GPU Undervolting Guide: How to Lower Temperatures Safely with Voltage Curves, Power & Temperature Limits

2026-10-19 ｜ GPU Repair ｜ 14 min read ｜ HK GPU Fix Editorial

Key Takeaways

Undervolting means running the GPU core at the same (or nearly the same) frequency with less voltage. The goal is efficiency: lower power draw and heat while performance stays similar. It is not a repair, and it only works if each small step is tested. If a setting proves unstable, revert to the last stable setting or the factory default.

What Undervolting Is

Undervolting means supplying less voltage to the GPU core while keeping the core frequency the same, or very close to the same. The goal is efficiency, not higher peak performance. If a given card can run its target frequency stably at a lower voltage, it draws less power and produces less heat at that frequency. Fan noise may also fall, because the cooler has less heat to remove. And because the card hits its power and temperature limits later, its boost frequency may hold steadier during sustained load.

That last point deserves care. Undervolting does not make a card "much faster." The accurate statement is that performance stays broadly similar, and sustained performance may become more consistent because throttling arrives later. Anyone expecting a large speed increase from undervolting is misunderstanding what the technique does.

Undervolting is also not a repair. It adjusts one layer — the relationship between core voltage and core frequency. It does not fix dust in the heatsink, dried thermal paste, a faulty fan, a power supply problem, a driver problem, or a hardware fault. If a card already shows artefacts, black screens, or instability at factory settings, the cause must be identified first. Using undervolting to mask a fault only makes diagnosis harder.

A related point: undervolting is not a matter of copying someone else's numbers. Cards of the same model can need different voltages to stay stable, because of manufacturing variation between individual chips. A voltage that is stable on one card may be unstable on another card of the same model. Every card must be tested in small steps, one change at a time, and any setting that proves unstable must be reverted.

This article covers conservative undervolting: finding a lower voltage that holds the target frequency, and verifying stability properly. It does not cover reverting a card after overclocking caused instability. That topic — returning an overclocked card to stable operation — is covered in a separate article. Overclocking appears here only as a point of comparison.

Why Lower Voltage Means Less Power and Less Heat

The dynamic power relationship

The dynamic power of a CMOS circuit follows a general relationship:

`P ∝ C × V² × f`

In this relationship, C is the equivalent capacitance, determined by the circuit design and treated as roughly constant; V is the voltage; and f is the frequency. The key point is qualitative but important: power scales with the square of voltage, but only linearly with frequency. At a fixed frequency, lowering the voltage therefore has a direct effect on dynamic power — more direct than lowering frequency alone. That is the theoretical basis for undervolting.

This formula must not be turned into a prediction about a whole graphics card. Real power draw also includes static leakage, memory power, power-delivery losses, and load-dependent behaviour. No single formula can be used to calculate how many watts or degrees a given voltage reduction will save on a given card. This guide gives no such conversion, because no general conversion exists.

DVFS and GPU Boost: the card already adjusts itself

A modern GPU does not run at one fixed frequency and one fixed voltage. It uses dynamic voltage and frequency scaling (DVFS) together with the manufacturer's boost algorithm — for example, NVIDIA GPU Boost — to monitor temperature, power draw, and workload in real time, and to decide the frequency and voltage for each moment.

In broad terms, the base clock is the guaranteed minimum operating frequency, and the boost clock is a higher frequency reachable under favourable conditions. The actual frequency floats between the two, depending on whether any limit is being hit. Voltage and frequency are physically linked: switching transistors faster generally needs higher voltage, but voltage also has limits, and excessively high voltage accelerates ageing and carries a risk of damage. The practical consequence is that the card is already balancing frequency against voltage, temperature, and power all the time. Undervolting works within that existing system by changing the voltage the card uses at a given frequency — it does not switch the system off.

Why factory voltage leaves headroom

Chips of the same model differ slightly because of microscopic variation in manufacturing. The stable voltage needed by one chip is not the same as the stable voltage needed by another. A manufacturer that must guarantee that all cards leaving the factory are stable can only do so by using a voltage table that is conservative — somewhat higher — for most cards.

That is why some individual cards can run stably at a lower voltage: the factory table was set to cover the weaker chips as well. A card that tolerates a lower voltage is not "better" in any general sense, and a card that needs the factory voltage is not "faulty." The difference only means that each card's stable voltage must be found by testing that card. Copying another card's numbers, even for the same model, is not a shortcut to stability.

Undervolting vs Downclocking vs Overclocking

These three terms are often confused. They describe different changes with different goals.

Two common confusions follow from this table.

First, lowering the power limit is not the same as undervolting. Lowering the power limit tells the card to stay under a lower total power ceiling, so it hits that ceiling earlier and reduces frequency. The effect may look similar — lower temperature — but the mechanism is different, and the performance cost is usually more direct. Power-limit adjustment and voltage-curve adjustment should be treated as separate changes, not as synonyms.

Second, changing the fan curve, the temperature limit, or the power limit is not undervolting either. A fan curve only changes how cooling responds. A temperature limit only changes the point at which throttling starts. Neither changes the voltage-frequency relationship itself.

Reading the Voltage-Frequency Curve

The voltage-frequency curve, often called the V/F curve, is the map the card uses to relate voltage to frequency. In a curve editor, the horizontal axis is voltage (in millivolts, low on the left and high on the right) and the vertical axis is frequency (in megahertz, low at the bottom and high at the top). Each point on the curve represents one pairing: at a given voltage, the card runs at a given frequency.

The general shape of the curve is worth understanding. In the low-voltage section, the curve is relatively flat: a small increase in voltage buys a substantial increase in frequency. In the high-voltage section, the curve becomes steeper: a large increase in voltage buys only a small increase in frequency. In other words, the voltage cost of each additional step of frequency is typically higher at the top of the range. This is a qualitative trend, not a claim about a specific voltage at which the curve bends.

The factory curve is a conservative setting designed so that most cards are stable. It is not the lowest stable voltage for any particular card. Undervolting means searching along the curve for a lower-voltage point at which this specific card still runs its target frequency stably.

Two qualifications matter. First, even after a target point is set, the GPU still adjusts in real time according to temperature, power, and available voltage, so the frequency may float slightly. A set value should not be described as a locked, unchanging absolute. Second, the curve editor only changes the core voltage-frequency relationship. Memory (VRAM) is a separate matter: core undervolting does not cool the memory. Memory under sustained heavy load — for example, long-running work with high memory occupancy — generates its own heat, and that heat cannot be solved by core undervolting alone. Undervolting must not be presented as a cure-all for the whole card.

Power Limit, Temperature Limit, and Other Limits That Trigger Throttling

The boost algorithm is constrained by several limits at once. Whichever limit is reached first triggers a reduction in frequency. The four main categories are:

1. Temperature limit (temp limit). When the core temperature reaches the set ceiling, the algorithm lowers frequency (and voltage) to prevent overheating.
2. Power limit. When total board power reaches its ceiling, the algorithm lowers frequency or voltage to stay within the ceiling. This is one of the most common causes of throttling.
3. Voltage limit. When the maximum voltage permitted by the VBIOS is reached, frequency cannot be pushed higher by adding voltage, even if temperature and power still have headroom.
4. Reliability / current limits. Firmware intervenes when it detects instability or excessive current. This is less common, and this guide does not elaborate on its internal detail.

Undervolting can make sustained performance more consistent because, at the same frequency, power draw is lower, so the power limit is reached later; lower power draw means less heat, so the temperature limit is also reached later. The two limits still exist. Undervolting does not remove them — it only delays when they are hit.

Telling throttling apart from a hardware fault

Before changing anything, it helps to know which kind of problem is present. A useful diagnostic sequence follows these principles:

1. Look at monitoring data, not impressions. Use a monitoring tool such as MSI Afterburner monitoring, GPU-Z, or HWiNFO64 to watch frequency, voltage, temperature, power draw, and fan speed together. Note what is at its ceiling at the exact moment frequency drops: temperature, power, or neither.
2. Throttling has a characteristic pattern. Frequency falls in steps as temperature or power rises, and recovers when temperature or load falls. The picture itself shows no artefacts, no black screen, and no driver crash.
3. Symptoms that do not look like throttling. Artefacts at idle or low load; black screens or restarts unrelated to temperature; frequent driver crashes; the same symptom appearing at factory default settings. These point toward the driver, power delivery, contact, cooling structure, or the hardware itself — not toward a limit that undervolting can address.
4. Change one thing at a time. If the core, memory, power limit, and fan curve are all changed together, it becomes impossible to tell which change caused a problem. Change one variable, test it, then move on.
5. Establish a baseline. Run the same test at factory defaults first, and record frequency, temperature, power draw, and frame rate. If abnormalities already exist at defaults, they are baseline problems, not effects of undervolting.

Conservative Adjustment: Principles Before Tools

The following principles apply whichever tool is used. They matter more than any specific sequence of clicks.

• Use one tuning tool. Do not run Afterburner, a manufacturer utility, and the driver's built-in tuning at the same time. Competing tools can fight over the same settings.
• Record the original settings first. Take screenshots or write down the factory curve, the factory power and temperature limits, the fan settings, and the driver version. The purpose is to be able to restore everything completely.
• Establish a baseline. At factory defaults, run a familiar game or benchmark and record load frequency, voltage, temperature, power draw, and frame rate. That baseline is the comparison for every later change.
• Make sure cooling is basically sound. Check that case ventilation is reasonable and that the cooler is not choked with dust. Undervolting cannot substitute for cleaning the cooler.
• Do not undervolt a card with an unexplained fault. If hardware failure is suspected, investigate first (see the "When Not to Undervolt" section below).
• Move in small steps. Many guides describe steps on the order of about 10–20 mV at a time as a working habit. That figure is an operating habit, not a guarantee that any particular step is stable. After each step, run a short test. Only if the setting is stable should the next small step be taken. If it is unstable, revert to the previous stable setting.
• Change one variable at a time. Test the core first, with memory at factory settings. Memory instability can be subtle — it may be masked by error correction and appear as performance going down rather than up — which makes diagnosis much harder if memory and core are changed together.
• Do not enable apply-at-startup immediately. Wait until a setting has proven stable across hours of real games or applications before saving a profile and considering automatic application at startup. Applying an unstable setting automatically at boot makes troubleshooting harder.

The NVIDIA Path: MSI Afterburner Curve Editor

MSI Afterburner is the most widely used tool for NVIDIA cards. The following describes the general operation of its curve editor. It is not a per-model recipe, and no voltage or frequency value given here should be copied as a guaranteed-stable setting for any specific card.

Opening the editor and checking voltage control

Open the curve editor with Ctrl + F, or through the curve icon in the main window. The horizontal axis is voltage and the vertical axis is frequency, with each point representing one voltage-frequency pairing.

If voltage control appears greyed out, check Settings → General for the "Unlock voltage control" and "Unlock voltage monitoring" options, then restart the program. If the hardware or tool does not support voltage control, do not force it.

The conservative procedure

1. Pick a target frequency from the baseline. Use the frequency observed under typical load at factory defaults as the target. Do not start by chasing a higher frequency.
2. Find that frequency on the curve and move it down a small step. Locate the voltage point corresponding to the target frequency, and move it down by a small step — on the order of about 10–20 mV as a working habit, not a guaranteed-stable amount.
3. Flatten the curve to the right of the target point. The section to the right covers higher voltages. Flattening it prevents the card from climbing to higher voltage on its own under load.
4. Apply the change. Press Apply. Do not jump in large steps.
5. Run a short test after each change. If the setting is stable, take the next small step. If it is unstable, revert to the previous stable setting.

Two editor functions are worth knowing as tool features, not as required steps. Selecting a point and pressing L locks that voltage-frequency point for isolated testing. Ctrl + D resets the curve offset. These are functions of the tool; they do not imply that every reader must use them.

The standing rules from the previous section apply throughout: do not change memory at the same time, and do not set the profile to apply at startup until the setting has proven stable over hours of real use.

Reverting

Reverting means restoring the factory defaults in the tuning tool, or loading the previously saved factory profile, and confirming that the restored settings have actually been applied. If automatic application at startup was enabled, disable it first.

If problems appear after undervolting, test again at factory defaults. If the problem disappears at defaults, the undervolted setting is the likely cause, and the correct action is to revert. If the problem persists at defaults, it was not caused by undervolting, and investigation should turn to the driver, cooling, power delivery, or hardware.

One rule must be stated plainly: do not respond to instability by pushing further in the same direction to "see what happens." The correct response to instability is to step back.

The AMD Path: AMD Software (Adrenalin)

AMD users do not need Afterburner. AMD Software (Adrenalin) includes tuning under Performance → Tuning, with an automatic Undervolt GPU preset as well as manual GPU Tuning with advanced control.

The manual principle is the same as on NVIDIA: choose a target frequency to hold, lower the voltage in small steps — many guides use steps on the order of about 10 mV as an example — apply each change, run a short test, and revert to the previous stable voltage if the setting proves unstable.

Three AMD-specific notes matter:

• Interface varies by generation and driver version. Newer generations use frequency and voltage sliders with advanced control; older generations (for example, RX 5700, Vega, and RX 500 / 400 series) use a curve graph or performance-state (P-state) interface. Readers should not assume their screen matches any particular screenshot, and should expect the layout to differ by generation and driver version.
• Per-game profiles exist. If a global undervolt is stable in most games but unstable in one particular game, a more conservative profile can be set for that game alone, rather than abandoning the whole setting.
• Power Tuning is separate from voltage tuning. The Power Tuning control adjusts the power ceiling as a percentage. It is not the same control as voltage adjustment, and the two should not be confused — the same distinction made earlier between lowering the power limit and undervolting.

Stability Testing: Three Layers, All Required

Stability testing is the core of this guide. A setting that has not been tested in all three layers below has not been shown to be stable.

Layer 1: Short load after each small change

After every small adjustment, run a short load test of roughly 10–15 minutes — a stress or benchmark test — to quickly screen out settings that are clearly unstable. Passing a short test only means the setting is worth continuing to test. It does not mean the setting is stable for long-term use.

Layer 2: Long load for the final setting

For the final setting, run a longer cycle test of roughly one hour, plus roughly one to two hours of real games or applications. Long testing catches a specific failure mode: heat soak, in which the case and cooler warm up gradually and instability appears only after sustained operation. A setting that passes a short test can still fail a long test.

Layer 3: Different applications

Do not rely on a single test. Different engines, different APIs (DirectX 11, DirectX 12, ray tracing), and different load characteristics (burst versus sustained, core-heavy versus memory-heavy, rapidly rising and falling load) trigger different instabilities. Passing one test does not mean all applications are stable.

Load change itself is a test. Modern GPUs often fail not during steady full load, but at the instant frequency or power draw jumps up or drops sharply. Testing should therefore include scenarios in which the load jumps around, not only one steady full load.

Comparing fairly

Before-and-after comparisons must use the same conditions: the same resolution, quality settings, frame-rate cap, and background programs. Record more than the highest frame rate: temperature, power draw, fan speed, frequency stability, and frame time — whether frame delivery shows stuttering — all matter.

FurMark is not enough on its own

FurMark is an extreme thermal and power stress test whose load characteristics differ from everyday gaming. Some cards automatically reduce frequency under FurMark as a protection measure. FurMark is useful as a thermal and power sanity check, but a stability conclusion should rest on several tests together — 3DMark, Unigine, OCCT, and real games — not on FurMark alone.

Instability Symptoms: Any One Means FAIL, Revert

If any of the following appears, the setting has failed. Revert to the previous stable setting or to factory defaults, and do not continue pushing in the same direction.

• Artefacts, flicker, or texture errors. Specks, colour blocks, lines, or garbled textures on screen. Note that artefacts can also come from memory, the driver, cabling, or cooling, so artefacts alone do not prove undervolting caused them. The correct step is to revert to defaults and test again to see whether the symptom disappears.
• Black screen with the machine still running. The display goes black while fans or sound continue. This is commonly associated with a driver timeout reset (TDR), power fluctuation, or an unstable GPU setting. Windows may also record an event of the "display driver stopped responding" type.
• Driver crash or application crash. A game or benchmark exits to the desktop without warning, or a driver crash message appears.
• Full system restart or freeze requiring a hard restart. This is a more serious instability signal, not a minor issue. Revert to defaults first, then investigate power delivery, temperature, and driver directions.
• Performance going down instead of up. If scores or frame rates fall after adjustment, error correction or throttling protection may already have been triggered. "No artefacts on screen" does not mean "stable."
• Delayed instability. A setting can pass a short test and still fail in a long test or in one particular game. That is precisely why passing a short test does not equal stability.

A test log worth keeping

A simple log keeps the process honest. It does not need example numbers filled in — blank cells are left for the reader to fill, so that no one is misled into copying figures.

When Not to Undervolt

1. Hardware failure is already suspected. For example, artefacts, black screens, frequent driver crashes, or full restarts already occur at factory defaults. Undervolting will not fix hardware, and it can make diagnosis messier. The correct sequence is: revert to defaults, record symptoms, and investigate driver, cooling, power delivery, contact, or hardware directions — or arrange an inspection.
2. The cause of artefacts or black screens is unknown. Do not use "try undervolting and see if it gets better" as a diagnostic method. Undervolting is an efficiency adjustment, not a fault-diagnosis tool.
3. The cooling structure itself has a problem. Heavy dust in the cooler, a fan that does not spin or makes abnormal noise, ageing thermal paste, or a case with no airflow at all. These must be dealt with as cooling problems first. Undervolting cannot compensate for them.
4. Power delivery is unstable or the power supply is inadequate. Undervolting cannot substitute for a suitable power supply and stable power delivery. If power delivery is suspected, investigate power delivery.
5. The original settings were not recorded. If there is no known way back, do not start.
6. Laptops and all-in-one machines need extra caution. Cooling headroom is smaller, and driver and manufacturer controls may differ. Do not apply desktop guidance directly to a laptop.
7. Warranty: check the terms for the specific card. Published guidance disagrees here. Some guides state that pure software undervolting generally does not void warranty; others state that in rare cases it may affect warranty. Because warranty terms differ by manufacturer and model, this guide does not adopt the claim that undervolting "definitely does not affect warranty." The accurate position is: pure software adjustment can generally be reverted to factory settings, but warranty terms vary by manufacturer and model, so check the terms for the specific card before changing anything. What is genuinely high-risk, and very likely to affect warranty, is modifying the VBIOS, flashing BIOS, or modifying hardware voltage — and this guide does not teach any of those.
8. Never treat VBIOS modification, BIOS flashing, or hardware voltage modification as part of undervolting. This guide covers only software-level curve and driver tuning. Any instruction involving firmware or hardware modification is out of scope and must not be followed as an extension of this guide.

FAQ

Is undervolting risk-free?

No. Pushing voltage too low causes instability. And if a fault already exists, undervolting can mask or confuse diagnosis. The correct attitude is "low risk when done carefully, with testing and a known way back" — not "no risk."

Can I just copy someone else's voltage numbers?

No. Stable voltage differs between individual cards of the same model. Figures found online are at most a reference starting point for that author's card under that author's conditions. They are not a guarantee for any other card.

Is undervolting the same as downclocking or overclocking?

No. See the comparison table above. Downclocking lowers frequency. Overclocking raises frequency above factory settings. Undervolting keeps frequency nearly unchanged and lowers voltage.

Is lowering the power limit the same as undervolting?

No. The power limit is a total power ceiling. Lowering it makes the card hit the ceiling earlier and reduce frequency. Undervolting changes the voltage-frequency curve. The mechanisms differ, and so do the performance costs.

Does passing a short test mean the setting is stable?

No. Heat soak, a particular engine, or a sharp load change can cause failure later. Stability needs all three layers: short tests after each step, a long test for the final setting, and different applications.

Does "no artefacts on screen" mean stable?

No. Memory instability can be masked by error correction and appear as performance going down rather than up. Driver crashes, application crashes, and restarts are also instability, even with no artefacts.

Can undervolting fix artefacts or black screens?

No. Undervolting is an efficiency adjustment, not a repair. Symptoms already present at factory defaults need diagnosis first — driver, cooling, power delivery, contact, or hardware — not a voltage change.

Does undervolting cool the whole card?

No. Core undervolting mainly affects the core. Memory generates its own heat independently, and sustained memory-heavy work cannot be solved by core undervolting alone.

Are fan curves, power limits, and temperature limits the same thing as undervolting?

No. A fan curve changes cooling response. A temperature limit changes the throttling trigger point. A power limit changes the power ceiling. None of them changes the voltage-frequency relationship itself.

Can undervolting replace cleaning dust or replacing thermal paste?

No. If the cooling foundation is poor, undervolting helps only to a limited extent. Cooling-structure problems must be fixed as cooling problems.

Does undervolting definitely leave my warranty unaffected?

This guide does not make that claim, because published guidance disagrees and warranty terms differ by manufacturer and model. Pure software adjustment can generally be reverted to factory settings. Check the terms for the specific card before changing anything. Modifying the VBIOS, flashing BIOS, or modifying hardware voltage is genuinely high-risk and very likely to affect warranty, and is not taught in this guide.

Will undervolting definitely make my card last longer?

No quantified lifespan claim is made in this guide, because no verifiable quantitative basis exists for one. Lower temperature and power draw are generally more favourable for long-term operation, stated qualitatively. That is the extent of the claim.

When to Stop Adjusting and Arrange an Inspection

Undervolting is a tuning exercise for a healthy card. If the card shows artefacts, black screens, driver crashes, or restarts at factory defaults — or if symptoms persist after reverting to defaults — stop adjusting and investigate the cause instead. A card with an unexplained fault, a failing fan, or a suspected power-delivery problem needs diagnosis, not a voltage change.

If diagnosis is needed, an inspection can identify whether the problem lies in cooling, power delivery, contact, or the hardware itself. Published repair pricing for this service covers a total range of HK$400–2,800, with the actual quote given after inspection. The workshop is at Shek Mun, Dali Plaza, Room 1209. For enquiries, call 2186 8010 or WhatsApp 6558 6806.