Start here
Do the free things first
Most guides open with the fiddly part. That’s backwards. These four moves are in order of payoff against effort, and if you stop after the first two you’ll already have most of the quiet.
Flip the graphics card to its Quiet BIOS
There’s a tiny physical switch on the card. One position is a calmer fan curve. Power target and clocks are identical, so you give up essentially nothing.
Cap the processor’s power, not its temperature
Set a watt ceiling and let the temperature settle wherever it lands. Capping temperature instead fights the boost algorithm rather than reducing the heat it has to shift, and on these chips that causes stutter.
Cap power on the graphics card
One number. No curves, no per-core anything. This is where most of the heat actually leaves the case.
Undervolt the voltage curve
The finicky one, and the only one that can make the machine unstable. Worth doing, but only if you’ll actually run the tests in the last section.
Undervolting means asking a chip to do the same work on less electricity. Less electricity in means less heat out, and less heat means the fans have less to do. You’re not making anything weaker; you’re removing the safety margin the factory left for the worst chip off the line. That margin is why this works, and it’s also why results differ between two identical parts.
The one diagram worth understanding
What “flattening the curve” means
Your graphics card holds a lookup table: at this voltage, run at that speed. Left alone it keeps climbing the table, spending a lot of extra voltage at the top for very little extra speed. Flattening it puts a ceiling on that.
Voltage against clock speed, before and after
Stock behaviour keeps climbing to the right, where each extra millivolt buys almost nothing. Pick a point partway along, then hold every point past it at the same speed. The card now refuses to go past that voltage, because there’s no longer any speed to gain by doing it.
The actual numbers
What to set
Conservative on purpose. These sit inside the normal spread between samples rather than at the edge of it, so they should hold on an average chip instead of only a lucky one.
ASUS RTX 4090
Graphics · TUF / Strix| Physical BIOS switch | Quiet (Q)Do this before anything else |
|---|---|
| Power limit | 70%Not 80%. See the note below |
| Voltage curve point | 0.900 V → 2,600 MHzThen flatten everything to the right |
| Memory clock offset | +0Leave it alone entirely |
| Fan ceiling | 60–65% |
| If it turns out unstable | 0.925 V → 2,600 MHzOr 0.900 V → 2,535 MHz |
| Expected draw in games | 300–330 WFrom about 340 W stock |
ASUS RTX 5090
Graphics · Astral / TUF| Before you touch anything | Run ASUS’s BIOS update toolIt retunes quiet mode; the launch version was criticised for not being quiet |
|---|---|
| Physical BIOS switch | Quiet (Q) |
| Voltage curve point | 0.900 V → 2,700 MHzThen flatten everything to the right |
| Power limit | 450 W (78%)As a backstop behind the curve |
| Memory clock offset | +0Actively avoid on GDDR7 |
| Astral only | Enable Power Detector+Per-pin current monitoring. Genuinely useful |
| Fan ceiling | 60–65% |
| Lowest possible power limit | 400 WFirmware floor. The slider stops there |
Ryzen 7 7800X3D
Processor · stock PPT 162 W, Tjmax 89 °CRyzen 9 9950X3D
Processor · stock PPT 200 W, Tjmax 95 °CThis is the most repeated mistake in 4090 guides. The card is rated at 450 W, but in actual games it already only draws about 340 W. So an 80% limit is 360 W, which is above what the card was drawing anyway, and it does nothing at all. You have to get down near 70% before the cap even starts to bite.
ASUS published an Astral undervolt at 950 mV and 2,902 MHz. That’s a hand-picked chip in a binned overclocking model, so treat their result as the target (102 W and 10.4 °C off) rather than their frequency as the setting. Aim at the outcome, not the number.
Turning the power limit down does not meaningfully protect the 12V-2x6 connector. The failures come from current spread unevenly across the individual pins, not from total wattage. A 5080 melted its connector at 360 W, and a 5090 melted at 100 W below its limit. Anyone telling you to undervolt for connector safety has the mechanism wrong.
What actually helps: seat the plug fully until it clicks, don’t bend the cable tightly right at the connector, and on the Astral turn on Power Detector+, which watches all six 12 V pins separately. That per-pin warning is the only real mitigation available to you in software.
PPT is just a watt ceiling for the processor. Stock is 162 W. Since the 7800X3D only reaches about 84 W flat out anyway, a 120 W ceiling costs you nothing in normal use and still clips the brief spikes that make the fans surge. Going much below 90 W starts to bind on real workloads, which is a different thing from saving heat.
An earlier version of this page told you to set Platform Thermal Throttle Limit to 80 °C on both chips, and gave the 9950X3D a 170 W PPT. Both were wrong, and applied together on a 9950X3D they produce in-game stuttering. A BIOS reset clears it.
170 W is the 9950X3D’s TDP, not its PPT. Stock PPT is 200 W, with TDC 160 A and EDC 225 A. So that setting cut 15% of the power budget while claiming to be stock.
The 80 °C cap is the bigger fault. Precision Boost is a closed loop that adjusts frequency up to a thousand times a second. Park it permanently against a ceiling well below its design target and it oscillates. Oscillation shows up as frame-time variance, which you feel as stutter, not as a lower average frame rate.
Before you tune anything
Is your temperature actually a problem?
These chips are designed to run to their limit and sit there. That is not a fault. But there is one reading that is, and telling the two apart decides whether you touch the BIOS at all.
Normal, by design
Hitting Tjmax under a sustained all-core load is how AMD intends these to behave. The chip boosts until it runs out of thermal headroom, then holds there.
Not normal · a cooling fault
Games load mainly the cache CCD and draw far less than an all-core render. TechPowerUp measures about 144 W gaming against roughly 200 W in heavy all-core work. On a working 360 mm AIO that should land near 70–80 °C.
Games load this chip lightly. Sustained ceiling temperatures while gaming, on a cooler that should easily handle 84 W, point at the mount rather than the settings.
So gaming at the thermal ceiling is a fault. Tuning around it hides it.
Tjmax is 95 °C, so a sustained 100 °C is not possible on a healthy part. Either a different sensor is being read, or something is genuinely wrong. In HWiNFO read CPU (Tctl/Tdie), not CPU Package, not the socket sensor, not a CCD hotspot.
If it is the cooling, check in this order
- Pump actually moving coolant. An RPM reading proves a tacho signal exists, nothing more. Feel both hoses; one should warm up.
- Cold plate contact and mounting pressure. The usual culprit on AM5.
- Paste coverage across the large AM5 heat spreader.
- Air lock in the loop, and radiator orientation.
- A contact frame is worth considering; AM5 socket flex is well documented.
- A 360 mm radiator is not undersized for either of these chips, so do not replace it first.
Measured
You give up very little until you go low
Both cards hold nearly all their performance well under their stock power. The curve only falls away sharply at the bottom end, which is exactly why a modest cap is such a good trade.
Performance kept against power limit
Each line starts at its card’s stock power on the right. The vertical bars show how much the answer moves between different games, which is the part a single average number hides.
At its 400 W floor the 5090 is drawing roughly what a stock 4090 draws, while still being comfortably faster. That’s the whole argument for capping it.
Where to click
Finding these settings
Menu names differ between boards and app versions, so the known variants are listed rather than one guess. Anything that couldn’t be confirmed from official documentation is flagged as such instead of being filled in.
ASUS BIOS · the processor settings Restart required
Get into the BIOS, then into Advanced Mode
Restart and press Delete or F2 as the ASUS logo appears. You’ll land in a simplified view. Press F7 for the full menus.
Leave the temperature ceiling alone
This field is here so you can find it and confirm it is on Auto. Do not set it to 80. Capping temperature makes the thermal limiter permanently active, and a permanently active limiter oscillates.
Label variants: "Platform Thermal Throttle Limit", "Thermal Limit", occasionally "TjMax". ASUS also exposes presets under "Enhancement" › "Thermal Limit"; leave those alone too.
Set the Curve Optimizer offset
Here’s the bit that catches people out. The sign and the number are two separate fields. You choose "Negative", then type 15. Don’t type −15.
› "CCD 1 Curve Optimizer Sign" › "Negative" › "CCD 1 … Magnitude" › 20
Set the power ceiling
PPT 120 W, TDC 90 A, EDC 140 A. Stock is 162 / 120 / 180.
PPT 190 W. Leave TDC at 160 A and EDC at 225 A, which are stock. Do not enter 170 W; that is the TDP, not the PPT.
Save it somewhere it will survive
Press F10 to apply and restart. Then save a copy, because a BIOS update or a cleared CMOS will wipe all of this.
Also export to a USB stick via "Load/Save Profile from/to USB Drive", and press Ctrl+F2 for a readable text copy. Note: only the .CMO file reloads, and ASUS only guarantees it on the same board with the same BIOS version, so the text copy or a photo is your real backup.
ASUS gives you two different pages for the same power settings, and they use different units. The Ai Tweaker page is in watts and amps. The AMD Overclocking page is in milliwatts and milliamps, labelled [mW] and [mA].
So don’t type 90 into a field asking for mW. Pick one page and stay on it, because the two don’t reliably mirror each other’s values.
MSI Afterburner · the graphics card curve Both builds
Version first. Afterburner 4.6.6 Final, build 16757, is the earliest with proper RTX 50-series support and is enough for a 5090. Older builds may not identify the card correctly.
Open the curve editor
Press Ctrl+F. It opens as its own window. Left to right is voltage in millivolts, bottom to top is clock speed in MHz. Closing this window does not apply anything.
Lift your chosen point
Click the node at 900 mV and drag it up to 2,600 MHz.
Click the node at 900 mV and drag it up to 2,700 MHz.
Arrow keys nudge one step, Ctrl with up or down moves ten, and pressing Enter on a selected point lets you type the number.
Flatten everything to the right of it
Hold Shift, click in empty space just left of your point, and drag a selection rectangle across to the far right edge. Then press Shift+Enter twice.
Note: the instruction you’ll see everywhere, "shift-click all the points to the right", isn’t quite right. MSI documents a shift-drag rectangle.
Apply, then check it took
Go back to the main window and click the tick. Then reopen Ctrl+F and look, because the card can round your nodes to its own steps.
Save the profile
Click the floppy icon to arm saving, then click a profile number. The Windows icon top right is "apply at startup", which is a different setting to "start with Windows". Don’t turn on apply-at-startup until it’s passed the tests below.
NVIDIA’s driver blocks the relevant hardware address range on most current 5080 and 5090 cards, so ticking "Unlock voltage control" in the settings cannot give you real voltage control on this card. The flattened curve is your voltage ceiling; there’s no separate maximum-voltage box to fill in.
ASUS GPU Tweak III · easier than Afterburner, honestly ASUS cards
If your card is an ASUS one, this is the gentler route. It has a single command that does the flattening for you, which Afterburner doesn’t.
Power and temperature targets live under "Home" › "Power", and the presets are Default, Overclocking, Quiet and Custom.
0dB only works in Auto fan mode. Switching the fans to Custom hands control to your own curve and removes the firmware’s stop-the-fans-entirely behaviour. If silence at idle matters to you, leave the fans on Auto with 0dB enabled and don’t draw a custom curve. Stock behaviour starts the fans above 55 °C or 100 W and stops them below 50 °C and 50 W.
Don’t run GPU Tweak III and Afterburner at the same time. Pick one.
The physical switch on the card Do this first
There’s a small slider on the card itself. The letters are printed next to it: P is Performance, Q is Quiet. Go by the printing, not by left or right, because which way round it sits depends on the model and on which side you’re looking from.
- Shut the machine down properly first.
- On the Astral it’s recessed along the backplate edge, near the power connector end.
- On TUF and Strix cards it’s on the exposed top edge, usually near the power connector.
- Slide it fully across to Q, then boot. The card reads the setting as it powers up.
ASUS’s own description is that Q keeps the same power target and top-end settings and only changes the fan curve. That’s why it costs you almost nothing.
Corrections
Where the popular advice is wrong
These all come up constantly, including from sources that are otherwise reliable. Each one below was checked against measurement.
Honest about the gaps
What isn’t settled
Six separate research runs went into this page and they didn’t agree on everything. Where they diverged, here’s the divergence rather than a confident average.
Which half of the 9950X3D takes the deeper offset
This chip has two halves, one with the extra cache and one that clocks higher. They want different offsets, and nobody agrees which way round.
- Common community reasoning: the cache half tolerates less, because the stacked cache limits voltage and heat.
- MSI’s own testing ran it the other way, at −25/−20 and then −31/−25, and reported about 6% better Cinebench.
No study across multiple samples exists either way. The −10/−20 above takes the cautious side of both, because getting the cache half wrong produces the nastiest symptom to diagnose: crashes that only appear when the machine is nearly idle. Start there, move in steps of 5, and test after each one.
What 105 W ECO Mode actually costs
Two outlets measured the same setting and got answers too far apart to be chip variation.
- TechSpot: about 25% off multi-core, with all-core clocks down nearly 30%.
- Club386: about 7% off, for a 38% smaller power budget.
Most likely one board applied a hard 105 W ceiling and the other mapped 105 W to a class with a roughly 142 W ceiling. You can settle it on your own machine in two minutes: turn it on, then watch PPT in HWiNFO. Whichever number you see tells you which measurement applies to you.
How much the 4090 loses at 70%
Reported as 2.2% by one outlet and 9% by another. That’s not a contradiction, it’s which games were in the test suite. Ray-traced scenes at 4K saturate the power budget, and rasterised ones frequently never reach it.
So the honest answer is a range, and where you land inside it depends on what you play. If you’re mostly in heavy ray-traced titles, expect the worse end.
Whether −15 is leaving something on the table
Probably, on a good chip. The point of −15 is that it should work on an average one.
I’d rather publish a number that holds for most samples than the best number I found in a forum thread. If you want more, move in steps of 5 and run the single-core test each time, because that’s the one that catches it.
How much of this is corroborated
Less than you’d hope. Across the six runs, source overlap was about 7%, and no single claim was made independently by more than one of them.
So treat each number as one well-sourced measurement rather than a consensus. The ranking used throughout: controlled multi-game power sweeps first, then chamber-measured thermal and acoustic testing, then vendor guides (real measurement, but one hand-picked sample), then forums. Forum data is genuinely the best evidence for how much chips vary, and useless for a single number.
Before you trust it
Testing that actually tests something
An undervolt that seems fine for a week and then crashes during a cutscene wasn’t tested, it was just used. These are the two tests that catch the failures each part actually has.
Graphics card
Don’t use FurMark. It slams into the power limit and drives the card down the curve, so it never visits the point you changed.
- Play something demanding for a couple of hours, ideally with loading screens and scene changes rather than one static scene.
- Add a 3DMark Speed Way or Steel Nomad stress run.
- Failure looks like: driver timeouts, a black screen that recovers, coloured artefacts, or crashes when a new area loads rather than under sustained load.
Processor
Cinebench isn’t enough. All-core loads run at lower voltage per core. Curve Optimizer problems show up at idle and under light load, where a single core boosts hardest.
Testing for stutter is a different test
A crash test will not find stutter, because stutter is not a crash. Average frame rate will not find it either; the average can be untouched while the experience is ruined.
- Capture frame times with CapFrameX or PresentMon and read the 1% and 0.1% lows, not the average.
- Log in HWiNFO: effective clock (not the reported clock), plus thermal throttling residency and PPT limit residency.
- Residency is the number that matters. A limit sitting active most of the time is the one causing the oscillation.
- Watch for clock stretching, where an over-deep undervolt quietly drops effective clock while still reporting nominal. Hitching, not a crash.
On dual-CCD X3D chips the AMD 3D V-Cache Performance Optimizer service intermittently fails to start, and games land on the wrong CCD. It happens on stock. If you tuned recently it is easy to blame the tuning, so check the service is running before you unpick BIOS settings.
If you ever move to a dual-CCD X3D chip, be aware that Windows CCD parking can misfire on its own and produce stutter that looks exactly like an unstable undervolt. Not applicable to this single-CCD part.
Change one thing at a time. If you set the temperature cap, the power cap and the curve together and it misbehaves, you’ve got three suspects and no way to tell which. Do them in the ladder order and test in between; it’s slower once and faster every time after.
AMD treats Precision Boost Overdrive, Curve Optimizer, undervolting and running EXPO outside published specification as overclocking, and says damage from it isn’t covered. That includes undervolting; "it’s gentler than stock" is not a warranty argument. On Threadripper 7000 AMD blows a fuse to record that it happened, although reporting suggests that alone doesn’t automatically void the cover.
Your call entirely. Just make it knowing that, rather than finding out later.
The 2023 Ryzen burnout episode was an overvolting fault, caused by too much SoC voltage. AMD’s firmware now caps that at 1.30 V. Undervolting doesn’t carry that risk, and you leave SoC voltage on Auto anyway. All you need to do is enable EXPO, then check in HWiNFO that CPU VDDCR_SOC hasn’t been pushed somewhere silly by the board.
If you remember three things
The short version
The free wins are the big wins
The Quiet BIOS switch costs under 1% and takes a minute. Everything after that is refinement, and the processor side is a power cap, never a temperature cap.
70%, not 80%
450 W, and 400 W is the floor
Because the card already draws about 340 W in games, a 360 W cap is theatre. Get to 70% before you expect anything to change.
450 W costs about 5%. The floor at 400 W averages around 10% but swings from 4% to 28% depending on the game, so check it against what you play.
Test the thing you changed
CoreCycler for the processor, real games for the card, and one change at a time. FurMark and Cinebench both pass undervolts that will bite you later.