How Are Chips Cooled? Heat Sinks, Heat Pipes and Limits

In everyday life

Play a game on your phone for a while, and the phone gets warm. After some time, it may even feel a little sluggish. On a laptop, the fan gets louder when you start a heavy task.

All of these come down to the heat the chips inside give off, and the parts that get rid of it. News about the power and cooling of data centers is the same story on a much bigger scale. This article follows heat from where it is made to where it leaves.

Why do chips get hot?

Inside a chip, billions of transistors switch between 0 and 1. Each switch puts a tiny amount of charge into the wiring and then lets it go. When current flows through a narrow path, the path's resistance turns electrical energy into heat. An electric heater warms up in the same way.

The power used by this switching is called dynamic power. The faster the switching, and the more transistors switch at once, the more heat there is.

A chip makes heat even when it is not switching. A small current keeps leaking through transistors that should be off. The power lost to this leakage is called static power. Leakage rises quickly as temperature goes up, so a hot chip tends to make even more heat.

A chip computes, but it hardly moves anything or gives off light. So almost all the electrical energy that goes in ends up as heat. "A chip that uses 100 watts" means nearly the same as "a chip that gives off 100 joules of heat every second." A joule is a unit of energy, and a watt is one joule per second. Here we leave out the tiny amount of energy that leaves the chip as signals.

What goes wrong when a chip gets hot?

Every chip has a temperature it must not go above. The hottest part is the layer where the transistors sit. Its temperature is called the "junction temperature," and each product has its own limit. According to Intel, the limit is 100–110 ℃ for many of its processors.

Near the limit, the chip lowers its own speed (frequency) and voltage to make less heat. This is called thermal throttling. If the temperature still does not drop, the chip shuts off its power automatically to avoid damage. Without enough cooling, a chip cannot reach its full speed.

The other problem is lifetime. At high temperatures, wear speeds up, such as metal atoms in the wiring slowly moving. For how temperature and the rate of wear are linked, see "Arrhenius Equation in Chips." The common saying "10 ℃ hotter halves the lifetime" is only a rough guide that depends on conditions.

The heat path: from chip to air

Let's follow the heat in a PC's CPU. The CPU itself is a small piece cut from a round disk of silicon, called a wafer. This piece is called the "die." The die sits upside down on a small board with wiring in it, called the substrate. So the back of the die faces up.

A metal lid sits on the back of the die. The lid spreads the heat of the small die over a wider area, so it is also called a heat spreader. On top of it, a heat sink is pressed down: a block of metal with rows of fins. Heat is handed on in order: die → lid → heat sink → air.

The "TIM" in the figure is a material that fills gaps. Metal surfaces look flat, but up close they are bumpy. When two are stacked, tiny pockets of air remain between them. Air conducts heat very poorly, less than 1/10,000 as well as copper. So a paste or other material that conducts heat well is put in the gaps. This is called a TIM, short for thermal interface material.

A heat sink has rows of fins to give it more surface touching the air. That matters because heat moves into the air only at the metal's surface. For what the package around a chip does, see "What Is Semiconductor Packaging?"

Thermal resistance: a number for how hard heat flows

Each step on the heat path makes it harder for heat to flow, to some degree. This is called thermal resistance. Its unit is ℃/W. It tells you the temperature difference between the entry and exit when 1 watt of heat passes through. The smaller the number, the more easily heat flows.

The temperature difference equals the amount of heat times the thermal resistance. For example, say the thermal resistance from the heat sink to the air is 0.3 ℃/W. Passing 100 watts of heat creates a difference of 100 × 0.3 = 30 ℃. If the room air is 25 ℃, the heat sink is about 55 ℃.

The part from the die to the heat sink also has thermal resistance. Say the die, TIM, lid and paste add up to 0.2 ℃/W. That adds another 20 ℃. Thermal resistances along a path add up, so the die is at 25 + 30 + 20 = 75 ℃. That still leaves room below the 100 ℃ limit. These numbers are only examples. Real values depend on the product and how it is mounted.

If the fan stops, the thermal resistance from heat sink to air rises to, say, 1.0 ℃/W. Then the die would reach 25 + 100 + 20 = 145 ℃, above the limit. In practice, thermal throttling kicks in before that and cuts speed until the heat drops. You can try this calculation from power and thermal resistance with the junction temperature calculator.

Fans and heat pipes

A fan replaces the air around the heat sink. Air warmed at the fins flows away at once, and cool air keeps touching them. So heat moves into the air more easily. In other words, a fan lowers the thermal resistance from heat sink to air. The trade-off is noise: the faster it spins, the louder it gets.

In laptops, the chip and the fan are often far apart. So a thin copper tube carries the heat over the distance. This is called a heat pipe. The air inside the tube has been pumped out, and it holds a little liquid, usually water.

At the hot end, the liquid evaporates into a gas. The gas moves through the tube to the cold end, where it cools and turns back into liquid. As it does, it gives the heat it carried to the fins. The liquid soaks into fine grooves or mesh on the inner wall and is drawn back to the hot end. It is the same force that pulls water up a tissue when you dip one corner in water.

This carries heat well because a liquid absorbs a lot of heat when it turns into a gas. It is the same reason your body cools as sweat dries. With the air pumped out, the pressure inside is low, so water evaporates readily below 100 ℃. The liquid returns by the wall's wicking force, not by gravity, so the pipe works in any position. But how much heat it can carry does change with its position.

Phones: cooling without a fan

Most phones have no fan and no fins. These would not fit in a thin body, and holes are hard to add when the phone must keep out water and dust. Instead, thin plates and sheets spread the chip's heat across the whole body. The heat then escapes slowly from the body's surface into the air.

One of the parts that spreads heat is a flat heat pipe. This is called a vapor chamber. A thin tube carries heat in one direction, but a plate can spread heat across a whole surface. For example, Apple announced the iPhone 17 Pro in September 2025. Apple says its vapor chamber, sealed with water, is welded into the aluminum body.

Even so, a phone can only get rid of so much heat. Its surface cannot get hotter than a person can hold. During long gaming sessions, thermal throttling lowers the speed to make less heat. Apple's support page says the same. If the phone gets too hot, it changes how it runs to control its temperature. If it gets hotter still, it shows a warning screen.

PCs and GPUs: hundreds of watts of heat

A PC's CPU or GPU gives off far more heat than a phone chip. Cooling is designed around a guide figure: the heat a chip gives off when it keeps doing heavy work. This figure is called TDP (thermal design power). Since its 12th-generation Core chips, Intel has called this value "Processor Base Power."

For example, Intel's Core i9-14900K, released in 2023, has a base power of 125 watts. When it speeds up, it uses up to 253 watts. Its junction temperature limit is 100 ℃.

GPUs, used for graphics and AI computing, give off even more heat. NVIDIA released the GeForce RTX 5090 in 2025. Its power limit for the whole card is 575 watts. The RTX 4090 of 2022 had a limit of 450 watts. When not doing heavy work, these cards use less power than this. Cards like these are cooled by a large heat sink with heat pipes or a vapor chamber, plus several fans. For why GPUs are used for AI, see "Why Are GPUs Used for AI?"

Data centers: from air to liquid

In a data center, thousands of servers holding chips like these stand in rows. Servers are stacked on tall shelves called racks. In racks packed with AI GPUs, air alone can no longer carry away all the heat. For how a whole building is cooled, see "What Is a Data Center?"

That is why using liquid to carry heat is spreading. For the same volume, water can store about 3,500 times as much heat as air. A common method presses a metal plate with liquid flowing inside it right on top of the chip. The heat the plate picks up travels with the liquid through pipes out of the server.

Another method sinks whole servers in a liquid that does not conduct electricity. This is called immersion cooling. In 2021, Microsoft announced it had started sinking servers in a liquid that boils at about 50 ℃. The chips' heat makes the liquid boil. The vapor is cooled by the tank's lid, turns back into liquid and falls back down. It is the heat pipe's "evaporate, condense and return" idea, used on a whole tank.

Common myths

By the numbers

Chip power figures change with each new generation. The product figures above are the specs at each launch.

Learn more

• Try the math: the junction temperature calculator (find a chip's temperature and margin from its power and thermal resistance), the CMOS power calculator (find switching and leakage power)
• See the differences in a table: BGA vs QFP vs QFN (the package shape also changes how heat escapes), Wire bonding vs. flip chip
• See it in figures: all figures (including a model of thermal resistances joined in a chain)
• Terms for the next step: DVFS (changing voltage and speed to match the workload), power capping (setting a limit on power and holding to it)
• Related articles: "What Chips Are Inside a Smartphone?" and "What Is a Power Semiconductor?"