The Ultimate NVIDIA GPU Undervolting & Overclocking Guide

NVIDIA GPU overclocking and undervolting helps you lower voltage, reduce heat, and keep strong performance. In this guide, we tune an NVIDIA GPU with MSI Afterburner, benchmark with Unigine Superposition and 3DMark Steel Nomad, and validate stability in real workloads.

The goal is simple: lower voltage while keeping the same performance, or in some cases gain performance, because the GPU can boost more consistently with less heat and lower power draw.

The screenshots and benchmark numbers in this guide are examples from one RTX 5070 Ti system. Your results can be different depending on your GPU model, cooler, BIOS, driver version, ambient temperature, power limit, case airflow, and the game or benchmark you are testing.

What Is GPU Undervolting?

GPU undervolting means asking the graphics card to run a specific core clock at a lower voltage.

On NVIDIA GPUs, this is usually done through the Voltage/Frequency Curve Editor inside MSI Afterburner. Instead of only moving the core clock slider, we choose a target voltage point, raise it to our target frequency, and flatten the rest of the curve to the right.

On modern NVIDIA cards, undervolting results depend on silicon quality, cooling, drivers, the active power limit, the temperature limit, and the workload you are testing. During our testing, the same RTX 5070 Ti behaved differently in Superposition, CS2, MSI Kombustor, and 3DMark Steel Nomad. The same curve can produce slightly different delivered clocks in different workloads.

An unstable CPU/RAM overclock, an unstable power profile, aggressive background software, overlays, unstable drivers, or system instability can mimic an unstable GPU undervolt. These problems can cause crashes, score drops, stuttering, or artifacts that have nothing to do with your GPU voltage/frequency curve. Before tuning your GPU, make sure your system is running from a known-stable baseline.

GPU Tuning Terminology

Required Tools

Download and install the following tools before continuing. Use trusted download sources only.

Initial MSI Afterburner Setup

Open MSI Afterburner.

Click on the settings menu.

On the General tab, enable the following options if they are available on your system:

• Unlock voltage control
• Unlock voltage monitoring

Press Apply.

Go to User Interface and select the skin called MSI Cyborg Afterburner. Press Apply, then press OK.

MSI Afterburner may need to restart. Restart it before continuing.

After MSI Afterburner opens again, check your sliders.

Benchmark Setup

Open Superposition. Click Benchmark, then Performance, and copy the settings shown below.

You may need to change the resolution before testing. In this guide, we use a 1080p monitor, so we test with 1080p settings.

If you have two monitors, place MSI Afterburner on one monitor and Superposition on the other.

Running the Stock Baseline

Start the Superposition benchmark.

During the first scene, look at MSI Afterburner after the clock speed has stabilized. Note down these values during the run:

• Core voltage
• Core clock

After the benchmark finishes, note down the final benchmark score.

Write these values down on your phone or in a spreadsheet, then wait for the benchmark to finish before recording the final score.

The image below shows our sample RTX 5070 Ti scoring around 15118 points at stock settings.

During the run, the GPU was clocked around 2860 MHz at 1055 mV.

For the first undervolt attempt, we will target 1020 mV, roughly 30 mV below the stock voltage we observed. You may want to start higher or lower depending on your GPU, but do not rush the process.

Opening the MSI Afterburner Curve Editor

Click on the monitoring graph area inside MSI Afterburner, then press CTRL + F.

You will now see the Curve Editor.

On the left side, we see frequency. On the bottom, we see voltage.

We are going to find the voltage point we want, raise it to the frequency we want, and flatten the rest of the curve to the right.

Finding the Target Voltage Point

Look at the bottom of the graph and find 1020 mV. Move your mouse upward from 1020 mV until you see the dot for that voltage point.

If you click on the dot and look on the left, you will see the clock speed highlighted.

In our example, the 1020 mV point is around 2827 MHz, which is a bit lower than what we want. So we need to raise that voltage point on the VF curve.

Click the 1020 mV dot and drag it upward until you see your target frequency. In our example, we want around 2860 MHz.

But we will not aim exactly at 2860 MHz. We will aim around 2880 to 2890 MHz, which is roughly a +30 MHz request. We will explain why we do that later.

This is called a single point undervolt.

Anything prior to this point is left alone. We change the points from our selection and onwards.

Make the VF Curve window smaller and find MSI Afterburner.

Press the Apply button.

Open the VF curve editor again and find your voltage point. For us, it is 1020 mV.

Now that you have raised the target point, you need to flatten the rest of the points to the right. This limits higher-frequency boost behavior beyond your chosen VF point. It does not hard-lock the GPU to one exact voltage or frequency in every workload, because GPU Boost can still adjust real operating behavior based on power, temperature, current, BIOS limits, and workload.

Focus on the selected dot. Place your cursor between the next points, hold SHIFT, and drag to the right until the rest of the curve points are selected. It will look like this.

Then click one of the selected dots inside the blue selection.

Drag the selected points down slightly, then let go of the mouse.

Now hold SHIFT and press ENTER two times in a row. The end result will be something like this.

Applying the First Undervolt

Close the VF Curve Editor and press Apply in MSI Afterburner.

On the Core Clock field, you should now see Curve. You have applied your first undervolt.

Retesting in Superposition

Open Superposition and start the test again. Use the exact same benchmark settings as before.

In our first undervolt run, the score increased to 15424 points, roughly 300 points higher than the stock baseline.

Afterburner reports that our card ran around 2865 MHz with roughly 1020 mV during the run. A small voltage difference can come from voltage-step granularity, polling rate, rounding, or GPU Boost behavior, and is usually not meaningful by itself.

Rinse and Repeat

Now repeat the process.

1. Open MSI Afterburner.
2. Click the reset button if you want to start from a clean curve.
3. Set the fans to the same fixed fan speed again for consistent testing.
4. Open the VF Curve Editor.
5. Try another voltage and frequency target.
6. Apply the curve.
7. Run Superposition again with the same settings.
8. Write down the score, voltage, and clock speed.
9. Repeat promising results 2 to 3 times to confirm they are not one-run variation.

For the next test, we tried around 2860 MHz at 1000 mV.

During the test, our GPU is not able to sustain the requested 2860 MHz at 1000 mV.

In this specific run, Superposition cannot maintain what we requested. That may mean the voltage is too low for that target frequency on this card, but it can also mean the card is hitting a power, current, temperature, BIOS, or workload-related limit. Check your limit reasons and test carefully before assuming voltage is the only cause.

To recover, we can add 10 or 20 mV and retest, or lower the target frequency slightly, then check whether the card can maintain the target more consistently.

Running a Kombustor test to see if 1000 mV is enough for 2860 MHz can point us in the same direction, but it is still only one workload. If the card cannot hold the target there either, increasing voltage slightly or lowering the target clock is usually the next step.

The Important Caveat: Requested vs. Delivered Frequency

Even though you have set a target point on the VF curve, your actual clock speed may be slightly lower during a benchmark or game.

This happens because when we edit the VF curve, we request a target. We do not force the GPU to hold that exact number in every workload. The GPU still chooses what it can actually hold based on silicon quality, workload, temperature, power behavior, current limits, BIOS limits, and driver behavior.

To explain it simply: just like a CPU can drop voltage and frequency during a heavy stress test, the same rule can apply to a GPU. A heavy workload may cause the card to run slightly below the requested VF point.

If you want to compensate for clock drop, do it in small steps and test multiple workloads. You can slightly raise the requested frequency, slightly raise voltage, or lower your target clock. The correct answer depends on your GPU and your workload.

Validating in Real Workloads

Here is how a target of around 2860 MHz at 1000 mV looks in CS2.

CS2, Superposition, MSI Kombustor, FurMark, and 3DMark Steel Nomad are different workloads. They can all request different amounts of power from the GPU, and the delivered clock can change from one workload to another.

In this example, we request around 2860 MHz at 1000 mV, but in CS2 we get around 2820 MHz. That does not automatically mean the undervolt is bad, but it does tell us that this workload is not delivering the exact clock we expected.

Compensating for Clock Drop

Be cautious with this approach. If a heavier workload drops your delivered clock, that may mean the requested point is not truly stable at that voltage. Lighter workloads that boost closer to the requested clock can then crash. Raising the request can mask instability rather than fix it. Verify in both heavy and light workloads before keeping the higher request.

In CS2, we now get around 2860 MHz at roughly 0.995 V. This is closer to what we expected.

Running MSI Kombustor also matches our expectations in this test.

When the Score Consistently Drops, Stop Chasing Voltage

This is the point where we stop chasing lower voltage in that direction. The undervolt may be too aggressive, the requested frequency may be too high for that voltage, or the card may be hitting a different limit. The fix is usually to raise voltage slightly, lower the target frequency, or return to the previous stable curve.

Do not forget that each benchmark run has some score variation. We are not looking for tiny differences. We are looking for large score drops, crashes, artifacts, or behavior that clearly shows the GPU is no longer performing correctly.

Validating Your Undervolt

To validate your undervolt, open MSI Kombustor and copy the following settings.

Make sure to change the resolution if you are not using a 1080p monitor.

Then press RUN.

Let the stress test run for at least one hour as an early stability check. Passing one hour of Kombustor does not guarantee that the undervolt is stable in every game or application, but it is a useful first test.

Focus on the bottom-right area where you see ARTIFACTS.

The artifact count should stay at 0/0. If you see the artifact count increase, treat the profile as unstable. First confirm that the same test does not artifact at stock settings. If stock settings are clean and only your undervolt artifacts, raise voltage slightly, lower the target frequency, or return to the previous stable curve.

Why We Use Kombustor Instead of Only FurMark

Some people may ask why we do not rely only on FurMark for testing.

FurMark and Kombustor are different applications, but they share some similarities. FurMark can be a much heavier power and thermal stress test depending on settings, driver behavior, cooling, and power limits. It can push the GPU into power, current, or thermal limits very quickly.

That does not mean FurMark ignores your VF curve. The curve is still there. However, because FurMark can create an extremely heavy load, the GPU may hit another limit first and boost down to a lower operating point than the one you expected from your curve.

Kombustor is useful because it lets us check the undervolt for artifacts and stability without relying only on the heaviest possible stress workload. However, Kombustor is still not enough by itself. You should also test the games and applications you actually use.

Overclocking the GPU

Overclocking the GPU is a bit easier to understand than undervolting, but it still needs careful testing. For a normal overclock, we usually increase the core clock and memory clock offsets, then validate that performance actually improves.

Open MSI Afterburner and click the reset button so you are starting from a clean profile.

For overclock testing, you can raise the Power Limit and Temperature Limit to the maximum your card allows. Raising the Core Voltage limit is optional and can allow higher boost voltage under load, which may increase heat and power draw. Do not treat it as required for every overclock.

Set the fan speed to the same fixed testing speed you used earlier, then start with a small core clock offset.

Recommended starting points:

If your card is known to have a lot of headroom, you can move faster, but beginners should not jump straight to large offsets. Test after each step.

When you are done, press the checkbox and apply your setting. Now open Superposition and start the test.

On our sample card, a +100 MHz core offset gives around 16490 points while clocking around 2970 MHz at roughly 1.055 V.

We will use this as a base and keep going further. This is an example from our card, not a universal beginner starting point.

Setting +150 MHz on the core scores around 16730 points with roughly 3030 MHz on the core and 1.055 V.

Setting +300 MHz on the core scores around 16320 points with roughly 3165 MHz and 1.055 V. This is a regression.

A lower score at a higher clock can mean the overclock is unstable, the GPU is hitting another limit, or the run is affected by throttling or errors. This is our sign to stop pushing the core clock in that direction and test smaller steps.

Going back to a +200 MHz core offset gives us 16635 points, around 300 more points than our regression run. This is a sign that the higher offset was not behaving correctly in this benchmark.

We will stick to +200 MHz on the core for this test session and focus on memory offset and voltage later on.

Setting +400 MHz on the core crashes Superposition during the run.

We found the rough limit of our core clock for this test. For daily use, it is usually better to back off from the maximum value. In this example, we would test values below the crash point and below the score-regression point, then validate them properly.

Memory Clock Overclocking

Memory overclocking can improve performance, especially in memory-sensitive games and benchmarks. However, memory overclocks can be deceptive. A bad memory overclock may not crash immediately. It may artifact, reduce scores, stutter, or behave differently across games.

Recommended memory offset starting points:

Some RTX 40-series and RTX 50-series cards may eventually handle much higher memory offsets, including +1000 MHz or more, but that should be discovered through testing, not assumed from the start.

Here is how our sample RTX 5070 Ti looks with a +1000 MHz memory offset and +200 MHz core offset.

Press Apply and start Superposition.

A +200 MHz core offset and +1000 MHz memory offset gives us around 16620 points.

A +200 MHz core offset and +2000 MHz memory offset gives us around 16834 points in Superposition.

A +200 MHz core offset and +1000 MHz memory offset gives us around 7265 points in 3DMark Steel Nomad.

A +200 MHz core offset and +2000 MHz memory offset gives us around 7365 points in 3DMark Steel Nomad.

A +200 MHz core offset and +3000 MHz memory offset gives us around 7299 points in 3DMark Steel Nomad. This is a regression, so for this workload we will fall back to +2000 MHz instead.

Combining Undervolting and Overclocking

Now we will combine what we learned in this article.

Only continue with this section after you have separately tested your undervolt, core overclock, and memory overclock. Otherwise, you will not know which setting caused a crash, artifact, or score regression.

On our sample card, the strongest settings that behaved well in our earlier Steel Nomad testing were around +200 MHz on the core and +2000 MHz memory offset. These are not universal recommendations. They are example values from one card.

Start Superposition or 3DMark Steel Nomad and observe your clocks and voltages.

In this example, a +200 MHz core offset gives us around 3040 MHz at 1.055 V during Superposition.

For context, with the same settings in an empty Dust2 map in CS2, the GPU can clock around 3080 MHz while running around the map.

This further shows that GPU load affects clock speed. We will not worry about a 30 MHz difference. It is not important by itself.

We will now try to chase around 3160 MHz at 1.045 V, which is a 10 mV reduction from the observed 1.055 V behavior.

Open the VF Curve Editor and look at the bottom of the graph to find your voltage target.

We want around 1.045 V, or 1045 mV. Depending on the exact voltage nodes shown by your version of MSI Afterburner, select the closest available voltage point. In this example, we click the point just before 1050 mV.

When we click on it, the left side shows our target is around 3180 MHz.

Remember what we said earlier: the requested frequency and delivered frequency may not be exactly the same. Because this card may deliver slightly lower clocks under load, we aim for around 3180 MHz on the curve to maintain around 3160 MHz in heavier benchmarks.

We highlight the points to the right by holding SHIFT and dragging after our selected voltage point. Then we click on a point inside the selection and drop the selected points down.

After that, we let go of the keys, hold SHIFT, and press ENTER twice. Then we exit the curve editor and press Apply.

If you have trouble doing this, scroll back to the earlier VF curve section and repeat the instructions slowly.

Now run Superposition or 3DMark Steel Nomad to check for score regression.

Our initial 3DMark Steel Nomad score is around 7505 points, almost the same as before the undervolt.

In this specific run, the GPU cannot always reach 3160 MHz and 1.045 V during the full benchmark load, so it sometimes drops to around 3060 MHz during the run.

On the other hand, if we launch an MSI Kombustor stress test, we can see our target is very close.

While running the Superposition benchmark, we can also see the target being reached during parts of the run.

This further shows the difference between benchmarking applications and workloads.

In a separate Superposition-focused run, we also tested around 3160 MHz on the core, +3000 MHz memory offset, and roughly 1.045 V. That scored around 17264 points.

We are not only looking for a higher score. We are looking for no steep regression in clocks or score, lower or controlled power draw, acceptable temperatures, and stable behavior.

We will now try chasing 1.035 V at the same target clock speed.

Running 3DMark now scores fewer points because our target cannot be maintained consistently, so clock speed and voltage drop.

Superposition also shows regression, although not an extreme one.

It follows the same pattern as 3DMark. During the run, the card cannot sustain around 3160 MHz at 1.035 V, so it drops to around 3120 MHz and 1.030 V.

At this point, the better daily profile is probably the previous voltage target, a slightly lower target clock, or a small reduction in memory overclock. Do not chase the lowest possible voltage if it creates score regression, clock drops, crashes, or artifacts.

The Formula

The basic process is:

1. Find a repeatable stock baseline.
2. Find a stable core clock offset.
3. Find a stable memory offset.
4. Observe the voltage and clock behavior during your benchmark or game.
5. Lower voltage in small steps, such as 10 mV at a time.
6. Benchmark and stress test after each change.
7. Stop lowering voltage when clocks drop, scores regress, artifacts appear, or crashes happen.

For example, if your core target is around 3000 MHz, your memory offset is +2000 MHz, and the card runs around 1.055 V, you can try around 1.045 V and benchmark again. If the score, clocks, temperatures, and stability still look good, continue carefully. If performance regresses or stability gets worse, return to the previous setting.

Validate Your Overclock

Validating your overclock is important. A benchmark pass does not prove the overclock is stable. It only proves that the card survived that specific run.

We will use two tools:

• MSI Kombustor
• OCCT

Start by running MSI Kombustor for at least one hour.

While running Kombustor, focus on the bottom-right artifact area.

The artifact field should remain 0/0 at all times. If you see the number increase, lower your clocks and validate again. Also test stock settings to confirm the artifact is caused by your overclock and not another issue.

If you pass Kombustor, run OCCT as an additional stability test.

During real gameplay, GPU load changes rapidly. These load transitions can create fast power/current changes and quick boost-state changes that a flat 100% stress test may not expose. OCCT 3D Adaptive Switch mode can help expose instability from changing GPU load, but it is still not a perfect replacement for testing your actual games and applications.

Open OCCT and go to the settings tab.

Enable the following options if they are available:

• Stop on error
• Stop on WHEA error

Navigate to Stability Test, select 3D Adaptive, and choose Switch.

Use the following starting settings:

• First intensity: 30%
• Second intensity: 100%
• Interval: 500 ms

You can change the settings later if needed, but these are a good starting point.

Press Start and wait a few seconds for the test to initiate.

Looking at MSI Afterburner, you will see OCCT's adaptive load behavior changing GPU load, voltage, and clock speed.

We suggest one hour of OCCT as a minimum early validation test.

Since VRAM tuning was also a part of the process, we suggest running a 98% stress on your GPU with OCCT's VRAM suite.

You can also run a combined test with the following preset to add some more pressure to the card.

You can set the first intensity percentage based around your game's average GPU usage. For us, we will use 60%.

How to Save a Profile in MSI Afterburner

1. Click on the lock.
2. Click on the save button.
3. Click on the Windows button if you want to autoload the profile on boot.
4. Click on profile slot 1.

Unstable GPU clocks can cause crashes, driver resets, artifacts, flickering, stuttering, hitching, or lower benchmark scores. However, stuttering can also come from CPU, RAM, storage, shader compilation, game engine behavior, or background software, so do not blame every stutter on the GPU overclock.

Ray-traced workloads stress different parts of the GPU than raster benchmarks. A profile that passes Superposition, Kombustor, and OCCT can still crash in path-traced games or AI workloads. If you use these, include at least one as part of your validation.

For CPU-side overclocking and stress testing, read our AM5 Infinity Fabric – FCLK overclocking and stability testing article.