Why undervolting works
Undervolting exploits the physics of CMOS chips and the conservative factory settings of GPU Boost. A short grounding in both makes the method easier to follow.
The physics in one equation
The dynamic power a digital chip uses is approximately proportional to capacitance multiplied by frequency multiplied by voltage squared. Because voltage is squared, a small reduction in voltage reduces power by a larger proportion. If voltage falls by 10 percent at the same frequency, power from this term falls to 0.9 squared, which is 0.81, a 19 percent reduction. Static power adds a further voltage-dependent term. These are approximations, but they explain why a modest voltage cut can lower power and heat a lot.
Why frequency does not fall with it
A chip needs a certain voltage to switch reliably at a given frequency. Manufacturers add margin to cover the worst silicon, the hottest temperatures and ageing. Many individual cards can run the same frequency at a lower voltage, so reducing the voltage point removes unused margin.
Why the factory curve is conservative
NVIDIA sets one voltage-frequency curve per product, with the same margins for every chip. Your sample may be better than average, so a custom curve can exploit that. A weaker sample may gain less, and that is normal.
What GPU Boost does with the curve
GPU Boost picks an operating point on the curve based on power, temperature and load. When a card is power limited, it runs at a lower frequency and voltage. An undervolt keeps frequency high at lower voltage, so the card draws less power at the same speed, or sustains a higher speed within the same power limit. The guide to RTX 5070 Ti overclocking explains the limiters in more detail.
What to expect from an RTX 5070 Ti undervolt
Set realistic goals.
Typical outcomes
Owners of recent NVIDIA cards often report meaningful reductions in board power, temperature and fan noise with a small loss in performance. The size of the saving varies by sample, game and chosen target. Your RTX 5070 Ti has a 300 W rating, so a drop of tens of watts is plausible, and the benefit is largest in power-limited or heat-limited situations.
What it will not do
It will not make the card faster by magic, and it will not rescue a card with poor cooling. It reduces power and heat, which may let clocks hold steadier, and which makes your computer quieter and cooler.
Who benefits most
Owners of small cases, quiet builds, hot rooms and compact coolers benefit most. The site's guide to light undervolting of the RTX 5090 shows the same ideas on a higher-power card.
Tools and preparation
You need a few free tools.
MSI Afterburner
Afterburner's curve editor lets you reshape the voltage-frequency curve. Download it only from the official site, and open the curve with its keyboard shortcut, usually Ctrl plus F.
A monitoring tool
HWiNFO or the Afterburner overlay shows frequency, voltage, board power, temperature and fan speed. Logging while you test makes comparisons reliable.
A repeatable test
Use a game scene or benchmark you can repeat, and record average frame rate, 1% lows, board power, core temperature, hotspot and fan speed at stock settings. Add 3DMark Steel Nomad stress testing for stability. The article on benchmarking tools and methods explains a fair procedure.
A wall power meter
A plug-in meter measures total system power, which includes the processor, and it provides a check on the card's own power readings.
Step 1: baseline measurements
Without a baseline, you cannot show a gain.
Record stock behaviour
Run your test scene for several minutes at default settings. Note the average frame rate, the 1% low, the board power in watts, the sustained core clock, the maximum voltage reached, the core temperature, the hotspot temperature and the fan speed.
Compute frames per watt
Divide the average frame rate by the board power. For example, 100 frames per second at 290 W gives about 0.345 frames per watt. This number is your efficiency baseline.
Note the ambient temperature
Room temperature affects results. Record it, and compare tests made in similar conditions.
Step 2: choose a target point
The target defines the undervolt.
Choose a frequency near stock boost
Look at your baseline log and note the sustained clock under load. A good first target is a frequency at or slightly below that sustained clock. This keeps performance close to stock while allowing a lower voltage.
Choose a candidate voltage
Open the curve editor and look at the curve. Voltage runs along the horizontal axis and frequency along the vertical axis. Pick a voltage point that is lower than the stock maximum, for example somewhere in the range of 900 to 950 millivolts for a first attempt. The exact range depends on your card and firmware, and you should read your own curve, not copy numbers from others.
Why start conservative
A higher voltage point is safer, and a lower point saves more power. Start with a conservative point, prove it stable, then try a lower one.
Understand the trade-off
For each voltage point there is a maximum stable frequency. Choosing a lower voltage reduces that maximum. The search is for the lowest voltage that sustains your target frequency stably.
Step 3: shape the curve
The curve editor method is reliable, with a few details.
Lift the chosen point
Click the chosen voltage point and drag it up to the target frequency. Afterburner moves the whole curve with it, so the other points shift too.
Lock the point
Select the chosen point and use the lock command, which many versions bind to the L key, so that the card treats that voltage as the maximum. Apply the settings.
Flatten the right side
Check that the curve to the right of the chosen point does not rise above your target frequency. If it does, drag those points down so they sit at or below the target. A flat curve to the right means the card never selects a higher voltage, so it never exceeds your cap.
Apply and verify
Apply the profile, then watch the overlay under load. The voltage should not exceed your chosen point, and the clock should hover near your target.
Step 4: stability testing at load and idle
Undervolts often fail in surprising places.
Load tests
Run the Steel Nomad stress test and verify the frame rate stability stays at or above 97 percent. Play demanding games for an hour or more, and include ray tracing and DLSS workloads, because they use different units. The guides to DLSS 4 and ray tracing versus path tracing describe those workloads.
Light-load tests
Test with video playback, browsing and light games, because the card runs at different clocks and voltages there. An undervolt that is stable in heavy games can crash during a video, because the voltage-frequency relationship at lower clocks may differ.
Sleep and wake
Also test sleep and wake cycles, and multiple-monitor setups, where idle states vary.
Read the logs
After testing, open Event Viewer and check for display driver errors and WHEA warnings. A clean log supports your conclusion.
Step 5: measure the gain
Numbers prove the result.
Re-run the baseline test
Run the same scene with the undervolt active. Record average frame rate, 1% low, board power, temperature, hotspot and fan speed.
Compute the changes
Power change percent equals the new power minus the old power, divided by the old power. Performance retention equals the new frame rate divided by the old frame rate. Frames per watt equals frame rate divided by power.
An illustrative example
Suppose stock gives 100 frames per second at 290 W, and the undervolt gives 97 frames per second at 235 W. Power fell by about 19 percent, performance retention is 97 percent and frames per watt rose from about 0.345 to about 0.413, a gain of about 20 percent. These numbers are illustrative, not measurements of any specific card.
Check temperature and noise
Compare the core and hotspot temperatures and the fan speed. A lower fan speed at the same temperature means quieter operation. Record your impressions of noise, because that is often the main practical benefit.
Working down: find the lowest stable voltage
Once your first undervolt works, you can refine it.
Lower the voltage in small steps
Move the chosen voltage point down by one step, such as 10 or 25 millivolts, keep the target frequency, apply and retest. Continue until you see instability, then return to the last stable setting plus a margin.
Lower the frequency instead
If the voltage cannot go lower at your target frequency, lower the frequency slightly. A small frequency reduction can allow a much lower voltage, and it may give a better efficiency result.
The efficiency sweet spot
Efficiency usually peaks at a point below the maximum clock. Plot frames per watt against your tested settings to find it.
Reading the curve editor like an engineer
The curve editor can look cryptic, and a clear mental model removes the mystery.
Axes and meaning
The horizontal axis is voltage, usually in millivolts, and the vertical axis is frequency in megahertz. Each point says: at this voltage, the card may run up to this frequency. The card moves along the curve to the right when it has power and thermal budget, and to the left when it must save power.
The curve is monotonic
The curve rises from left to right, because higher frequencies need higher voltages. When you lift one point, the editor lifts the points beside it to keep the curve rising, which is why you must flatten the right side after you lock your point.
The first and last points
The leftmost points govern idle and light loads, and the rightmost points govern the heaviest loads. An undervolt that only changes the right side leaves light-load behaviour alone, which is usually what you want.
Why flattening works
When you lock a voltage point and set every point to its right to the same frequency as your target, the card gains nothing by moving right, because the frequency does not rise. It therefore stays at or near your chosen voltage. Voltage readings in the overlay should confirm this.
Why locking works
The lock tells the boost algorithm to treat the selected point as the highest it may use. Combined with the flat curve, it caps both voltage and frequency.
Interactions with power limit, temperature and boost
An undervolt interacts with the other limiters.
Power limit
Because the card draws less power at your chosen point, it hits the power limit less often. If your card was power limited, the sustained clock may rise toward your target. If it was not, the main change is lower power, not higher clocks.
Temperature
Lower power means lower heat, so the card stays further from its temperature target and the fans run slower. The thermal limiter rarely intervenes.
Boost bins
GPU Boost adjusts frequency in bins of about 15 MHz. Your target frequency may land between bins, and the card will round. Choose targets in multiples of the bin size to avoid surprises.
Reliability limiter
The firmware also restricts the highest frequency allowed at each voltage. If your target exceeds that restriction, the card ignores the excess and runs lower than requested. If the overlay shows a lower clock than you set, reduce the target.
Measuring power correctly
Good power numbers need care.
Board power readings
The card reports estimated board power through its sensors. These readings are consistent for comparisons on the same card, and they differ a little from wall measurements.
Wall power
A wall meter measures the whole computer. Subtracting the idle reading gives an approximate GPU contribution, though the processor and fans also vary. Use wall readings to confirm trends.
Averages and peaks
Record the average power over a minute of steady load, and the peak. A lower average with similar peaks may still need a good supply for transients. The guide to high-wattage supplies discusses transient behaviour.
Reporting results
Report the test scene, the settings, the room temperature and the driver version with every result, so that you or another reader can reproduce it.
Why two cards of the same model undervolt differently
Silicon variation is the heart of the matter.
Manufacturing variation
Chips from the same wafer differ slightly in how much voltage they need for a given frequency. NVIDIA tests and bins chips into products, and within a product some samples are better than others. A strong sample can run the same clock at a noticeably lower voltage.
Partner board differences
Power delivery quality, cooler contact and thermal pads also vary by partner design. A card with a better cooler stays cooler, which lowers leakage current and allows lower voltages at the same frequency.
Temperature dependence
A chip needs more voltage when it is hot. An undervolt that is stable in a cool case may need a small margin for a hot summer day, which is why a conservative setting is wiser than the edge of stability.
What this means for you
Treat the numbers other people post as a hint, not a target. Work down from a safe point, and stop when the card tells you to.
Undervolting across the RTX 50 family
The same method applies to other cards, with different numbers.
Higher-tier cards
The RTX 5080 and RTX 5090 draw more power, so an undervolt can save more watts. The guides to RTX 5080 undervolting and light undervolting of the RTX 5090 show examples.
Lower-tier cards
The RTX 5060 and RTX 5050 draw less power, so the savings in watts are smaller, though the percentage can be similar. Their small coolers already run cool, so the main benefit is quiet. The guide to RTX 5060 overclocking and undervolting covers the entry tier.
Laptops
Laptop chips often lock voltage control in the firmware, and the vendor's software may offer limited controls. Check the laptop maker's guidance, and be careful with heat.
Documentation and rollback
A tidy record makes tuning safe.
Keep a log
Record each setting you try, with the test results and the date. A simple text file works well. When something goes wrong, the log shows the last good state.
Save profiles
Afterburner can store profiles. Save your stock profile, your first undervolt and each refinement under clear names.
Know how to reset
Use the reset button in Afterburner to restore defaults. If the system will not boot normally because of an unstable startup profile, boot into safe mode and remove the profile.
Do not apply at startup until stable
Enable apply-at-startup only after a week of stable use. This avoids boot loops from an unstable setting.
Frequently asked practical questions
Does undervolting void my warranty?
Manufacturers generally allow software tuning within the limits their tools enforce, and they exclude hardware modification and physical damage. Check your partner's terms, and keep a stock profile in case a service request needs it. Check your partner brand's terms before you start, and keep a stock profile ready.
Will the undervolt survive a driver update?
Afterburner stores the profile on disk, so it persists, but a driver install may reset the applied state until you reload the profile. Reapply and retest briefly after driver updates. Reapply the profile after each update, confirm the voltage and clock in your overlay, and rerun a short stress test before gaming.
Can I undervolt in the NVIDIA App?
The NVIDIA App has offered performance tuning tools on supported cards, which are simpler and more conservative than a hand-built curve. Check the NVIDIA App guide for what your version offers. Features vary by version, so check what your current App offers, and use a curve editor if you need precise control over voltage points.
Does an undervolt help with coil whine?
Sometimes it reduces coil whine by lowering current, and sometimes it changes the pitch. The effect is unpredictable, so treat it as a possible side benefit and not a guaranteed fix. Results vary, so test the profile and listen at the same fan speed, and remember that a frame cap often matters more for whine.
Undervolting and DLSS workloads
Your test set should reflect real use.
Tensor and RT loads
DLSS, Ray Reconstruction and frame generation load Tensor cores, and ray tracing loads RT cores. These units draw power differently from shader-only loads, so an undervolt stable in a rasterised test should still be checked in a ray-traced, upscaled game. The DLSS 4 overview lists the features to test.
Efficiency gains from upscaling
Upscaling lowers the work per frame, so it already improves efficiency. Combine it with an undervolt for the quietest, coolest result, then cap the frame rate to your monitor's refresh rate.
A short troubleshooting table
| Symptom | Likely cause | What to try |
|---|---|---|
| Crash at the start of a game | Voltage point too low for the frequency | Raise the voltage point or lower the target frequency |
| Artefacts under load | Frequency too high at this voltage | Lower the target by one or two bins |
| Card still boosts past your voltage | Curve to the right not flattened | Flatten and reapply |
| Clock lower than your target | Reliability limiter at this voltage | Lower the target frequency |
| Crash in video playback | Unstable low-clock states | Raise the voltage slightly, or lower low-clock points |
| Crash after sleep | Profile reapplied incorrectly | Reapply the profile after resume, retest |
| No change in power | Card was not power limited | Check that the lock and flat curve applied |
How to use the table
Make one change at a time and retest. The table lists the fastest fixes, and your log shows which one worked.
Efficiency in your electricity bill
Power savings translate into running costs.
The formula
Energy in kilowatt-hours equals power in kilowatts multiplied by hours of use. Cost equals energy multiplied by your price per kilowatt-hour. For example, a saving of 50 W over 1,000 hours of gaming equals 50 kilowatt-hours. At a hypothetical rate of 30 pence per kilowatt-hour, that is £15. Substitute your own rate from your bill.
Heat in the room
Every watt the card saves is a watt not heating your room, which matters in summer. A 50 W saving is similar to turning off a small bulb, and it adds up in a small, warm room.
Noise
Lower power allows slower fans. Many owners value the quiet more than the saving.
Stability tests in detail
Stability testing deserves a methodical approach.
Stress test settings
In the Steel Nomad stress test, run the full set of loops. A pass requires at least 97 percent frame rate stability. A lower value suggests throttling or instability, and you should investigate temperature and power.
Game tests
Use games with different load types. A ray-traced game stresses RT cores, a DLSS and frame generation test stresses Tensor cores, and an esports title stresses clocks at high frame rates. Each exposes a different weakness.
Long sessions
Play for an hour or more. Some instabilities appear only when the card and case are fully heat-soaked.
Warm room test
Test again when the room is warm. A margin that works in a cool room may fail in a hot one.
Simpler alternatives
Not everyone wants curve editing.
The power limit slider
Reducing the power limit to about 90 percent forces the card to run lower clocks and voltages. It is simple and safe, but it is less precise than a curve, and it often costs more performance for the same power saving.
Frame caps
A frame cap reduces work in lighter games. Cap at your refresh rate minus three frames per second with variable refresh rate on, as the article on VRR, G-Sync and V-Sync explains.
Automatic tuning tools
Some tools offer automatic scanning and conservative profiles. Check the NVIDIA App guide for what your version provides.
Cooling and noise after an undervolt
An undervolt changes the thermal picture.
Fan curve adjustment
A card that draws less power runs cooler at the same fan speed, so you can lower the fan curve for quiet. The guide to fan curves shows how to build one.
Case airflow
Lower heat output eases case airflow demands, and it may make a small case workable. Keep good intake and exhaust anyway.
Zero-RPM mode
Cards that stop their fans at low temperatures may stay silent longer after an undervolt, which is a pleasant side effect.
Common mistakes
Lowering voltage without capping the curve
If the curve to the right still rises, the card may boost past your chosen point. Flatten it.
Testing only in games
Check idle and video, because instability appears there too.
Copying other people's numbers
Your card is different. Use others' numbers only as a hint.
Skipping the margin
A setting that passes today may fail in summer. Leave a margin of one or two steps.
Forgetting the baseline
Without a baseline, you cannot show that the undervolt helped. Record it.
Further reading
If you want to see how the same method scales to a flagship, read the light undervolting guide for the RTX 5090 and the hardcore overclock and undervolt walkthrough. Both explain how voltage, clock and temperature interact on a much larger power budget, and the logic carries straight over to a 300 W card. Compare their stability tests with yours, note where your sample differs, and keep a written record of every profile so that a driver reinstall never costs you an evening of careful work.
Summary
Undervolting the RTX 5070 Ti exploits the fact that power rises with the square of voltage and that factory curves carry safety margin. Reshape the voltage-frequency curve in Afterburner by lifting a chosen voltage point to a target clock and flattening the curve to its right, then test for stability at load and idle and measure the result with frames per watt. Expect lower power, heat and noise with a small performance cost that varies by sample. Keep a margin, keep notes and, if you want more speed, see the guide to RTX 5070 Ti overclocking and the complete RTX 5070 Ti guide.
Frequently asked questions about undervolting the RTX 5070 Ti
What is undervolting?
Undervolting means running the GPU at a lower voltage for a given frequency than the factory curve allows. Because power rises with the square of voltage, a small cut reduces power and heat disproportionately, while performance changes little if the chip is still stable. On NVIDIA cards it is done by reshaping the voltage-frequency curve.
Will undervolting reduce my RTX 5070 Ti performance?
It can, slightly, if you choose a target frequency below stock boost, but many owners choose a target close to stock and lose only a small amount. Some see no loss because the card was power or heat limited before. Measure frame rate and frames per watt before and after to judge your own result.
Is undervolting safe for my GPU?
Yes, it is considered safe, because you lower the voltage rather than raise it. The risk is instability, which shows as crashes or artefacts and is resolved by raising the voltage point or lowering the frequency. It does not stress the silicon more than stock settings do.
How do I undervolt an RTX 5070 Ti in MSI Afterburner?
Open the curve editor, choose a voltage point such as the one near 900 to 950 millivolts, raise it to your target frequency, lock it, and flatten the curve to its right so the card does not boost beyond that voltage. Apply, test for stability and measure power. The exact point depends on your card.
How much power does an undervolt save?
It varies by card and target, and owners commonly report savings in the tens of watts on high-power cards. Measure your own card by logging board power at stock and at the undervolt in the same scene, then compute the percentage difference.
Can I undervolt and overclock at the same time?
Yes, with care. Undervolt first and verify stability, then add a modest memory offset and retest. A large core offset on top of an undervolt complicates the curve and makes diagnosis harder. Measure frames per watt to confirm the combination is worthwhile.