CPU Undervolting 2026: -50mV in 12 Steps, 60 Min

Overclocking gets the headlines. Undervolting gets the results. One of them is a teenager in 1998 gluing a fan to a Celeron 300A to hit 450 MHz on an Abit BH6 and posting the screenshot to Usenet; the other is a quiet, boring adult who wants that same chip to run ten degrees cooler, hold its boost clock through a two-hour Baldur's Gate 3 session, and never spin the fans loud enough to be heard over the emulator in the next room. This guide is for the adult.

Here is the thesis, stated plainly so you can leave now if you disagree: undervolting is the closest thing to a free lunch that modern silicon offers, and it is the mirror image of overclocking. Where overclocking pushes frequency up until the chip demands more voltage and more heat to stay stable, undervolting pushes voltage down until the chip runs the same frequency on less power. You do not lose performance. In most cases you gain a little, because every modern CPU is throttled by the first limit it hits — temperature, power, or current — and cutting voltage buys back headroom against all three at once.

It is also the one tuning knob that Intel once spent an actual security advisory trying to take away from you, which we will get to in detail. For now: reboot into your firmware, or don't, because there is a software path and a Linux path as well. This is the sixty-minute, twelve-step version, and it is honest about the parts that can quietly brick your afternoon. Nothing here voids a warranty the way a soldering iron does, but a bad offset can cost you a save file, so we test like we mean it.

Why Undervolt at All

The instinct, if you have ever tuned anything, is to ask what number goes up. Undervolting is unusual because the honest answer is none of the headline ones. Your advertised boost clock does not change. Your core count does not change. What changes is everything the marketing box does not print: how long the chip holds that boost, how loud the cooler has to work to let it, and how much of your power budget is wasted heating the room instead of running the game.

The point is headroom, not a bigger number

Modern CPUs are not clocked by you; they are clocked by an algorithm — Intel Turbo Boost, AMD Precision Boost — that ramps frequency up until it slams into the nearest ceiling. That ceiling is whichever of three walls it reaches first: a thermal wall (the temperature limit, typically 90–100°C), a power wall (Intel's PL1/PL2 or AMD's PPT), or a current wall (ICCMAX, or AMD's TDC/EDC). A stock chip at full tilt is usually pinned against one of them, throttling clocks down to obey it.

Undervolting attacks all three walls with one move. Less voltage at the same frequency means less power drawn, less current pulled, and less heat produced. The boost algorithm sees the extra headroom and does what it always does — clocks higher, or holds its clock longer, than it did before. That is why a correct undervolt frequently reads as a small performance gain in sustained multi-core workloads, most visibly on laptops and handhelds where the power and thermal walls are close and brutal. It is the same V/F-curve discipline behind GPU overclocking, run in reverse.

Thermals, acoustics, and the living-room emulation box

The retro crowd has a specific reason to care that the benchmark tourists do not. If you have built a small-form-factor HTPC to sit under the television running RetroArch, Dolphin, and a wall of shaders, the enemy is not frame time — a Wii game from 2008 is not troubling a modern CPU — the enemy is fan noise. An undervolt of 50–100 mV can drop package temperature enough to keep the fans in their quiet curve permanently, turning a whining shoebox into a silent one. Same clocks, same emulation accuracy, no audible cooling. For a box whose entire job is to disappear into the furniture, that is the whole point.

Longevity, and the reliability question nobody enjoys

Lower voltage and lower temperature mean less electromigration stress on the silicon over years of use — undervolting is, if anything, gentler on a CPU than running it stock. But the reliability conversation in 2025–2026 got complicated, and it is worth being precise. On August 25, 2025, Ars Technica reported that AMD, echoing Intel's earlier troubles, blamed motherboard makers for a run of burnt-out CPUs, with experts McAfee and Kirsch advising users to install the latest BIOS updates from their board vendors. The lesson for the undervolter is not don't touch voltage — it is that the voltage floor and the board's default behavior are set by firmware you do not control, and stale firmware is the actual hazard. Update first, tune second.

The Physics: V² Is the Game

You do not need a semiconductor degree to undervolt, but you need one equation, because it explains why the payoff is disproportionate to the risk. Everything about why this works — and why a tiny voltage cut yields a large power cut — is hiding in a single exponent.

Dynamic power scales with the square of voltage

The dynamic power a CMOS chip burns is P = α · C · V² · f — activity factor times capacitance times voltage squared times frequency. The term that matters is V². Because voltage is squared and frequency is linear, shaving voltage is a lever with mechanical advantage: a roughly 5% cut in voltage yields something like a 10% cut in power, at the same frequency. That is the entire reason this is worth an hour of your evening. You are not trading speed for efficiency; you are exploiting a nonlinearity that the stock configuration leaves on the table.

Guardband and the silicon lottery

Why is there voltage to cut at all? Because the manufacturer cannot test-and-bin every chip for your exact cooling, your exact board, and your exact ambient temperature. So they ship a guardband — a voltage margin above what a typical sample actually needs, wide enough that the worst-acceptable chip in the batch is stable in the hottest reasonable case. Your specific chip is almost never the worst case, which means your specific chip is running on more voltage than it needs. Undervolting reclaims that margin. How much margin exists is the silicon lottery: two identical SKUs off the same wafer can differ by 30–50 mV in how far they will go. This is why every number in this guide is a starting point, never a target. Your neighbor's stable −30 is your instant crash.

Boost is a race to the first limit

Tie the two ideas together and the strategy writes itself. Boost clocks are gated by temp, power, and current. Cutting V² lowers all three simultaneously. So the chip that was throttling at, say, 4.9 GHz because it hit its power limit can now sustain 5.0 GHz on the same wattage — you did not raise a clock, you removed the reason it was being held down. The effect is strongest exactly where the walls are tightest: single-threaded boost (one core at maximum voltage), thin-and-light laptops, and battery-limited handhelds. On a desktop with a 360 mm radiator you will mostly buy silence and a few degrees; on a gaming laptop you can buy real, measurable sustained clocks.

Prerequisites: Silicon & Software

Before you reboot into anything, sort out three questions: can your specific CPU even be undervolted, is your board's firmware current, and do you have the right tools installed. Getting these wrong wastes an hour chasing a setting that silently does nothing.

Which CPUs can actually undervolt (the Plundervolt problem)

This is the part every lazy tutorial skips. In 2019, researchers demonstrated Plundervolt (CVE-2019-11157, CVSS 7.9) — an attack that abused the software voltage interface to fault Intel's secure enclave. Intel's December 2019 microcode fix let firmware lock the voltage-control register, MSR 0x150, so that writes to it silently no-op. That microcode shipped for 6th- through 10th-generation Core and Xeon, and it is the reason the Intel undervolting story pivots at the 12th generation. In practice: most modern Intel desktop K-SKUs on Z-series boards undervolt fine, and mobile HX chips generally do — but the majority of non-HX 11th–14th-gen laptops are locked unless the OEM chose to expose a BIOS toggle. Intel's own support note on undervolt protection (updated July 4, 2026) states plainly that when undervolt protection is not configured, the minimum voltage is defined by BIOS settings — the board vendor sets your floor, not you. AMD Ryzen has no equivalent lock: Curve Optimizer is a first-class, documented feature on Ryzen 5000, 7000, and 9000.

Boards, firmware, and clearing CMOS

You want a board that exposes voltage or Curve Optimizer controls, and you want it on current firmware. Intel offsets realistically require a Z-series chipset and an unlocked (K) CPU; AMD is more forgiving, as PBO and Curve Optimizer appear on most B- and X-series AM4/AM5 boards. Update the BIOS before you start — this is the single most important prerequisite per the Ars Technica reporting above, and if you are shopping the board itself, note that even 2026's flagship PCIe 6.0 motherboards ship undervolting controls buried in the same overclocking menus. Then, before you change one setting, find your CMOS-clear method: the physical CLR_CMOS jumper, the rear I/O button, or a 30-second coin-cell pull. You will POST-fail at least once. Knowing the recovery in advance turns a panic into a shrug.

The software toolbox and its 2026 versions

Install these before rebooting so a fresh, stable desktop is one click away:

• HWiNFO64 (8.5x; the 8.51 build restored the Custom and WHEA sensors) — your source of truth for Vcore, package power, temperature, and effective clocks.
• Intel XTU — v7.14.x for 14th-gen and older (7.14.2.93, July 2026); v10.0.x for Core Ultra Series 2 (10.0.1.45, March 2026). Needs a Z-chipset.
• AMD Ryzen Master 3.1.x for AM5 / Ryzen 9000, with a built-in stress test. Run it as administrator or Curve Optimizer greys out.
• ThrottleStop 9.7 (stable) / 9.7.3 (beta, April 2025) — the standard tool for locked-down Intel laptops.
• Stress tools: Cinebench 2026 (Maxon; note that its scores are ~3× higher than the 2024 release and are not cross-comparable), OCCT, and Prime95 for Small FFTs.

Intel Offsets vs AMD Curve Optimizer

The two platforms expose undervolting through two different mental models, and the single most common beginner mistake is treating a number from one as if it means the same on the other. It does not. Learn both units before you touch either menu.

Intel: a negative voltage offset, in millivolts

On Intel you apply a negative voltage offset measured in real volts. Set the mode to Offset, the sign to negative, and a magnitude — the 2025–2026 consensus conservative start is −0.050 V (50 mV), refined toward −0.075 V and beyond if it holds. This offset is curve-wide: it shifts the entire voltage/frequency curve down by that amount at every operating point. That is the behavior you want for a set-and-forget undervolt. Avoid Override mode (it pins one fixed voltage and kills boost behavior); Adaptive tunes only the top of the curve. Offset is the honest default. A 2026 guide from Evezone frames the practical Intel range for 12th-gen and later — Alder Lake, Raptor Lake, Arrow Lake — as roughly −50 mV to −100 mV via XTU or BIOS core-voltage offsets.

AMD: Curve Optimizer counts, not volts

AMD does not hand you volts. Curve Optimizer works in "counts" — step units along the per-core voltage/frequency curve — and each count is worth only about 2–5 mV, and the exact conversion shifts with your AGESA firmware version. The standard starting point in 2025–2026 guides is All Cores, Negative, magnitude 10 ("−10"), then testing whether −15 or −20 holds; the same Evezone guide notes many Ryzen 5000/7000 owners begin more aggressively at −20 to −30 all-core before validating. Ryzen 9000 samples commonly settle in the −15 to −25 all-core range. Because the count is not a volt, −10 on AMD is nowhere near −10 mV — it is closer to −20 to −50 mV of effective offset. Confuse the units and you will either undershoot badly or crash on the first boot.

Why the unit mismatch matters when you compare notes

Multiple 2026 guides go out of their way to stress that Curve Optimizer values are step units, not voltages, precisely because forum advice mixes the two constantly. When you read "I run −30 and it's rock solid," your first question must be which platform. −30 counts on a lucky Ryzen is routine; −30 mV on Intel is a mild, cautious offset. They are not the same claim, they are not the same risk, and they are not transferable between your machine and anyone else's regardless — silicon lottery sees to that.

The BIOS Method: 12 Steps

BIOS is the persistent path. Software tools are fine and we cover them next, but a value you set in firmware applies before the OS loads, survives a Windows reinstall, and does not depend on a background process staying alive. For a long-term gaming build, that is the setting you want. Here is the full sequence, each step with its reason.

The twelve steps, with rationale

1. Update the BIOS to the current release. Board vendors ship AGESA and voltage-behavior fixes constantly; the Ars Technica reporting tied burnt CPUs to board defaults, and Intel's floor is BIOS-defined. Start from current firmware or your baseline is a moving target.
2. Learn your CMOS-clear procedure now. Jumper, button, or coin-cell pull — confirm which one your board uses before you change a setting. You will POST-fail at least once, and you want the fix to be muscle memory, not a 1 a.m. web search on your phone.
3. Record a baseline. Note stock Vcore, idle and load temps, and a stock Cinebench 2026 score. You cannot prove you gained anything without a before number, and a sane baseline confirms your cooling is not already the bottleneck.
4. Enter firmware setup. Tap Del, F2, or F12 at POST. This is the persistent, OS-independent layer — the reason we are here rather than in a Windows app.
5. Leave Load-Line Calibration on Auto. A May 23, 2026 guide's advice, and correct: LLC injects extra voltage under load and muddies which change caused an instability. Tune one variable at a time; touch LLC only much later, if at all.
6. Set your first offset. Intel: find CPU Core Voltage, set Mode = Offset, Sign = negative, Magnitude = 0.050 V. AMD: open Precision Boost Overdrive → Curve Optimizer, set All Cores, Negative, Magnitude = 10. These are the conservative 2026 starting rungs for each platform — deliberately timid so the first boot is boring.
7. Save and boot to the desktop. If the machine will not POST, your very first offset is already too deep for a cold boot — clear CMOS and halve it. A clean boot to desktop is the gate for everything after.
8. Idle-soak for a few minutes. Sit at the desktop and watch for a spontaneous reboot or freeze. Ryzen in particular fails at idle and single-thread first, not under load — do not skip this because "it booted fine."
9. Run a short stress test. Cinebench 2026 multi-core, or OCCT, or Prime95 Small FFTs, for 10–15 minutes minimum. This validates the all-core, full-power, full-heat case that idle-soaking cannot.
10. If stable, deepen by one step and repeat 7–9. Intel to −0.075 V; AMD to −15. Walk it down in small increments — never jump straight to someone's forum number. Each rung earns its own boot-idle-stress cycle.
11. When you find the edge, back off one full step and add margin. The step that crashes tells you the limit; retreat one rung past it. "Stable for fifteen minutes" is not "stable for a six-hour raid," so leave headroom on purpose.
12. Save a named BIOS profile and run a long soak. Store it as, e.g., "UV-75mV" so a CMOS clear does not erase your work, then run a multi-hour or overnight validation before you trust it with a save file.

Reading your current Vcore

You cannot tune what you cannot see. Before and after each change, watch Vcore, package power, temperature, and effective clock in HWiNFO64 on Windows, or with sensors and turbostat on Linux. An illustrative Linux readout after a clean 75 mV cut — your silicon will differ — looks like this:

$ sensors
coretemp-isa-0000
Adapter: ISA adapter
Package id 0:  +58.0°C  (high = +100.0°C, crit = +100.0°C)
Core 0:        +54.0°C
Core 1:        +55.0°C
Core 2:        +53.0°C
Core 3:        +54.0°C

$ sudo turbostat --interval 1
CPU   Avg_MHz  Busy%  Bzy_MHz  PkgWatt  PkgTmp
-       4785   99.4     4812     92.1    58

What a good result looks like

A successful undervolt reads as: identical or slightly higher effective clocks, several degrees lower package temperature, and a visibly lower PkgWatt figure at the same load — with zero crashes, zero freezes, and (crucially) zero WHEA corrected-error events, which we validate in the stability section. If temps and watts did not move at all, the setting did not take — jump to the troubleshooting table, because you are very likely locked.

The Software Route: XTU & Ryzen Master

If your board's BIOS is a hostile maze, or you want to iterate quickly before committing a value to firmware, the vendor software does the same job from inside Windows. The trade-off, per every 2026 guide, is persistence: BIOS survives everything, whereas a software offset depends on the tool relaunching at login. Prototype in software; commit the winner to BIOS.

Intel XTU

Intel Extreme Tuning Utility gives you a live Core Voltage Offset slider without a reboot, which makes the boot-stress-repeat loop far faster to iterate. Use v7.14.x for 14th-gen and older and v10.0.x for Core Ultra Series 2. Set the offset negative, start at −0.050 V, and lean on the same increments as the BIOS method. Remember XTU realistically wants a Z-series chipset; on a locked board the slider will move and change nothing. When you settle on a value, transcribe it into BIOS as an Offset and stop relying on the app.

AMD Ryzen Master

Ryzen Master 3.1.x exposes Curve Optimizer per-core and all-core from the desktop, and it ships a built-in stress test that is convenient for a first sanity pass. Two rules: run it as administrator or the controls grey out, and make sure PBO is set to Auto or Enabled — Curve Optimizer is a modifier on top of Precision Boost Overdrive, so with PBO off the curve does nothing. Validate your memory overclock first: an unstable EXPO profile will masquerade as an unstable undervolt and send you chasing the wrong variable. If you are still deciding on kit, our DDR5 vs DDR6 breakdown covers why EXPO validation matters before any curve tuning.

ThrottleStop for locked laptops

On Intel laptops where the BIOS hides everything, ThrottleStop is the escape hatch — when the silicon allows it. Open the FIVR window, tick Unlock Adjustable Voltage, and drag the Offset Voltage slider negative, matching your CPU Core and CPU Cache offsets to the same value. If Unlock Adjustable Voltage is greyed out, you have met the Plundervolt lock in person and no software will move MSR 0x150. On a locked chip your fallback is thermal, not electrical: use Speed Shift EPP (0–255) and the TPL power limits to shave sustained power indirectly. It is not a true undervolt, but on a throttling ultrabook it recovers real clocks.

Undervolting on Linux

If your emulation box or gaming rig runs Linux — and a startling number of retro setups do — the toolchain is different but the physics and the locks are identical. The ArchWiki CPU undervolting page (updated August 4, 2026) is the canonical reference, and it gives a usefully broad range: decreasing Intel CPU and CPU-cache voltage by 100 to 200 mV is usually stable — a wider window than the Windows guides quote, but still bounded by your silicon lottery.

intel-undervolt and the MSR 0x150 domains

The tool of record is intel-undervolt, which writes the same MSR 0x150 the Windows tools use — meaning the exact same Plundervolt lock applies, and on a locked chip the writes silently no-op here too. It addresses five voltage planes by index: 0 = CPU core, 1 = GPU, 2 = CPU cache, 3 = System Agent, 4 = Analog I/O. A conservative, ArchWiki-aligned config:

# /etc/intel-undervolt.conf
# domain index: 0=CPU core, 1=GPU, 2=CPU cache,
#               3=System Agent, 4=Analog I/O
# values are millivolts; negative = undervolt
apply undervolt 0 'CPU'          -100
apply undervolt 1 'GPU'           -50
apply undervolt 2 'CPU Cache'    -100
apply undervolt 3 'System Agent'    0
apply undervolt 4 'Analog I/O'      0

# optional package power limits (watts): short long
apply power package 35 45

Apply and verify from a shell. If rdmsr 0x150 reads back nothing resembling your write, or temps do not budge, you are locked:

# Arch Linux
sudo pacman -S intel-undervolt

# read the current interpreted values
sudo intel-undervolt read

# apply the config above
sudo intel-undervolt apply

# Plundervolt-lock sanity check
sudo modprobe msr
sudo rdmsr 0x150

RyzenAdj: power limits, not voltage

On AMD, the Linux equivalent people reach for is RyzenAdj — but be clear about what it does. RyzenAdj tunes power limits (STAPM, fast, slow) and thermal targets, not Vcore. There is no Curve Optimizer on Linux; the closest analogue lives in BIOS. What RyzenAdj gives you is an undervolt by proxy: cap the power envelope and the boost algorithm holds lower voltages to fit inside it. This is exactly the lever that matters on AMD handhelds, where it trades a sliver of peak clock for meaningfully longer battery in emulation — the same tuning that stretches a session on the Retroid Pocket handhelds and their APU cousins:

# handheld / APU power tuning (milliwatts; tctl in °C)
# RyzenAdj tunes POWER, not vcore -- undervolt by proxy
sudo ryzenadj \
  --stapm-limit=15000 \
  --fast-limit=18000 \
  --slow-limit=15000 \
  --tctl-temp=90

Persisting across suspend and reboot

The single most common Linux failure is the offset vanishing after a suspend/resume cycle, because the MSR resets and nothing re-applies it. intel-undervolt ships a systemd service and a resume hook for exactly this — enable it and the value survives sleep and reboot:

# re-apply on boot AND on resume-from-suspend
sudo systemctl enable --now intel-undervolt.service

Stability Testing

An undervolt is not "done" when it boots. It is done when it has survived a battery of tests designed to find the instability that a five-minute glance never will. The core loop, unchanged since the guardband existed, is boot → stress → repeat after every single change. Skipping it is how you get silent save-file corruption a week later.

The boot-stress-repeat loop

After each offset change: boot to desktop, idle-soak, then stress for at least 10–15 minutes per step — 30 minutes once you are near the edge. Run each rung to completion before deepening. The discipline is boring and non-negotiable; the crash you are hunting frequently takes twenty minutes and a specific instruction mix to surface, and it will surface in your game at the worst possible moment if you did not provoke it on your own terms first.

The tools and what each one catches

Prime95 Small FFTs is the maximum-heat, maximum-current torture test — it finds the thermal and current edge fast. OCCT is friendlier and includes an error-detecting mode that flags computational faults, not just crashes. Cinebench 2026 multi-core is the realistic all-core workload and doubles as your before/after performance number (remember: its scores are not comparable to Cinebench 2024's). And on Ryzen, CoreCycler — sp00n's script — is essential, because it drives Prime95 single-core across every core in turn at maximum boost, catching the per-core weakness that all-core tests hide. No single tool is sufficient; the failure modes are different and so are the tests that expose them.

Reading WHEA: the silent tell

Here is the part beginners never check and experts never skip. A machine can be "stable" — no crashes, no freezes — while quietly logging corrected hardware errors, meaning the CPU made a mistake and the error-correction machinery caught it. That is your undervolt failing silently, and it will eventually escalate to corruption. On Windows, hunt for WHEA-Logger Event ID 18 and 19 in the System log after every stress run:

# PowerShell: find silent machine-check corrections
Get-WinEvent -LogName System -MaxEvents 200 |
  Where-Object { $_.Id -eq 18 -or $_.Id -eq 19 } |
  Format-Table TimeCreated, Id, ProviderName -AutoSize

An illustrative hit — one is one too many — looks like:

TimeCreated            Id  ProviderName
-----------            --  ------------
8/14/2026 11:42:07 PM  19  Microsoft-Windows-WHEA-Logger

See a single Event 18 or 19 and you back off — roughly 5 counts on AMD, or 0.010–0.020 V on Intel — and re-test until the log stays clean through a full soak. A clean WHEA log is the real pass condition, not "it didn't crash."

Five Ways This Goes Wrong

Most failed undervolts are not failures of the silicon. They are failures of method — the same handful of mistakes, repeated across every forum thread, that turn a sixty-minute tune into a weekend of chasing ghosts. Here are the ones that cost people the most time, and the fix for each.

• Chasing someone else's number. The single biggest one. "−30 is stable on the 7800X3D" is a statement about their chip, not the SKU. Fix: treat every published value as a starting rung and walk your own curve down. The silicon lottery is real and it does not care what Reddit says.
• Testing only under full load. All-core stress makes an aggressive undervolt look stable because Vdroop and lower per-core boost hide the fault. Fix: idle-soak and single-core test (CoreCycler) as seriously as you all-core test. See the idle paradox below.
• Stacking variables. Enabling EXPO/XMP, a manual overclock, and an undervolt in one sitting, then blaming the undervolt when it crashes. Fix: change exactly one thing at a time; validate your memory profile as stable on its own before you touch voltage.
• Trusting a ten-minute test. Short tests miss the instability that needs heat soak or a specific instruction mix. Fix: a real validation is 30 minutes of Prime95/OCCT plus an overnight soak plus a clean WHEA log — and then a few days of actual gaming before you call it permanent.
• Forgetting the offset is curve-wide. On Intel, Offset mode moves the whole V/F curve; a value that is fine at 5.0 GHz can starve the chip at a low-power idle point. Fix: understand Offset vs Adaptive, and always validate low-load and idle states, not just the boost ceiling.
• Not writing down the CMOS-clear first. The classic. You will need it, usually at midnight. Fix: locate the jumper or button before step one, so recovery is thirty seconds instead of a spiral.

The idle paradox, expanded

The counter-intuitive Ryzen failure deserves its own paragraph because it fools nearly everyone. Under a heavy all-core load, voltage droops (Vdroop) and no single core is boosting to its absolute peak, so an over-aggressive curve looks rock solid. Sitting at the desktop doing nothing, the CPU boosts one core to maximum frequency at maximum voltage-minus-your-offset — and if the offset is too deep, that lone high-frequency core starves and the whole system freezes or reboots "for no reason." If your machine crashes browsing the web but never in Cinebench, this is why. Validate 1T and idle explicitly.

Change one thing, write it down

The discipline that prevents most of the above is a boring log: one change, one test, one line recorded. Offset value, test run, result, WHEA count. When something breaks three steps later, the log tells you exactly which change to blame instead of forcing a full reset. Tuners who keep the log finish in an evening; tuners who wing it are the ones posting "help, random crashes" a week later.

Troubleshooting Table

When something misbehaves, match the symptom to its most likely cause before you start changing values at random. The table below covers the failures you will actually hit, in rough order of frequency.

The symptom-to-fix table

Clearing CMOS without panic

The failures at the top of that table share one fix, and it is not scary: clearing CMOS resets firmware to defaults and undoes whatever you just did. Power off, pull the wall plug, then either short the CLR_CMOS jumper for a few seconds, press the rear button if your board has one, or remove the coin cell for about thirty seconds. The machine boots to stock, your undervolt is gone, and you start again one rung shallower. Nothing is damaged. This is the whole reason undervolting is low-risk — the worst common outcome is a reset, not a repair.

When nothing changes: you are locked

The most frustrating row is the one where the undervolt does nothing. No temperature drop, no wattage drop, no crash — the setting simply has no effect. On Intel that is almost always the Plundervolt lock: the microcode is silently discarding your writes to MSR 0x150. Confirm it by checking whether ThrottleStop's "Unlock Adjustable Voltage" is greyed out, by reading MSR 0x150 on Linux, or by looking for CFG Lock / Overclock Lock in BIOS. On a locked non-HX laptop there is no software workaround; occasionally an older BIOS restores the interface, but that is a decision to weigh against the security fix it removes. Pivot to power-limit and Speed Shift tuning instead.

Advanced: Power Limits & Per-Core

Once the basic all-core offset is stable, three advanced levers extract the rest of the margin: power-limit tuning, per-core curves, and — for handhelds — pure power-envelope control. None are mandatory. All are where the last few degrees and the last hundred megahertz live.

Power-limit tuning on AMD

A 2026 approach pairs the Curve Optimizer undervolt with gentle power-limit tuning rather than one dramatic cut. On AMD, lower PPT (package power), TDC (sustained current), and EDC (peak current) together in small steps — this reshapes where the boost algorithm spends its budget and can drop temperature and noise further without touching the curve at all. The key word is gradual: shave, test, repeat, exactly as with voltage. On Intel the analogue is PL1/PL2; leaving them stock while you tune the offset first, then tightening them, keeps the variables separated.

Per-core Curve Optimizer

All-core is a compromise pinned to your weakest core. Per-core Curve Optimizer lets you undervolt each core to its own limit — and the counter-intuitive truth is that the best cores tolerate the least. Your highest-boosting preferred cores run closest to their own edge already, so they take only −5 to −15; the lazy, weaker cores that never boost as high will happily swallow −25 to −35. Identify preferred cores (Ryzen Master stars them), give the strong ones a conservative curve and the weak ones an aggressive one, and validate every core individually with CoreCycler. It is the most tedious and most rewarding phase.

Handhelds and emulation battery life

On an AMD handheld the goal is not silence, it is minutes. Because there is no Curve Optimizer on the Linux side, the RyzenAdj power-envelope approach from earlier is the lever: cap STAPM and the fast/slow limits low enough that the SoC holds lower voltages, and a PS2 or GameCube session that drained the battery in ninety minutes can stretch considerably with no perceptible drop in emulation smoothness. Modern emulators rarely need the full power budget; handing that budget back is free runtime. This is the same efficiency argument as the desktop, just measured in battery percentage instead of decibels.

A Complete Working Configuration

Here are three known-good starting templates — one Intel, one AMD, one Linux — that encode everything above. They are starting points, not magic numbers: apply the shallower value, validate with the full stability loop, then deepen. Your silicon will land somewhere near these, rarely exactly on them.

The Intel desktop profile

=== Intel (Z-series board, K-SKU) — persistent BIOS profile ===
CPU Core Voltage Mode ......... Offset
Offset Sign ................... -  (negative)
Offset Magnitude .............. 0.075 V   (start 0.050, refine)
CPU Cache / Ring Voltage ...... follow core, -0.075 V
Load-Line Calibration ......... Auto
Long Duration Power  (PL1) .... stock
Short Duration Power (PL2) .... stock
Undervolt Protection / IA CEP . Disabled  (only if exposed)
Save as profile ............... "UV-75mV"
# ArchWiki notes core+cache -100 to -200 mV is often stable;
# 0.075 V is a conservative, WHEA-clean daily target.

The AMD Ryzen profile

=== AMD (AM5, Ryzen 9000) — persistent BIOS profile ===
Precision Boost Overdrive ..... Advanced
PBO Limits .................... Motherboard (or manual below)
  PPT (W) ..................... 120
  TDC (A) ..................... 90
  EDC (A) ..................... 140
Curve Optimizer ............... All Cores
Curve Sign .................... Negative
Curve Magnitude ............... 20   (counts; ~2-5 mV each)
Max CPU Boost Clock Override .. 0 MHz  (leave stock first)
EXPO / DDR5 ................... validated separately, first
Save as profile ............... "CO-20"
# Start All-Core -10, walk to -15/-20/-25; a single WHEA
# Event 18/19 means back off ~5 counts and re-test.

The Linux profile

On Linux the complete configuration is the intel-undervolt.conf from the Linux section, paired with the enabled systemd service so it survives suspend and reboot. Together they are the full persistent setup:

# /etc/intel-undervolt.conf  (core + cache, conservative)
apply undervolt 0 'CPU'        -100
apply undervolt 2 'CPU Cache'  -100
apply undervolt 1 'GPU'         -50
# then make it permanent:
#   sudo intel-undervolt apply
#   sudo systemctl enable --now intel-undervolt.service
# verify clean: no crash, no WHEA/MCE in `journalctl -k`,
# temps and PkgWatt down at the same effective clock.

That is the whole discipline: cut voltage, not frequency; move one variable at a time; validate until the error logs are silent, not merely until the crashes stop. Done properly, undervolting is the rare tuning move with no downside on the box and a genuine payoff off it — cooler, quieter, longer-lived silicon that clocks exactly as high as it did before, and holds it longer. The teenager gluing a fan to a Celeron had the more exciting hobby. The adult undervolting a Ryzen has the better-run machine.