Sustained performance is the throughput a system holds at thermal and power equilibrium. For any workload longer than about ten seconds, it is the number that determines how long the job takes.
- Most machines have measurable sustained headroom available for free through airflow, fan curves, power plans, and closing background processes, before any hardware is bought.
- Undervolting raises sustained throughput by reducing the heat produced at a given frequency, which lets the chip hold that frequency for longer; it is not overclocking and usually costs no peak performance.
- Lowering a power limit can increase sustained throughput, because a lower, stable clock held for the whole run beats a high clock that collapses after fifteen seconds.
- SystemCheck measures sustained throughput and decay directly rather than inferring them from a peak, so before-and-after comparison is straightforward, but it cannot read temperatures, only behaviour.
Why the plateau is the number that matters
Every boost algorithm is designed to make the first few seconds look good. There is nothing dishonest about it. Short bursts are genuinely what most interactive computing consists of, and a chip that can respond quickly feels fast. But the moment your work extends past the boost window, the peak becomes irrelevant and the equilibrium clock takes over entirely.
The arithmetic is unforgiving. A ten-minute video export is spent almost entirely in the plateau; the boost window contributes maybe two percent of the total. If your machine peaks 20% higher than a rival but sustains 15% lower, the rival finishes first and the peak difference bought you nothing.
This is the gap that spec sheets and short benchmarks systematically hide, and it is the reason SystemCheck runs a sustained multi-core load for a continuous stretch rather than reporting the best interval it saw.
What limits sustained throughput
Three constraints bind, usually in this order of frequency:
- Thermal: the chip reaches its temperature limit and steps clocks down to stay there. Fixable with cooling and airflow.
- Power: the platform enforces a sustained power budget lower than its short-term budget. Common on laptops and small-form-factor systems, and largely a firmware decision rather than a physical one.
- Contention: other processes take cores, memory bandwidth, or GPU time, so your workload gets less of the machine than the hardware could give. Free to fix, and frequently the real cause.
The free wins, in order
Start with the things that cost nothing, because they are also the things most likely to be wrong on a machine that has been in use for a while.
Airflow first. A case with intake blocked by a filter clogged with dust, or a laptop sitting on a duvet, is losing sustained throughput to a problem that takes five minutes to fix. Look at the whole path: air must get in, cross the heatsink, and get out. Any of those three being obstructed produces the same symptom.
Then the fan curve. Stock curves are tuned to be inaudible at idle and frequently do not ramp hard until the chip is already above 80°C, by which point it has been stepping down for a while. Moving the curve so full RPM arrives by 70-75°C trades noise for sustained clocks, and the trade is usually worth making on a machine you use for long jobs.
Then the power plan. Windows' balanced plan parks cores and caps minimum processor state; on a machine used for sustained work, the high-performance or ultimate plan measurably changes the plateau. Confirm the laptop is on mains power, and check that any vendor performance-mode toggle is set to the performance rather than the quiet or balanced profile.
Then contention. Sync clients, game launchers, browser tabs running background timers, and scheduled anti-malware scans all take cores from a job that wants all of them. This is the least glamorous item on the list and routinely the largest.
Undervolting: the highest-leverage tuning change
A CPU or GPU is shipped with a voltage curve that guarantees stability across every chip in the production distribution, including the worst one. Most individual chips are better than that worst case and will run stably at meaningfully lower voltage for the same frequency.
Because power scales roughly with the square of voltage, a modest voltage reduction produces a disproportionate reduction in heat. Less heat means the chip stays below its temperature limit for longer, which means it holds higher clocks for longer, which means higher sustained throughput. Peak performance is typically unchanged, because you have not touched the frequency curve.
The caveat is that undervolting is per-chip and must be validated. Too aggressive an offset produces instability that shows up as crashes under load rather than as a clean error. Move in small steps, test each step under a real sustained load, and back off at the first sign of instability. SystemCheck's consistency figure is a useful early warning here. An undervolt that is marginally unstable often shows up as degraded consistency before it produces an outright crash.
The counter-intuitive one: lower the power limit
This feels wrong and is often right. Frequency-versus-power is strongly non-linear at the top of the curve: the last few hundred megahertz can cost 30-40% of total package power. A chip pushed to that region heats rapidly, hits its thermal limit, and then throttles down past the point it would have held voltage-stably in the first place.
Capping the sustained power limit produces a lower, flatter curve that the cooling solution can actually hold. You lose a little peak. You frequently gain sustained throughput, and you always gain consistency and acoustics. On thin laptops in particular this can be the difference between a machine that sawtooths between fast and slow and one that simply works at a predictable rate.
The way to evaluate it is with a before-and-after sustained measurement, not with a peak score, which is exactly the measurement a peak-oriented benchmark will tell you got worse.
Measuring the change honestly
Every change in this guide needs to be validated against a baseline taken under identical conditions, and the standard should be strict. Run three times before, three times after, from a comparable thermal starting point each time. A difference under about 3% is noise. Change one variable at a time or you will not know which one worked.
The metric to watch is sustained throughput, not decay. Decay is a diagnostic, not a goal. You can trivially improve decay by making the machine slower at the start, and that is not an improvement. If sustained throughput goes up, the change worked, whatever happened to the decay percentage.
And note again what the measurement is: SystemCheck observes throughput over time in a browser. It cannot read your die temperature, your package power, or your fan RPM, because no browser can. Pair it with HWiNFO or your vendor's utility when you want to see the physical cause alongside the behavioural effect.
Step by step
- 01
Take a baseline you can defend
Close all other applications, connect mains power, and let the machine idle for two minutes. Run the full SystemCheck benchmark three times with a few minutes between runs, and record sustained CPU throughput, decay, consistency, and the Stage 2 and Stage 3 figures each time.
- 02
Clear the airflow path
Inspect intake and exhaust, clean dust filters and heatsink fins, and make sure nothing is blocking the route air takes through the machine. On a laptop, get it off soft surfaces and clear the underside vents. Re-measure.
- 03
Set a fan curve that ramps early
Replace a quiet-biased stock curve with one that reaches high RPM by 70-75°C rather than waiting for 80°C or above. The chip should never be stepping down while the fans still have headroom. Re-measure.
- 04
Set the power plan and performance profile
Switch to a high-performance Windows power plan, confirm mains power on a laptop, and set any vendor performance-mode toggle to its performance profile rather than balanced or quiet. Re-measure.
- 05
Eliminate background contention
Close sync clients, game launchers, and browser windows; pause scheduled scans; and check Task Manager for anything consuming CPU at idle. Re-measure, and watch consistency in particular: contention shows up there before it shows up in throughput.
- 06
Undervolt in small, validated steps
Apply a modest negative voltage offset using your vendor's utility, then run a sustained load and check both stability and SystemCheck's consistency figure. Repeat in small increments, backing off one step from the first sign of instability. Do not skip validation between steps.
- 07
Try a lower sustained power limit
Cap the sustained power limit modestly below stock and re-measure. If sustained throughput rises or holds while consistency improves, keep it. A flatter curve the cooler can hold usually beats a peak it cannot.
- 08
Compare against the baseline and keep only what won
Run three final measurements under the same conditions as the baseline and compare sustained throughput. Keep changes that produced a gain larger than about 3%; revert the rest. Improved decay with lower sustained throughput is a worse machine, not a better one.
Terms used here
Questions
- Is undervolting risky?
- It carries no risk of hardware damage, because you are reducing voltage rather than increasing it. The risk is instability: too aggressive an offset causes crashes or errors under load. Move in small steps, validate each one under a sustained load, and back off one step from the first instability you see.
- Why would lowering a power limit make my machine faster?
- Because frequency versus power is non-linear at the top of the curve, and the last few hundred megahertz can cost a third of total package power. A chip pushed there heats fast, throttles, and settles below where a capped chip would have held steadily. You lose a little peak and gain sustained throughput.
- How much sustained performance can I gain without buying anything?
- It depends entirely on how the machine was set up, but airflow, fan curves, power plans, and background contention together account for the majority of real-world sustained shortfalls. Machines that have been in service for a few years without maintenance are where the largest free gains live.