Thermal throttling is the chip deliberately reducing clock speed to stay inside its temperature limit. It is a protection mechanism, not a fault.
- The measurable symptom is decay: throughput that starts high and falls over a sustained run, then stays low until the load stops.
- SystemCheck runs in a browser and cannot read a temperature sensor. It infers thermal behaviour from throughput decay across a sustained multi-core load; for actual degrees Celsius you need HWiNFO, HWMonitor, or your vendor's software.
- Decay under roughly 5% over a sustained run is normal. Decay above 15% means the cooling solution is the limiting factor, not the silicon.
- The cheapest fixes are almost always case airflow and a fan curve change. Repasting and undervolting come after, not before.
What throttling actually is
Every modern CPU and GPU boosts opportunistically. The advertised boost clock is a ceiling the chip will reach when temperature, current, and power budgets all permit, not a frequency it holds. When the die temperature approaches its limit (commonly around 95-100°C for recent desktop CPUs, and 83-90°C for GPUs depending on vendor), the firmware steps the clock and voltage down until temperature stabilises. That is thermal throttling.
It is worth being precise about this, because throttling is routinely blamed for problems it did not cause. A chip that boosts to 5.4 GHz for two seconds and settles at 4.6 GHz for the next ten minutes is behaving exactly as designed. A chip that settles at 3.1 GHz because the heatsink is choked with dust is not. The difference is not visible in a peak number. It is only visible in the shape of the curve over time.
How SystemCheck detects it, and what it cannot see
State this plainly because it governs how you should read every number below: browsers expose no API for temperature, power draw, fan RPM, VRAM size, or voltage. There is no permission prompt that unlocks it, no flag, no vendor extension. A web page fundamentally cannot read your thermal sensors.
What SystemCheck can do is measure work completed per unit time, continuously, under a load heavy enough to saturate every logical core. Stage 1 spawns one Web Worker per logical processor and runs a sustained floating-point workload, retaining throughput samples for the whole run rather than reporting an average. Stage 2 applies calibrated GPU raster pressure and records frame time and stability. Stage 3 measures path-tracing samples per second and their consistency.
From those series SystemCheck derives a decay percentage: how far sustained throughput at the end of the run sits below the early peak. Decay is a behavioural proxy for thermal and power limiting. It tells you the machine slowed down and by how much. It does not tell you the die hit 97°C, because SystemCheck does not know that and will not claim to.
When you need actual sensor values (per-core temperatures, package power, VRM temperature, fan RPM, hotspot delta), use a native tool. HWiNFO64 is the most complete on Windows; HWMonitor is lighter; vendor software (Ryzen Master, Intel XTU, MSI Afterburner, AMD Adrenalin) exposes the same telemetry with fewer sensors and a friendlier layout. Run one of those alongside a SystemCheck run and you get both halves: measured behaviour and measured temperature.
Reading the decay number
Decay is a percentage, and the useful bands are wider than people expect. Treat these as directional, not as thresholds with hard physical meaning:
- Under 5%: normal. Cooling is comfortably ahead of the load. Most of this is ordinary boost settling, not a thermal problem.
- 5-15%: expected on laptops and small-form-factor builds, and on any air-cooled desktop under a genuinely heavy all-core load. Worth noting, rarely worth acting on.
- 15-30%: the cooling solution is now the limiting factor. You are leaving real sustained performance on the table and the fixes below will pay off.
- Above 30%: something is wrong rather than merely suboptimal. Blocked intake, dried-out thermal interface material, a heatsink that was never seated correctly, or a fan that has stopped.
Separating thermal decay from other causes of decay
Decay is a symptom, and thermal limiting is only one of its causes. Before you buy a cooler, rule out the cheaper explanations.
Power limiting looks almost identical in a throughput graph but responds differently: it appears within seconds rather than building over a minute, and it produces a hard plateau rather than a gradual slope. Laptops on battery are the classic case. Many drop to a fraction of their plugged-in power budget the moment the charger comes out, and no amount of cooling will change that.
Background load produces noisy, irregular decay rather than a smooth curve. A browser update, a cloud-sync client indexing files, or a game launcher updating in the background will all steal cores mid-run. SystemCheck reports consistency alongside throughput for exactly this reason: a low consistency score with a high decay number usually means contention, not heat.
Thermal decay has a signature: gradual onset over 20-60 seconds, a stable lower plateau, and recovery to full speed after the machine has been idle for a minute or two. If you re-run immediately and the second run starts lower than the first started, that is heat soak: the cooler has not shed the previous run's energy yet, which is itself strong evidence for a thermal limit.
The fixes, in the order worth trying them
Work cheapest-and-most-reversible first. The order below is roughly the order of impact-per-effort for a typical desktop or laptop that has been in service for a couple of years.
When throttling is the correct answer
Not every decay number needs fixing. A thin-and-light laptop with a 15W sustained power budget and a 45W boost budget is engineered to throttle. That is the entire design premise, and 25% decay on such a machine is the manufacturer's intent, not a defect. Similarly, a CPU running at its temperature limit while still holding a high sustained clock is being used efficiently; the limit exists to be approached.
The question to ask is not "am I throttling" but "is my sustained throughput acceptable for what I do". A machine that decays 20% but still delivers more sustained work than the one you are comparing it to is the faster machine. This is why SystemCheck grades on sustained behaviour rather than on peak: the peak is a marketing number, and the plateau is the one you actually live in.
Step by step
- 01
Establish a clean baseline
Close every other application, pause cloud-sync clients, and disconnect from any active download. Plug a laptop into mains power. Let the machine idle for two minutes so it starts cold, then run the full SystemCheck benchmark and record the Stage 1 decay and consistency figures.
- 02
Confirm the decay is repeatable
Wait five minutes and run again from cold. If both runs show similar decay, you have a real behavioural signal. If the numbers differ substantially, background load was interfering and the first result should be discarded.
- 03
Get actual temperatures from a native tool
SystemCheck cannot read sensors. Install HWiNFO64, HWMonitor, or your vendor's utility, start logging, and run the benchmark again. Watch package temperature, per-core clocks, and package power together. The moment clocks fall while temperature is pinned at its limit is thermal throttling confirmed.
- 04
Clear the airflow path
Check that intake and exhaust are unobstructed, that dust filters are clean, and that the heatsink fins are not packed with dust. On a laptop, lift it off soft surfaces and clear the underside vents. This single step resolves more throttling complaints than every other fix combined.
- 05
Fix the fan curve
Many stock curves are tuned for quiet rather than for sustained load and only ramp aggressively above 80°C, by which point the chip is already stepping down. Set a curve that reaches high RPM by 70-75°C and re-measure decay.
- 06
Renew the thermal interface material
Paste dries out over roughly three to five years, and a poorly seated cooler can be bad from day one. Repasting a CPU is a 30-minute job on a desktop; on a laptop it is a teardown and should be weighed accordingly. Expect a meaningful improvement only if the machine is old or was assembled carelessly.
- 07
Reduce heat at the source
Undervolting keeps the frequency curve but lowers the voltage needed to reach it, cutting power and therefore heat with little or no loss of throughput. Alternatively cap the sustained power limit outright: a modest cap often costs a few percent of peak and returns far more in sustained output.
- 08
Re-measure and compare against the baseline
Run SystemCheck again under identical conditions and compare sustained throughput, decay, and consistency against your recorded baseline. Sustained throughput going up is the result that matters. A lower decay percentage with lower sustained throughput is a worse machine, not a better one.
Terms used here
Questions
- Can SystemCheck tell me my CPU temperature?
- No. Browsers expose no temperature API, so no browser-based tool can read your thermal sensors. SystemCheck infers thermal behaviour from throughput decay over a sustained load. For actual degrees Celsius, use HWiNFO64, HWMonitor, or your vendor's utility.
- Is thermal throttling damaging my CPU?
- No. Throttling is the protection mechanism that prevents damage. Modern chips are designed to run at their temperature limit indefinitely. Sustained high temperatures accelerate the ageing of the thermal interface material and nearby components, but the chip itself is not being harmed by hitting the limit its firmware enforces.
- What decay percentage is normal?
- Under 5% is normal on a well-cooled desktop. 5-15% is expected on laptops and small-form-factor builds. Above 15% indicates cooling is the limiting factor, and above 30% usually means a specific fault such as blocked airflow or degraded thermal paste.
- My decay is high but the machine feels fast. Should I care?
- Only if your workloads are long. Decay is invisible in short bursts: browsing, launching applications, compiling a small project. It becomes the dominant factor in video export, long compiles, extended gaming sessions, and anything else that keeps the chip loaded for minutes at a time.