, so the answer is conduction: a thermal pad from the drive to the case, or a heat spreader bonded to the chassis wall. Several of our systems cool their M.2 drives into the enclosure this way. Our guide to M.2 key types covers drive formats, and industrial temperature drives keep their rating at the top of the range.

GPUs and Jetson modules

Discrete GPUs reduce their clocks at a thermal slowdown temperature or when they reach their board power limit, and most GPU systems keep active cooling for that reason; our FAQ on iGPU vs dGPU explains the difference in heat. NVIDIA Jetson modules work within a power mode chosen with nvpmodel, and lower their clocks when the module approaches its thermal limits. A Jetson system that holds 15 W comfortably in a sealed case may throttle in a higher mode such as Jetson Super mode unless the chassis or a fan kit can take the extra heat. Our Jetson thermal design guide covers power modes and heatsink choices in detail.

In a fanless PC the chassis is the heatsink. A copper block or heat pipes carry heat from the processor to the case, and fins on the outside pass it to still air by natural convection and radiation. Our guide to how a fanless PC moves heat walks through that path in detail. Some rugged designs go further and use conduction cooling, passing heat through the mounting face into a cold plate or the vehicle structure.

The chassis designer chooses the fin area, the mass and the heat path so the system can remove its rated processor TDP at its maximum rated ambient temperature while keeping the die below Tj max with some margin. That is why a fanless chassis lists a maximum processor TDP, and why the same chassis is often offered with lower-power processor options for wider temperature ranges.

Turbo bursts above TDP are absorbed for a while by the thermal mass of the chassis. Once it has warmed through, sustained power settles at what the fins can shed. A well-matched fanless system runs at or near PL1 all day without reaching Tj max. Running cooler helps reliability as well as speed, as our guide to MTBF explains.

These are the makers’ own ratings, and they show the trade-off plainly: for a given chassis, power and ambient temperature pull against each other.

Every operating temperature range on a datasheet assumes something about the installation. A fanless system rated across the industrial temperature range may reach the top of that range only with particular processor options, only with industrial temperature SSDs and memory, or only with a stated airflow across the fins. Read the footnotes before you rely on the headline range.

Ambient means the air immediately around the chassis, not the room. Common reasons a system runs hotter than expected:

• Enclosed cabinets. The air inside a closed cabinet sits above the room temperature, more so when other equipment shares it; our guide to electrical enclosure cooling shows how to calculate it.
• Mounting. Fins mounted horizontally, pressed against a wall or boxed in by cable trunking lose much of their natural convection.
• Sunlight. Solar gain on a kiosk, vehicle or outdoor enclosure can add far more heat than the processor does.
• Dust. A layer of dust on the fins insulates them and raises the temperature over months.

Occasional throttling does no harm: it is the processor protecting itself, and it will not damage the CPU. Persistent throttling is a sign that the system is running at its limit, which costs performance and, because electronics age faster when hot, some reliability. In an industrial system the goal is a design that never needs to throttle at the hottest expected ambient.

Avoiding throttling comes down to four questions asked before the hardware is ordered.

• Environment. What is the hottest the air around the computer will get, including cabinet temperature and sunlight?
• Performance. Size the processor for the sustained workload, not the burst. A lower-TDP processor that runs at full speed all day can deliver more than a faster one that spends the afternoon throttled.
• Expansion. GPUs, NVMe drives and add-in cards add heat inside the same chassis, and NVMe SSDs throttle themselves when they get hot.
• Lifecycle. Plan for dust and access. A system with margin on day one keeps that margin as conditions change.

Will the computer run flat out for hours at a time, in a warm cabinet? If it will, thermal headroom belongs in the specification alongside cores and memory. Where the sustained power is beyond what a sealed chassis can shed, our comparison of active vs passive cooling sets out the alternatives, and our guide to fanless mini PCs covers the smallest designs.

We specify fanless industrial computers with the processor matched to the chassis and the installation temperature, set the power limits in the BIOS where the application calls for it, and configure, image and test each system before it ships.

These systems publish the conditions behind their ratings, which is what you need to judge sustained performance.