most guides on this treat undervolting as one task with two tools. it isn’t. undervolting a graphics card works on nearly everything and the method is documented by the people who make the software. undervolting a processor depends entirely on whose chip you own, and on intel there’s a good chance voltage control has been closed off.
so this guide splits the two, and starts with the question nobody asks first: can you actually do it on your hardware?
what undervolting is, and what it isn’t
every chip ships with a voltage and frequency curve calibrated so the weakest silicon in the batch runs stable. your particular chip is probably better than that worst case, which means it’s being fed more voltage than it needs to hit the same clocks.
undervolting reclaims that margin. you lower the voltage while holding the frequency, and the chip does the same work with less power and less heat.
a conservative undervolt can reduce voltage, power draw and heat while preserving most or all of the original performance. but when clocks or power limits are also raised, undervolting becomes part of a broader tuning profile rather than a purely conservative change.
what it doesn’t do
it doesn’t repair a chip that’s already unstable, it doesn’t compensate for inadequate cooling, and it isn’t free performance in every case. results vary a great deal more than most guides admit.
if heat is the real problem, the mechanical side of cooling comes first: cleaning, paste and airflow fix what no voltage change can.
on amd graphics cards, undervolting can raise performance
before you start, it’s worth knowing that “success” looks different depending on whose card you own.
how the two vendors differ
| nvidia | amd | |
|---|---|---|
| what you edit | individual points on the voltage-frequency curve | a voltage offset applied to the whole curve |
| tool | msi afterburner curve editor | adrenalin driver, built in |
| what happens to the curve | you set a target point and limit behaviour beyond it | the entire curve shifts |
| typical outcome | same clocks at lower voltage | clocks can rise above stock |
| realistic goal | lower temperature and power at similar fps | lower temperature and potentially more fps |
as tom’s hardware explains, changing the voltage offset in amd’s radeon software doesn’t simply lower the operating voltage. it shifts the whole voltage-frequency curve, so the card can end up clocking higher than it did at stock.
what that looked like in testing
der8auer tested a powercolor rx 9070 xt red devil with a 170 millivolt negative offset and the power target raised to 110 percent. clocks went from 2.90 ghz at stock to 3.36 ghz, and cyberpunk 2077 at 4k ultra went from the card losing to an rtx 5080 to narrowly beating it, 66 frames against 65.
that isn’t a simple lower-power undervolt. it’s a hybrid tuning setup combining a negative voltage offset with a higher power limit to unlock additional clock headroom, so if your goal is to cut power draw, it isn’t the result you’re aiming for and isn’t what you’ll get.
tom’s hardware raises two caveats itself. the testbenches used across the comparison weren’t identical, and an rtx 5080 can also be overclocked by a similar margin. none of that makes one card better than the other. what it does show is that the headroom on rdna 4 is real, and larger than most people assume.
undervolting your graphics card
at least here the method comes from msi, who make afterburner, rather than from forum consensus.
before you start
you need msi afterburner 4.6.6 or newer. msi’s own guide notes that older builds don’t correctly recognise rtx 50 or rx 9000 cards.
you also need to unlock voltage control, which afterburner hides by default. it’s in settings, general tab, and two checkboxes matter: unlock voltage control and unlock voltage monitoring. leave force constant voltage off, since msi describes it as a diagnostic that pins voltage high and raises heat.
record a baseline before you change anything. clocks, voltage, temperature and power draw under a consistent load. without those numbers you can’t tell whether the undervolt did anything at all.
choosing your method
| power limit slider | curve editor | |
|---|---|---|
| what you change | a single percentage ceiling | individual voltage and clock points |
| difficulty | minimal | moderate, needs patience |
| time to a stable result | minutes | an hour or more with testing |
| performance cost | usually a small loss | can be none if tuned well |
| risk of getting it wrong | very low | instability until you find your point |
| best for | lower heat and noise with no fuss | squeezing out the best balance |
the simple method: the power limit slider
msi’s guide presents this as the starting point if your goal is lower temperatures, less fan noise and reduced power draw without much tinkering.
you drag the power limit slider down, testing in five percent steps between roughly 70 and 90 percent, applying and checking the effect on framerate and power each time. you’re not editing the curve at all. you’re setting a ceiling and letting the card work out how to stay under it.
it’s blunt, and you’ll usually give up a small amount of performance for the privilege. it’s also very hard to get wrong.
the precise method: the curve editor
press ctrl and f in afterburner to open the voltage-frequency curve editor. you’ll see a graph of core clock against voltage, which is your card’s factory-programmed behaviour at every voltage step.
the approach is to choose a voltage lower than your card’s default boost voltage, set the clock you want it to hold at that voltage, then limit the behaviour beyond that point so the card doesn’t climb back up the curve. how you constrain the upper portion varies with the card and the result you’re after, and there’s no single correct shape.
work in small steps and test after each one. moving a point by a large margin and hoping is how people end up with a card that benchmarks fine and then crashes in a menu.
msi’s guide reports reaching around 2.4 ghz at 800 millivolts on their test card without crashing. treat published voltage-and-clock examples as demonstrations of the method, not recommended settings. even two cards of the same model can have different stable limits.
testing it properly
msi recommends running a benchmark for ten minutes while watching for glitches, flickering, blackouts or crashes, then moving to real games.
the games matter more than the benchmark. an undervolt can be perfectly stable under sustained load and fail in a loading screen or a menu, where the card behaves differently. test across several titles before you trust it.
afterburner lets you save settings as numbered profiles, and there’s a lock icon that stops you overwriting them by accident, which is worth using on a configuration you spent an hour finding.
undervolting your processor: check first, download second
here’s where this stops being one article.
what’s actually available to you
| your cpu | can you undervolt | how | notes |
|---|---|---|---|
| amd ryzen | yes | curve optimizer, in bios under pbo or via ryzen master | official amd feature, can be set per core |
| intel k / hk / hx | usually | voltage offset in bios | the model lines that kept voltage control |
| intel non-k desktop | often not | bios option if the board exposes one | frequently locked at platform level |
| intel laptop | depends on the maker | bios option, if the manufacturer exposes one | mostly removed after 2019, returning on some models |
on amd, it’s a supported feature
ryzen processors expose curve optimizer through precision boost overdrive, available in your motherboard bios or through amd’s ryzen master software. it’s an official amd feature rather than a workaround.
curve optimizer doesn’t apply a fixed voltage offset. it shifts the voltage-frequency curve, and it can be set per core, which matters because cores within the same chip bin differently.
start conservatively. a small all-core negative value, in the region of minus five, is a reasonable first attempt rather than a target, and you work down from there in steps, testing at each one. copying a number from a forum tells you nothing, since the value that’s stable on someone else’s chip has no bearing on yours.
why “it didn’t crash” isn’t stability
ryzen catches people out here more than any other platform.
curve optimizer instability often doesn’t announce itself with a blue screen. it shows up as corrected hardware errors logged quietly in the background, or as crashes at idle and light load rather than under a stress test.
check the windows event viewer for whea-logger entries after each change, under windows logs and system. a machine that passes a stress test while logging corrected errors isn’t stable. it’s just not failing loudly yet.
on intel, voltage control is often closed off
in 2019 a vulnerability called plundervolt, tracked as cve-2019-11157, demonstrated that briefly dropping processor voltage through software could be used to inject faults and compromise intel’s secure enclave.
intel’s response was to disable the interface that made it possible. as the arch wiki documents, the recommendation to manufacturers was to enable the overclocking lock bit, which disables voltage control outright. on many newer intel systems, especially non-k desktop chips and laptops, undervolt protection can be enforced through processor and platform-level security controls.
in practice you may run into any of several barriers. your board’s firmware simply doesn’t expose the option, your laptop manufacturer removed it, oc lock is enabled, or the platform blocks it at a lower level. from the outside these look identical, and the outcome is the same either way.
so look in your bios first. if there’s no voltage offset or undervolt option there, a windows utility probably won’t find a way around it either, and guides telling you to install one anyway are worth treating with suspicion.
it’s slowly coming back
there’s some movement in the other direction. notebookcheck reported that xmg re-enabled bios-based undervolting on its raptor lake-hx laptops, after what they describe as a four-year hiatus following the plundervolt mitigations.
if you have a 13th or 14th gen intel desktop chip, read this
getting this part wrong costs real money, so it’s worth being precise about it.
what actually happened
intel’s 13th and 14th generation desktop processors suffered a well-documented instability problem caused by the chips requesting excessive voltage. intel shipped a series of microcode updates in response, of which 0x12b was the one identified at the time as the full mitigation.
the critical detail, reported by techpowerup, is that the microcode prevents the problem on chips that aren’t yet affected. it does not repair chips that already are, because the instability comes from irreversible physical degradation.
tom’s hardware reported the same conclusion and added something worse: the degradation can be symptomless, so a chip can be damaged without its owner knowing.
what to do instead of tuning
update your bios to the latest version available from your motherboard or system manufacturer. for affected 13th and 14th gen desktop processors, intel currently recommends a bios containing microcode 0x12f or later, together with intel default settings where available.
if your 13th or 14th gen chip is crashing, undervolting may reduce the symptoms for a while, and that’s precisely the danger. intel extended the warranty on affected processors by two years and offers replacement or refund for degraded chips, so masking instability with a voltage tweak burns that window on a part you may be entitled to have replaced.
if instability persists on current firmware, the right call is intel support rather than the bios.
does undervolting affect your warranty
there’s no single answer here, and anyone who gives you one is simplifying.
for processors, intel has stated that adjusting pl1 and pl2 power limits beyond the recommended values does not void warranty coverage, provided you stay within the iccmax threshold specified for the processor. that’s a clear position on power limits specifically, not a blanket statement about every kind of tuning.
for graphics cards the position varies by board partner rather than by gpu vendor, and consumer law in your country may matter as much as the printed policy. what can reasonably be said is that lowering voltage reduces electrical stress rather than increasing it, which is a different situation from overvolting.
and if a card is past saving and you’re weighing a swap, what a replacement card actually costs right now is worth knowing before you decide.
the distinction that matters most is reversibility. a software undervolt can be undone; a vbios flash can’t. flashing a modified bios onto a graphics card to achieve an undervolt sits well outside normal practice, carries a real risk to both the card and its warranty, and isn’t necessary to get the result described here.
what to expect, honestly
published results vary enormously, and most of the specific numbers circulating online come from sites with no testing behind them. the credible evidence supports something narrower than the headlines.
on amd rdna 4 cards there is real headroom, demonstrated in testing, and an undervolt can raise clocks rather than simply lowering power. on nvidia cards the realistic outcome is lower temperature and power draw at similar performance, which is worth having on its own terms.
on processors, the gain is meaningful on amd and frequently unavailable on intel.
if someone quotes you an exact voltage and an exact temperature drop for your specific chip, they’re guessing. the variation between individual chips is the entire reason any of this works.
the short version
check whether you can undervolt your cpu before you plan around it. on amd you can, through curve optimizer, starting small and watching for whea errors rather than just crashes. on intel you may not be able to at all, and if you have a 13th or 14th gen desktop chip with stability problems, this isn’t the tool for that job.
on the gpu side, start with the power limit slider, move to the curve editor if you want precision, and test in real games rather than benchmarks alone.
and record a baseline first, because the whole exercise is meaningless without a number to compare against.
