The question was “is my PC running ok?” The first number we read was 92°C at twenty percent load, and it turned out to be the least useful number of the afternoon.

The Situation

A desktop that had been running Windows until the day before. Same hardware, same dusty cooler, same room. On Linux it sat at 92°C with the fan audible and the CPU throttling continuously. Nothing crashed. Nothing logged an error. It was just hot, permanently, at almost no load.

There were five separate causes. Four of them looked like the whole answer at the time, which is why this took a day instead of an hour.

The machine: an i7-13700F, 24 threads, 62GB of RAM, an RTX 5060 Ti, in a compact MSI prebuilt with a 500W PSU. Uptime five hours, zero failed systemd units, 11% disk used, 40GB of RAM free. Nothing was broken in the way you would normally check.

Ninety-Two Degrees at Twenty Percent

The first read looked like a cooling emergency.

load average: 4.24, 4.71, 4.52        (24 threads, so ~18% busy)

Package id 0:  +91.0°C  (high = +80.0°C, crit = +100.0°C)
Core 16:       +92.0°C

Sampling the throttle counters thirty seconds apart confirmed it was not a stale sensor:

package_throttle_count: 85609 -> 85626   (+17 events in 30 seconds)

Eighty-six thousand package throttle events since boot. On a box doing almost nothing. The obvious read was dust, and the box was in fact dusty. But dust raises a steady-state temperature. It does not make a CPU throttle at 18% load, and it cannot produce the sawtooth we were about to find.

It Was a Power Profile

The lever that mattered had nothing to do with hardware.

$ powerprofilesctl get
performance

$ cat /sys/devices/system/cpu/cpufreq/policy0/energy_performance_preference
performance

That same value on all 24 policies. The distro shipped with power-profiles-daemon sitting in the performance profile, which drives EPP to performance, which tells the CPU to hold the highest frequency it can and boost instantly at the first sign of work. Windows’ default Balanced plan does the opposite: it clamps boost hard and lets clocks fall. Same silicon, same dust, completely different power policy. That is the entire explanation for “it was fine on Windows.”

Switching to balanced moved EPP to balance_performance:

profile package temp throttle events
performance 96 °C ~296 / min
balanced 78 °C 0–24 / 20 s

Seventeen degrees and effectively all of the throttling, from one command, with no reboot and no config file. The profile persists in /var/lib/power-profiles-daemon/state.ini, so it survives restarts.

It also made every later measurement meaningful, because the box was no longer artificially hot.

The Two Cores That Were Only Video

With the power profile fixed, one process still stood out. Discord’s renderer was burning two full cores.

Process-level sampling pointed at threads rather than rasterisation:

VideoDecoderQueue   39.6% of a core
VideoDecoderQueue   22.8% of a core
ThreadPoolForeg x50  ~9% each

And the GPU was idle the whole time:

utilization.gpu: 16%   utilization.decoder: 0%   power.draw: 35 W

Discord’s Linux Electron build has no hardware video decode path. The “Hardware Acceleration” toggle in settings controls GPU rasterisation, which draws the UI, and nothing else. Incoming stream and screen-share h264 is decoded in software, in VideoDecoderQueue threads, on the CPU. Flipping that toggle back on changed the total by nothing. Turning the screen share off dropped the renderer from 209% to 8%.

That is not a Linux problem you can flag your way out of. The app is already launching with --use-angle=vulkan, which is the correct backend. The choice was to not watch streams, or to accept two cores of decode.

The Spikes That Weren’t Dust

Removing the stream load left a puzzle. Temperatures still jumped to 95°C for a sample or two, then fell back to the high 60s. The instinct was to blame the cooler again. Sampling at 4Hz settled it.

45 seconds, 4Hz, 180 samples
temp:  min 62  median 65  max 87 °C
samples >= 90°C: 0 of 180
throttle counter: 0 over 45s

max temperature change between samples 0.25s apart: 21 °C

A heatsink cannot move 21 degrees in a quarter of a second. A lump of copper and aluminium has far too much thermal mass. The on-die sensor can, because the die itself is a sliver of silicon with almost no heat capacity. Those spikes are boost transients: a couple of cores pull 200W for 300ms, the die heats, the sensor reports it, the burst ends and it collapses. Dust raises the floor. It cannot manufacture a spike, and the earlier “it’s the dust” diagnosis was wrong.

The number that matters is the sustained one, not the peak of a 250ms burst.

The Plugin Eating a Core

One process was genuinely misbehaving:

pid 1714737  103% CPU  running 2.5 minutes
wakatime-cli-linux-amd64 --entity ~/assets/BLENDS/MikuTakeshita.blend

A coding-time tracker, normally taking milliseconds, stuck spinning a full core on one .blend file. Killing it dropped the package from the high 90s to 67–74°C with Blender, both world clients and the chat app still open, and the throttle rate from 125/sec to 47/min.

Worth being precise: the plugin is legitimate and is supposed to run while Blender is open. It was stalled, not malicious. A watchdog that kills it only after 60 seconds of accumulated CPU would catch the pathology without touching normal use.

The Fans Were Already at 100%

This is where the hardware story finally held up, and it needed the fan controller to prove it. The board’s Nuvoton chip exposes tachometers and PWM duty:

fan1:  1444 RPM @ pwm 140   (idle)
fan3:  2479 RPM @ pwm 140

3 P-cores pinned:
  pwm1 / pwm3:  245 -> 255        both fans to 100%
  fan3:         3488 -> 3519 RPM
  package:      89 -> 100 °C      and stayed there

The fan curve works perfectly. It ramps from 55% at idle to 100% under load. And with both fans flat out, three loaded cores still parked the package on Tjmax.

Then the spec sheet explained it:

MSI MAG Infinite S3 13th
  CPU Cooler:  Air cooling 65W
  Chassis:     173.8 x 446.97 x 407.17 mm  (32L)

A cooler rated for 65 watts, on a chip with a 219W turbo envelope. It was never going to hold three cores. Every software lever in this post was real and worth doing, and none of them were the hardware problem. The hardware problem was that the cooler was mismatched from the factory.

What Didn’t Work

Worth recording, because each one sounded reasonable:

Idea Result
Run the world in a browser instead of Electron Actually 35% cheaper (165–180% vs 250–270% of a core), but a different world and a newer Chromium, so not a clean comparison
Pin the heavy process to E-cores 2.2x worse: 605% vs 272%. Fixed-timestep catch-up — slower frames, more work, further behind
Hide the window to throttle rendering No effect: 236% vs 256%. The app’s render loop ignores it
power-saver profile to stay cool 65°C, but clocks collapsed to 1.1–2.2 GHz. Machine crawls
Undervolt Not available. H610 chipsets do not expose a CPU voltage offset
Toggle Discord’s hardware acceleration Nothing. It never touched video decode

The Takeaway

Three things generalise to any machine with this symptom.

Check the power profile before you clean anything. A CPU in a performance EPP profile will run 15–20°C hotter at idle-ish load than the same chip under Windows Balanced, with no hardware fault at all. powerprofilesctl get is one command and it is usually the answer.

Distinguish spikes from sustained temperature. A 20°C swing inside 250ms is a die-level boost transient and is normal. A high temperature that holds for ten seconds is a cooling problem. Sampling at 1Hz cannot tell those apart, and single-point readings will send you chasing dust that was never the cause.

Prove the fan curve before blaming the cooler. Read the PWM duty and the tachometer under load. If duty is at 255 and the CPU still sits at Tjmax on three cores, the fan is fine and the heatsink is undersized or badly mounted. Those are different repairs.

What To Run

# power profile (the 17°C win)
powerprofilesctl get
powerprofilesctl set balanced

# is it actually throttling, or is it a burst?
a=$(cat /sys/devices/system/cpu/cpu0/thermal_throttle/package_throttle_count)
sleep 30
b=$(cat /sys/devices/system/cpu/cpu0/thermal_throttle/package_throttle_count)
echo "throttle events in 30s: $((b-a))"

# prove the fan curve
sudo modprobe nct6687
sensors | grep -E 'fan|pwm'

The replacement is a 148mm single-tower with six heatpipes, which is the tallest that fits the case’s roughly 152mm of internal clearance. Total cost of fixing a factory power mismatch: £16.90.