You open a laptop, kick off a build and for two minutes everything flies. Then, with nothing broken and no application closed, the fan roars and the machine seems to be driving with the handbrake on. The temptation is to blame the fan, the manufacturer or the operating system. The reality is more interesting: inside the silicon itself there is a network of sensors and a set of rules that decide, thousands of times per second, at what frequency and voltage each core is allowed to run. The fan is not in charge. It takes orders.
Five diodes measuring the temperature of the silicon
A processor does not have a thermometer, it has many. Every core carries one or more thermal diodes: semiconductor junctions biased with a constant current whose forward voltage drop (VBE) falls almost linearly with temperature, roughly two millivolts per degree Celsius. An analog-to-digital converter reads that voltage and turns it into degrees. We call them DTS, Digital Thermal Sensors, and they do not measure the air in the case or the surface of the heatsink: they measure the silicon where the heat is actually generated, at the hottest point of the core, the so-called hot spot.
That detail explains why a processor can sit at 45 °C on the outside and 90 °C on the inside. It also explains why different monitoring programs disagree: there is no single “processor temperature”, there are dozens of sensors spread across cores, caches, the memory controller and, on modern chips, each block of the integrated GPU. A single aggregated figure is always an average, and an average hides the point that matters.
The hardware also has its own channel to report what it sees. Intel defines PECI (Platform Environment Control Interface), a single-wire serial bus that the chipset and the motherboard’s embedded controller use to ask for temperatures without going through the operating system. That controller, the EC, is the one driving the fan curve from that data: which is why on many machines the fan starts spinning before the operating system has finished booting. Firmware knows first; the operating system finds out later.
A cry for help called PROCHOT
Each chip family defines its ceiling: Tjmax, the maximum junction temperature. It is commonly 100 °C on many Intel mobile and desktop processors and 95 °C on AMD’s. As the chip approaches it, it does not shut down at once: first it warns. It does so by asserting a signal called PROCHOT (or its internal equivalent), which triggers the TCC, the Thermal Control Circuit. That circuit does not lower voltage or switch anything off; it modulates the clock, letting some cycles through and discarding others. The processor still “runs” at its nominal frequency, but with the clock active only a fraction of the time, in discrete steps of one eighth. What the user feels is exactly that: everything slows down while no classic metric gives it away.
If the temperature keeps climbing anyway, the last resort kicks in: THERMTRIP. That one really does cut power and shut the system down. It is designed as a safety net, not as a control tool: it fires when the previous step has already failed, for instance if the heatsink has come loose.
P-states: the voltage and frequency dial
Before heat ever forces a slowdown, the chip is already choosing where to operate. Each operating point is a P-state: a frequency and voltage pair. Voltage is not set with a potentiometer; it is requested through a voltage identification table (VID) that the processor itself sends to the motherboard’s regulators (VRM) over a fast serial bus. Dropping 100 MHz without dropping voltage saves almost nothing; dropping voltage is what really cuts consumption, because dynamic power depends on the square of the voltage. That is why fine-grained voltage tuning is the heart of DVFS, Dynamic Voltage and Frequency Scaling, and why the scaling is not a ladder of frequencies but a search for the minimum stable voltage at any given moment.
On Linux that policy is carried out by a governor: schedutil on current machines, previously ondemand or powersave. You can watch what is happening with two or three commands:
cat /sys/devices/system/cpu/cpu0/cpufreq/scaling_cur_freq
cat /sys/class/thermal/thermal_zone*/temp
cat /sys/devices/system/cpu/cpu0/cpufreq/scaling_governorThe first value gives the actual frequency in kilohertz; the second, each thermal zone’s temperature in thousandths of a degree. Comparing them while a load runs is the most honest way to tell “the chip is limited by temperature” from “the chip is limited by power”.
The real limit is not temperature, it is power
This is the part that usually surprises people. A modern processor rarely slows down because it is hot. It slows down because it has used up its power budget. That budget is declared through the limits of the RAPL interface: PL1 is the sustainable power (the one it can hold indefinitely), PL2 is the peak allowed for a limited time and Tau is how long that peak lasts. These are the famous “45 W” figures on laptop spec sheets: they are not a performance measure, they are a heat-dissipation promise the manufacturer uses to size the cooler. A chip can run at 60 W for 28 seconds and then drop to 35 W without ever reaching its maximum temperature.
This has a practical consequence: on a machine with good cooling but conservative power limits, you will see low temperatures and low frequencies at the same time. And conversely, in a thin laptop with a powerful chip, you will see the classic benchmark pattern: a few seconds of glory and then a long, lower plateau. The plateau is not a defect. It is the contract the firmware signed with the cooler.
What also counts in a phone
In a phone the problem is inverted. There is no fan and the chip must not exceed a certain temperature because the user is holding it. That is why phones carry surface temperature sensors in addition to the silicon ones, and the operating system shares the budget between CPU, GPU and modem: if the camera records in 4K and the network is uploading data, all three compete for the same watts. The famous “it gets hot and slows down” is, in fact, real-time resource sharing.
Why it pays to understand this
Knowing that the clock is modulated in steps, that P-states hunt for minimum voltage and that PL2 and Tau define the real peak is useful for something very concrete: interpreting what you measure. A low frequency with low temperature points to power limits or an undersized charger. A low frequency with high temperature points to dissipation: dried paste, dust in the vents, a badly tuned fan curve. And a high frequency with a light load usually means a background process is keeping the chip out of a deep idle state. A processor’s thermostat is not there to annoy you: it is the only reason a 100-watt chip fits into something we rest on our laps. Understanding it turns a vague complaint (“it feels slow”) into a diagnosis with a name and a surname.
