# Our CPU read 210 °F, and I almost bought a cooler

My phone showed a build machine at 210 °F. Before buying coolers we checked how long heat had held the chip back. Over 18 days, 342 seconds.

Topic: Hardware
Tested: 2026-11-03
Revision: 1
By Roman Vaisman
Page: https://eng.vaisman.me/entries/measure-before-you-buy/

## In plain English, for moms

My phone told me one of our computers was running scary hot, and I was ready to buy it a better fan. Then we checked how often the heat had actually slowed it down: almost never. These computers are built to run hot, like a kettle that is supposed to boil. The fan money stayed in my pocket.

## In 30 seconds

- Our phone dashboard showed a CI build machine at 210 °F (about 99 °C) and 85% load. My first thought was new coolers.
- A modern desktop chip boosts until it reaches its heat limit, so hot is often by design, not a fault.
- The kernel's own throttle counters said heat held it back for 342 seconds in 18 days, about 0.02% of the time.
- We changed nothing. Before you buy a cooler, check how much time the chip actually spends throttled.

## Problem

I opened our phone dashboard and one of our CI build machines was sitting at 210 °F, about 99 °C, at 85% load. It made me nervous, and my first thought was new coolers.

It has an Intel Core Ultra 7 270K Plus under a dual-tower air cooler, a Thermalright Phantom Spirit 120 SE, in a Lian Li LANCOOL 207 Digital case. It runs at the board's default power limit of 250 W, which is inside Intel's spec of 125 W base and 250 W max turbo.

So was it cooking itself, or was it just doing its job?

## What I tried

| Idea | What I found |
| --- | --- |
| Buy a 360 mm liquid cooler | An Arctic Liquid Freezer III Pro 360 costs about $70 to $100. Reviews put it roughly 8 to 15 °C cooler, worth about 3 to 6% on long all-core runs and 0 to 2% on short CI bursts. |
| Cap the CPU at 200 W | Tom's Hardware tested this on a 265K: about 5% slower. That gives speed away for a problem I had not proved. |
| Turn the fans up | On this board the fan curves live in the BIOS only, because Linux cannot see the fan controller. Not something to fiddle with on a hunch. |

The review numbers are estimates, not measurements on our machine. They come from published tests of the 265K, 285K and 270K Plus by Tom's Hardware, TweakTown, The FPS Review, Tech4Gamers, Club386 and Igor's Lab.

## What worked

**Ask the chip, not the dashboard.** A modern desktop chip boosts until it reaches its heat limit, so hot under load is often just the design working. What matters is how often the heat actually forces it to slow down.

**The kernel already keeps score.** Its own throttle counters add up the time the CPU spends held back by heat. Over 18 days of uptime, ours was held back for 342 seconds in total, about 0.02% of the time. The longest single episode lasted 36 ms. The machine was about 26% busy on average across those days.

**Its twin agrees on the scale.** We run a twin with the same board and chip and different RAM. Over its own uptime it logged 180 seconds held back by heat.

**Then we changed nothing.** By the reviews' estimates, a new cooler would have bought a few percent at best, for something that barely happens.

## Numbers

| Label | Value | Kind |
| --- | --- | --- |
| Dashboard reading that started it | 210 °F (about 99 °C), 85% load | measured |
| Time held back by heat, 18 days of uptime | 342 s (about 0.02%) | measured |
| Longest single throttle episode | 36 ms | measured |
| Average busy over those 18 days | about 26% | measured |
| Time held back by heat on its twin, over its own uptime | 180 s | measured |
| CPU power limit (board default, within Intel's spec) | 250 W | configured |

## When it stops being true

These counters describe how we use the machine today. If it starts running long all-core jobs for hours, the reviews' gain on long runs starts to matter, and I would read the counters again before deciding.

If the throttle time climbs from seconds to something you notice in a day, that is a real heat problem, and a better cooler earns its price.

If a cooler swap ever happens here, I would measure our own machine before and after rather than trust the reviews.

## Agent brief

```agent-brief
Help me decide whether a hot desktop CPU actually needs a better cooler, before I spend money. "(general)" marks advice not from my setup.

First ask me:
- The CPU, the cooler, the case and the power limit it runs at.
- What the machine does: short bursts of work or long all-core runs.
- What reading worried me, and where it came from.

Constraints:
- No hardware purchases, BIOS changes or power caps without my go.
- Read only. Do not change fan curves, power limits or firmware while measuring.
- Treat review numbers as estimates from similar chips, and say so wherever you use them.

Steps:
- Confirm the power limit is within the chip maker's spec.
- Read the kernel's own thermal throttle counters for every core: the total time held back by heat and the longest single episode.
- Compare that time to the uptime and to how busy the machine is on average.
- If there is a twin or similar machine, read its counters too, as a sanity check.
- Only if throttling is a real share of the time, look up published reviews for the cooler upgrade or a power cap and weigh the estimated gain against the cost (general).

Pitfalls:
- A modern desktop chip boosts until it reaches its heat limit, so a high temperature alone is not a fault.
- A dashboard reading is a single moment. The throttle counters cover the whole uptime, so trust them over the dashboard (general).
- A power cap costs speed on its own: about 5% at 200 W on a 265K in one review.
- On some boards Linux cannot see the fan controller, so fan curves live only in the BIOS.

Done when:
- You can tell me how many seconds the chip spent throttled, out of how much uptime, and the longest episode.
- You give me a plain recommendation: change nothing, or the one change worth its cost, with the estimated gain marked as an estimate.
```

## Related

- [My stepdad's 2012 PC is now my fleet's coordinator](https://eng.vaisman.me/entries/old-pc-in-the-fleet.md)
- [CI that shares agent boxes and yields to their owner](https://eng.vaisman.me/entries/ci-that-yields-to-its-owner.md)