This piece walks through that method and what it can and can't prove. It also flags where the accompanying Python script is weaker than the manual steps. This is a procedure, not a new Windows feature or a hardware release.
Throttling is a decision, not a temperature
A modern CPU has several independent brakes, and each leaves a different fingerprint. The fix for one does nothing for another.
| Limit | What triggers it | Typical fingerprint | Usual fix |
|---|---|---|---|
| Thermal (TCC / PROCHOT) | A sensor reaches the thermal limit | Temperature plateaus at the limit and clocks fall | Cleaning, fresh paste, airflow |
| Intel PL1 / PL2 | Package power passes the long-term or short-term budget | Clocks drop after the boost window while temperature stays below the limit | Firmware limits, or accept the design |
| AMD PPT / TDC / EDC | Package power or current reaches a cap | One limit percentage sits near 100 while clocks sag | Firmware tuning, cooling headroom |
| External PROCHOT | A voltage regulator or platform sensor trips | Clocks fall while CPU temperature looks moderate | Check VRM cooling, the power adapter, OEM thermal mode |
| OS or firmware profile | Power saver, battery policy or an OEM quiet mode | Clocks never reach boost, even when cool | Change the power plan or performance mode |
Two clarifications keep you from chasing ghosts:
- Low idle clocks are not throttling. Compare clocks under load, never at rest.
- Thin and fanless designs are built to boost briefly and then settle. A plateau below the peak isn't automatically a fault. What matters is how far it sits below the burst, whether it holds, and whether it moves after you clean or repaste.
Set the stage
Shattered recommends a controlled setup, because a result only means something if you can repeat it:
- Plug a laptop into its original charger.
- Pick one power mode and leave it alone for the whole test.
- Close browsers, chat apps and sync clients, and pause Windows Update and indexing.
- Reboot and let the machine idle for about five minutes.
- Write down the room temperature.
- Find your CPU's vendor-published maximum temperature. You will pass it to the script as
--tlimit.
For the model name, Windows can use Get-CimInstance Win32_Processor, Linux lscpu, and macOS sysctl -n machdep.cpu.brand_string. Check that the fans spin and the vents are clear before you start. Keep the first run short, and save your work first, because an emergency shutdown gives no warning.
Windows 11: no built-in indicator
Task Manager shows utilization and a speed figure. It doesn't show CPU temperature or limit flags. Shattered also notes that Microsoft's preview notes describe a Task Manager CPU speed display problem on recent Insider builds. Treat that figure as a convenience, not evidence.
The suggested approach has two parts:
- HWiNFO in sensors-only mode. Reset the min/max/average columns, then run an all-core load (Cinebench, Prime95 or y-cruncher) for 10 to 30 minutes. Watch effective clocks, package temperature and power, and any throttling or limit rows. Labels vary by CPU and HWiNFO version. A throttle flag flipping to Yes while effective clocks fall is the pairing that counts as evidence.
- Performance counters.
typeperforGet-Countercan read% Processor PerformanceandProcessor Frequency. Multiply nominal frequency by the percentage and divide by 100 to get the effective clock. If a counter isn't found, a localized Windows install may have translated the names.typeperf -qx "Processor Information"lists the local ones. - The active power plan.
powercfg /getactiveschemematters because a Power saver plan or OEM quiet mode can produce a sagging graph identical to thermal throttling.
On version context: Microsoft's release information lists Windows 11 26H2 as a General Availability Channel release dated 2026-09-29. The same page also lists 26H1. That version is scoped to newer devices and isn't offered as an in-place update. So call 26H2 the latest feature update for existing devices, not the newest in every sense. None of this procedure depends on the version.
Linux: the most direct view
The recommended command is:
sudo turbostat --quiet --summary --interval 2 --show Avg_MHz,Busy%,Bzy_MHz,PkgTmp,PkgWatt
The turbostat man page describes Bzy_MHz as the average clock rate while the CPU was not idle. It describes PkgTmp as the degrees Celsius reported by the per-package thermal monitor. That backs Shattered's advice to watch Bzy_MHz rather than Avg_MHz, which falls whenever the load does.
There is a caveat on the power side. The same documentation says the PKG_% throttling field is model specific, and that some hardware counts only thermal responses. It adds that comparing PkgWatt and PkgTmp to system limits is necessary. A flat power line with falling clocks suggests a power limit. A temperature pinned at the vendor limit suggests heat. Treat both as strong hints, not proof.
You can also snapshot the cumulative kernel counters before and after the run. These are thermal_throttle/core_throttle_count and package_throttle_count under /sys/devices/system/cpu/. The turbostat documentation says events since boot can be found in those thermal_throttle files. Intel systems commonly expose them. An empty result on AMD means the counters are unavailable, not that nothing happened.
If turbostat complains about the msr device, the suggested fix is sudo modprobe msr and running the tool as root. Also compare the cpufreq governor against the performance setting before blaming cooling.
The Linux kernel's own index lists the 7.3 series, with 7.3.5 dated 17 November 2026. Kernel version isn't a prerequisite, though. Driver and sensor support matter more.
macOS: thermal pressure, not degrees
Apple's support page lists macOS 27 Golden Gate at 27.2.1, with Tahoe at 26.7.3 and Sequoia at 15.8.3. The manual steps are:
pmset -g thermfor a snapshot, andpmset -g thermlogfor a stream. A CPU speed limit below 100 means macOS is holding the processor back. A note that no thermal warning level has been recorded is a clean result.sudo powermetrics --samplers cpu_power,thermal -i 5000 -n 12for the thermal pressure level, plus cluster frequencies and package power on Apple silicon.
Apple doesn't document a degree value for pressure levels, so don't try to map them to temperatures. A fanless MacBook Air and a fan-cooled MacBook Pro will behave differently, so compare each Mac only against its own earlier runs.
The Python logger, and where it falls short
Shattered's script uses psutil (tested on Python 3.12.3 and psutil 7.2.2) and starts one busy worker per logical CPU. It samples once a second through an idle baseline and a load phase, and writes a CSV. The default run is 30 seconds idle and 600 seconds under load.
The verdict logic compares the median clock of the first and last stretches of the load. The default threshold is an 8% drop, adjustable with --drop. The labels work like this:
- THERMAL THROTTLING: the clock dropped and either the kernel counter rose or the tail temperature came within 2 °C of your limit.
- POWER-LIMIT THROTTLING (likely): the clock and package watts fell by at least 10% with no thermal evidence.
- UNEXPLAINED CLOCK DROP: the clock fell with no thermal or power evidence.
- HOT BUT HOLDING CLOCKS: temperature is at the ceiling, but clocks held.
- CLEAN: no sustained clock loss was detected.
These rules are the author's heuristics, not a vendor standard, and the probes are uneven across platforms:
- Temperature. It reads
psutil.sensors_temperatures(), which exists only on Linux and FreeBSD. Windows and macOS runs will show no temperature. - Linux clocks. The script uses
psutil.cpu_freq(), not the turbostat Bzy_MHz measure the article recommends for manual work. For power it relies on the RAPL energy counter, which may be root-only. - macOS. The script reads only the
CPU_Speed_Limitfield frompmset. It doesn't collect the thermal pressure level or the powermetrics clock and power data. The article says the script treats thermal pressure as direct evidence, but the code doesn't do that. - Load. The built-in load is lighter than Prime95 or y-cruncher, so a clean result may flip on a heavier workload. Use
--no-loadand run your own job for a more realistic test.
The article's example outputs are explicitly illustrative. Its laptop run (4,210 MHz burst, 3,380 MHz sustained, a 19.7% drop, 96 °C peak) is not a measurement. Its only real run was on a virtual machine with a fixed 2,693.7 MHz and no sensors. That result shows the script handles missing data gracefully. It says nothing about cooling quality.
A repeatable repair loop
The most useful habit in the guide is changing one thing at a time:
- Run the full test and save
before.csv. - Make a single change, such as cleaning the intake, repasting, or switching the power mode.
- Rerun the same test, saved as
after.csv. - Compare the sustained clock and the peak temperature.
If the sustained clock rises and the peak falls, the change worked. If temperature fell but the clock didn't rise, a power limit is now in charge. The guide also suggests a scheduled weekly check, using cron or schtasks. It treats that as a trend alarm, since a five-minute run is shorter than ideal.
Common ways to ruin the result
- Reading temperature alone. A hot CPU holding full clocks is fine. A warm one that has dropped 20% is not.
- Testing for about a minute. Many laptops hold boost briefly, so a short run reports a clean chip. Ten minutes is a sensible floor.
- Testing on battery. Lower power caps there are policy, not a fault.
- Changing two variables. You won't know which one helped.
- Ignoring ambient temperature and background tasks. Both skew clocks and power.
Bottom line
The method is sound. Log effective clocks against temperature and power, run long enough to reach the plateau, and compare like with like. Use the script for repeatability, but lean on HWiNFO, turbostat and powermetrics for evidence about which limit is responsible. Remember that "clean" only means no sustained clock loss was detected by the available probes during that run.
References
- CPU Throttling Test: 12 Steps, 50 Min - shattered.io shattered.io · 2026-11-20T09:24:55+00:00
- Find out which macOS your Mac is using – Apple Support (UK) support.apple.com
- Index of /pub/linux/kernel/v7.x/ kernel.org