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

Let us begin with the disappointment, because it is load-bearing. If you searched for a 2026 undervolting method expecting a new tool, a new proprietary API, or a magic number that a decade of tinkerers somehow overlooked, you have been misinformed by the calendar. There is no 2026 method. There is a 2016 method that people keep re-dating and re-uploading, and the physics underneath it has not so much as twitched since Intel's Skylake and AMD's first Zen taught a generation of enthusiasts that the voltage stamped on the box is a suggestion, not a fact. This guide will teach you that method properly, in the order of operations that actually keeps your machine bootable, and it will tell you the two things that genuinely did change in the last eighteen months, both of which are reasons to update your BIOS before you touch a single millivolt.

The 2026 Non-Event

Every year a fresh crop of guides appears with the year in the URL and a promise of new secrets. The secret is that there are no new secrets. The negative voltage offset, AMD's Curve Optimizer, the FIVR window in ThrottleStop — these are the same three levers your predecessors pulled in 2018, and the silicon responds to them in exactly the same way. If your goal is a cooler, quieter, longer-boosting CPU, the good news is that the technique is mature, documented, and free. The bad news, for anyone hoping to feel like a pioneer, is that you are walking a paved road.

Nothing New Under the Heatsink

The utilities did get version bumps, and it is worth knowing which ones, because a version number is the only thing in this hobby that legitimately carries a date. ThrottleStop sits at 9.7 stable with a 9.7.3 beta that added Arrow Lake support in April 2025. Intel's own Extreme Tuning Utility jumped to version 10, exclusive to the Core Ultra 200S "Arrow Lake" family, while older Intel platforms stay on the 7.x branch. AMD's Ryzen Master and its Curve Optimizer got quietly refined for Zen 5. But a version bump is not a paradigm. The dial is the same dial. Anyone advertising a "revolutionary 2026 approach" is selling you a screenshot with the year edited and the substance untouched.

Who Actually Benefits

Undervolting pays the largest dividend to the machines with the least thermal headroom. Laptops throttle within seconds of a sustained load; shaving 80–120 mV off the core can be the difference between a chip that holds its boost clock and one that sags to base. Small-form-factor desktops fighting a single 92 mm fan benefit for the same reason. And anyone running a furnace of a GPU — the sort of build where a 575-watt RTX 5090 dumps its heat straight into the case and back onto the CPU cooler — will notice that a cooler processor means the whole thermal budget stops fighting itself. Undervolting is, in the most literal sense, the inverse of the GPU overclocking we covered earlier: same voltage-frequency curve, same stability discipline, opposite direction of travel.

The One Thing That Did Change: Raptor Lake

Here is the genuine 2024–2026 development, and it is a cautionary tale, not a feature. Intel's 13th and 14th Gen "Raptor Lake" desktop chips — the 65 W-and-higher parts — suffered a Vmin-shift bug in which the CPU requested progressively more voltage over its life and degraded irreversibly. Intel issued four public statements, then two microcode updates (0x125 and 0x129) consolidated into the final 0x12B mitigation in September 2024, and extended the warranty on affected models. The relevance to you is direct: if you undervolt a Raptor Lake chip that has not received 0x12B, you are chasing a moving target, because the chip's voltage demand is itself unstable. Update the BIOS first. This is the one 2026-specific instruction in the entire guide, and it exists precisely because the tools did not change but the errata did.

What Undervolting Actually Is

Before you touch a menu, you should understand what you are negotiating with, because undervolting is a negotiation with a factory that deliberately overpaid. The processor does not run at one voltage. It runs along a curve, and every point on that curve was set by a binning process that had to account for the worst chip that would ever ship under that model number. You almost certainly did not get the worst chip. Undervolting is the act of proving it.

The V/F Curve and the VID Table

Every modern CPU carries an internal voltage-frequency table — Intel calls the requested value the VID, AMD exposes it through the VFT curve. For each frequency the chip intends to hit, the table specifies a voltage the chip will request from the motherboard's VRM. That requested voltage includes a guardband: extra margin baked in so that the slowest, leakiest, least cooperative die in the entire production bin will still be stable at that frequency, at high temperature, after years of aging, on a cheap board. Your specific chip, on your specific board, at your specific temperature, does not need all of that margin. The margin it does not need is heat you are paying for and clock speed you are throwing away.

Why the Free Performance Is Real

Dynamic power scales with the square of voltage. Drop the voltage 10% and, holding frequency constant, you shed roughly a fifth of the dynamic power as heat. That is not a rounding error; it is the entire game. On a thermally or power-limited chip — which in 2026 is nearly every chip, because both vendors ship parts that boost until they hit a wall — lower voltage at a given clock means the wall arrives later, which means the boost algorithm holds higher clocks for longer. This is why a correct undervolt frequently produces an identical or slightly higher Cinebench score while running fifteen degrees cooler. You did not lose performance to save heat. You converted wasted heat back into sustained frequency. The phrase to internalize is "same performance, lower temperature," and on constrained hardware it becomes "more performance and lower temperature."

The Silicon Lottery Is Your Actual Opponent

There is no universal undervolt. The -50 mV Intel offset and the -10 Curve Optimizer count you will see quoted everywhere — including in this guide — are not targets; they are opening probes, the conservative first tap on a door to see whether it opens. Two identical chips off the same wafer can differ by a factor of two in how far they will undervolt. SkatterBencher's own tuning notes describe a single Ryzen part with four cores stable at a Curve Optimizer value of -30 and two cores that fell over at -13. That spread is normal. It means the only way to find your number is to walk toward the failure point in small steps and stop one step before it. Anyone who hands you a fixed value and calls it "the" undervolt for your model is describing their chip, not yours.

Prerequisites & Versions

Undervolting is cheap in money and expensive in patience. Every tool below is free. What you are actually spending is an afternoon of change-one-thing-and-test discipline. Assemble the kit before you start, because the worst time to go hunting for a stability tester is thirty minutes into a session when you have forgotten which offset you last applied.

Hardware You Need

On the Intel side, full offset control lives in the UEFI of a Z-series board paired with an unlocked K-series chip, though many B- and H-series boards still expose a core voltage offset even when they lock the multiplier. On the AMD side, Precision Boost Overdrive and Curve Optimizer are available across the B650/X670/B850/X870 range and even most B550 boards, because AMD treats tuning as a platform feature rather than a chipset upsell. Laptops are their own category, handled in the ThrottleStop section, and many of them will refuse you outright for reasons explained there. In all cases you need a cooler good enough to establish an honest baseline; you cannot measure a thermal improvement against a heatsink that was already saturated.

Software and Exact Versions

Pin these down before you begin. Use the latest stable build of each unless noted — there are no paid tiers and no 2026 "pro" editions, whatever the ad copy claims.

• HWiNFO64 (latest) — your source of truth for real Vcore, per-core VID, package power, and temperature. If a sensor and a menu disagree, believe the sensor.
• Cinebench 2024 (Maxon) — the current release, built on the Redshift engine. Note the name: it is Cinebench 2024, not "R24"; the old R-numbered line ended at R23. Its ten-minute stress option exposes throttling and instability under sustained load.
• OCCT (latest, from ocbase.com) — the workhorse stability tester, with a small-data-set AVX2 mode and, critically, built-in error detection that catches corrected machine-check errors most other tools ignore.
• Prime95 30.x (Small FFTs) — maximum current and heat, the brute-force final boss of stability.
• CoreCycler (sp00n on GitHub) — essential for AMD. A PowerShell script that pins single-worker Prime95 to each physical core in turn, catching single-core boost instability that all-core tests structurally cannot.
• Intel XTU — version 10.x for Arrow Lake / Core Ultra 200S, 7.x for prior Intel desktop generations.
• AMD Ryzen Master (latest) — AMD's own tuning suite, useful for live Curve Optimizer changes without a reboot.
• ThrottleStop 9.7 stable, or the 9.7.3 beta if you own an Arrow Lake laptop — Intel mobile only.

The Pre-Flight Checklist

Do these three things in order and you will save yourself a weekend. First, update your motherboard BIOS to the newest release — mandatory on 13th/14th Gen Intel for the 0x12B microcode, and good hygiene everywhere, because half of what beginners diagnose as "undervolt instability" is a firmware bug the vendor already fixed. Second, record a baseline. Boot HWiNFO64, run Cinebench 2024, and write down the load Vcore, peak package power, peak temperature, and the score. You cannot claim a victory you never measured. Third, know your architecture, because it dictates the tool: Skylake through Raptor Lake use the BIOS offset or ThrottleStop; Arrow Lake uses XTU 10 or BIOS; AMD Zen 3/4/5 use Curve Optimizer. Here is what a baseline capture should look like.

Baseline capture (HWiNFO64, Cinebench 2024 load)
------------------------------------------------
CPU Core Voltage (VID, load)     1.312 V
Core VIDs (max observed)         1.340 V
CPU Package Power (max)          241.0 W
CPU Package Temp (max)           98 C   (thermal throttle hit)
Cinebench 2024 Multi-Core        2038 pts
Cinebench 2024 Single-Core        141 pts

Ninety-eight degrees and a throttle flag. That is a chip leaving performance on the table to boil itself. It is also a perfect undervolting candidate, and everything below is the process of reclaiming that margin.

The 12-Step Method

This is the universal loop, written Intel-first because the negative offset is the cleanest illustration of the principle. AMD's Curve Optimizer follows in the next section and uses the same discipline with a different control. Do not skip steps and do not deepen two variables at once; the entire method depends on changing exactly one thing between tests so that when something breaks you know what broke it.

The FIVR, the Offset, and Adaptive Mode

On Intel, voltage is delivered through the Fully Integrated Voltage Regulator, and you influence it with an offset — a value added to or subtracted from every point on the VID curve. This matters enormously versus a fixed override voltage. An override pins the core to one voltage at all times, which destroys idle power savings and can actually run hotter at light load. An offset, ideally in Adaptive-plus-Offset mode, scales the whole curve down while preserving the chip's ability to drop to near-nothing at idle. Always undervolt with an offset. The override is a tool for a different job.

The 12 Steps

1. Update the BIOS and microcode. Rationale: on Raptor Lake this installs 0x12B and stops the chip's voltage demand from drifting; everywhere else it removes firmware bugs that masquerade as instability.
2. Capture the baseline in HWiNFO. Rationale: without a before, the after is just a vibe. Record Vcore, power, temperature, and Cinebench score.
3. Enter the UEFI — tap Del, F2, or F12 during POST — and locate CPU Core Voltage. Rationale: these keys have opened firmware for over a decade; the mode you want is Adaptive or Offset, never a fixed override.
4. Set a negative offset of -0.050 V (-50 mV). Rationale: this is the conservative opening probe, stable on the overwhelming majority of chips, and it confirms the whole pipeline works before you get ambitious.
5. Save and boot to Windows. Rationale: a failed boot here is the cheapest possible failure, fully reversible by clearing CMOS; better to fail now than after you have stacked five changes.
6. Confirm in HWiNFO that Vcore actually fell versus baseline. Rationale: some boards silently ignore offsets or apply them only in certain modes; trust the sensor reading, not the fact that the menu accepted your number.
7. Run Cinebench 2024 multi-core and compare score and peak temperature to baseline. Rationale: an equal-or-higher score at a lower temperature is proof the undervolt is doing exactly what it should.
8. Run OCCT CPU, small data set, AVX2, for 30–60 minutes, watching for errors and WHEA events. Rationale: this is where marginal offsets die quietly; Cinebench passing is necessary but nowhere near sufficient.
9. If stable, deepen the offset to -0.075 V and repeat steps 6 through 8. Rationale: you refine toward the wall in controlled increments rather than leaping and guessing.
10. When you hit a crash, reboot, or WHEA storm, back off one step. Rationale: the last value that survived a full test, minus a small safety margin, is your daily driver — not the value that crashed.
11. Add a separate cache/ring offset only after the core is settled, and go shallower there — start around -0.050 V. Rationale: core and cache are different voltage domains; over-cutting the cache produces silent, hard-to-trace instability that looks like a bad core offset.
12. Lock the final value in BIOS and run a long soak — two-plus hours of OCCT, a Cinebench loop, and a day of real workloads. Rationale: "stable for ten minutes" is not stable; cold-boot failures and edge-case crashes surface only over time.

Here is the BIOS state you are aiming for at the end of a successful Intel run.

[ Intel UEFI - CPU Voltage / Overclocking ]
CPU Core Voltage Mode        : Adaptive + Offset
Offset Prefix                : -  (negative)
Core Voltage Offset          : 0.075 V
CPU Cache (Ring) Offset Mode : Adaptive + Offset
Cache Voltage Offset         : 0.050 V
CPU Load-Line Calibration    : Level 4 (medium)
Long Duration Power Limit    : 253 W (stock)
Save and Exit (F10)

Arrow Lake and the No-Auto-Boot Rule

Two Intel-specific traps. First, if your CPU Core Voltage menu is greyed out entirely, that is not a bug — it is Plundervolt. The 2019 vulnerability (CVE-2019-11157) let attackers fault Intel SGX by undervolting, so Intel shipped microcode that lets vendors disable the undervolting interface, and many Skylake-and-later business, locked, or SGX-dependent systems ship with it disabled. No amount of tapping Del brings it back short of an older BIOS you probably should not run. Second, on Arrow Lake, XTU version 10 deliberately removed the ability to auto-apply an undervolt on boot; you can tune interactively, but for a daily-driver setting you must commit the value in BIOS. Intel is steadily pushing undervolting toward being a firmware-only affair, and the utilities are following.

AMD Ryzen: Curve Optimizer

AMD does not want you subtracting a flat voltage. It wants you reshaping the per-core voltage-frequency curve through Curve Optimizer, and because Precision Boost Overdrive is opportunistic, the payoff is different: on Ryzen, a good undervolt usually manifests as higher boost clocks, because PBO spends the reclaimed thermal and electrical headroom on frequency. You undervolt to overclock, without ever typing a frequency.

Curve Optimizer Versus the Old Offset

Curve Optimizer shifts the VFT table by a number of "counts," positive or negative, per core or across all cores. You will see the count described as roughly 3–5 mV, and that rule of thumb is fine for intuition, but be precise about what it is not: SkatterBencher's testing is explicit that a count is not a fixed millivolt subtraction across the whole curve. The documented behavior is that a -30 count shifts the frequency above 5.6 GHz by more than 11% while at 4.7 GHz the same -30 shifts it only 8%. The impact is non-linear across the 600-to-6000 MHz range, which is precisely why single-core boost stability and all-core stability are different questions with different answers.

The All-Core Starting Probe

The conservative, beginner-safe path: enter the BIOS, enable Precision Boost Overdrive, set PBO limits to Motherboard, and open Curve Optimizer. Set it to All Core, Negative, magnitude 10 — the standard opening probe, comfortably stable on nearly every chip. Boot, and validate with the same loop as Intel: HWiNFO to confirm behavior, Cinebench 2024 for a quick sanity check, then OCCT. If it survives, increase the magnitude in increments of 5 — to -15, then -20 — retesting at each step. If your board offers Global C-state Control and you experience random freezes specifically at idle, disable it; certain Ryzen chips freeze on C-state transitions when undervolted, and this single toggle is the fix.

Per-CCD and Per-Core for Zen 4 and 5

The all-core number is set by your worst core, which is a waste of every other core's potential. Two refinements recover it. On dual-CCD parts — the Ryzen 9 7950X, 9950X and their kin — switch to Per CCD, because each chiplet has its own curve and bins differently; SkatterBencher's tested Zen 5 starting split is -25 on CCD0 and -20 on CCD1, the better die taking the deeper cut. Then, for the last mile, go Per Core: set every core to -30, boot, and run an all-core Prime95 load in HWiNFO while watching for cores that crash or clock-stretch, walking each unstable core back until the whole set holds. Expect real spread — four cores at -30 and two at -13 is an ordinary result, not a defective chip. The AMD-specific gotcha is that this per-core stability must be tested per-core, which is exactly what CoreCycler exists to do.

[ AMD UEFI - Precision Boost Overdrive ]
PBO Mode                     : Advanced
PBO Limits                   : Motherboard
Curve Optimizer              : Per Core (or Per CCD)
CCD0 Sign / Value            : Negative / 20
CCD1 Sign / Value            : Negative / 15
Max CPU Boost Clock Override : +100 MHz
PBO Scalar                   : 1X (Auto)
Global C-state Control       : Auto (Disable on idle freeze)
Save and Exit (F10)

Note the Max Boost Clock Override at a modest +100 MHz and the PBO Scalar left at 1X. Add boost override only after the undervolt is proven, in +25 to +50 MHz steps; raising the scalar just pours voltage and heat back in for marginal gains and accelerates the aging you undervolted to avoid. And run CoreCycler with a config like this before you call it done.

# CoreCycler config.ini (excerpt)
[General]
stressTestProgram   = PRIME95
runtimePerCore      = 6m
[Prime95]
mode                = SSE     ; SSE surfaces CO instability fastest
FFTSize             = Smallest
[Error]
stopOnError         = 1       ; halt on the first failing core

Laptops & ThrottleStop

Laptops are where undervolting stops being a nicety and becomes the point. A thin-and-light chassis cannot dissipate what a modern mobile chip requests, so the CPU throttles within seconds, the fans scream, and the boost clock you paid for evaporates. A correct undervolt buys sustained frequency, a quieter fan curve, and measurable battery life, all at once. It is the single highest-value tweak available to a laptop owner, which is why it is the first thing worth trying on any machine — including the value champions like the Legion 5i or the flagship Legion Pro 7i, whose cooling is good but never infinite.

Why Laptops Are the Best Candidates

The mechanism is the same square-law relationship, but the constraint is tighter, so the reward is larger. Where a desktop undervolt might shave ten degrees off a chip that was fine anyway, a laptop undervolt can move the chip from a hard thermal throttle to a clean sustained boost, turning a stuttering all-core workload into a smooth one. Because the tool for Intel laptops is a live utility rather than a firmware menu, you can also watch the effect in real time, adjusting while a stress test runs and seeing the temperature curve bend.

ThrottleStop, FIVR, and the .ini

For Intel mobile CPUs, ThrottleStop is the standard. Open the FIVR window, tick "Unlock Adjustable Voltage," and apply a negative offset to CPU Core — start around -50 mV and walk toward -100 to -125 mV, testing as you go — then apply a matching offset to CPU Cache, since on most Intel mobile parts the two are linked and a mismatch causes instability. On unlocked 10th-gen-or-newer HX and K mobile silicon, ThrottleStop 9.7 exposes full core and cache V/F tuning in that same window. The step everyone forgets: click "OK - Save voltages immediately" and enable saving to the ini, or your painstaking offset vanishes the moment the program exits. Then use Task Scheduler to launch ThrottleStop minimized at boot so the setting survives a restart. The relevant ini keys look like this.

; ThrottleStop.ini (relevant keys)
FIVR=1
UnlockAdjustableVoltage=1
OffsetVoltageCPUCore=-100.6      ; mV, Core domain
OffsetVoltageCPUCache=-100.6     ; mV, Cache/Ring domain
SaveVoltagesImmediately=1        ; survives program exit
Task=1                           ; launch at boot via Task Scheduler

The odd fractional value — -100.6 rather than a round -100 — is not a typo; the FIVR quantizes offsets to specific steps, and ThrottleStop reports the real applied value rather than the one you asked for. Believe the number it shows you.

When the BIOS Is Locked, and Plundervolt Locked It

Here is the laptop heartbreak. Apply your offset, check HWiNFO, and if the core voltage does not move a millivolt, your FIVR is locked. This is Plundervolt again: after the 2019 SGX disclosure, a great many laptops shipped microcode that disables the undervolting interface, and ThrottleStop's own FIVR window will report "Undervolt Protection" when that is the case. There is no software trick around a microcode lock; your only options are an older BIOS, if one exists and you accept the risk, or acceptance. AMD-based laptops are a separate world with no ThrottleStop-style FIVR access at all — tuning there runs through the Universal x86 Tuning Utility or the manufacturer's own tool, with considerably less reach. Check for the lock before you invest an evening.

Proving Stability

An undervolt that has not been proven stable is not an undervolt; it is a delayed crash with good manners. The failure mode that makes this hobby genuinely hazardous is not the blue screen — that is loud and obvious — but the silent one, where the chip computes a wrong answer, corrects it invisibly, and logs a machine-check error you never looked for. Testing is the discipline of forcing those errors to the surface where you can see them.

The Test Stack

Use layers, from fast to brutal. Cinebench 2024 multi-core is your thirty-second sanity check and a decent thermal probe, especially in its ten-minute stress mode. OCCT with the small data set and AVX2 enabled is the electrical stress that draws the stability boundary with granularity, and its built-in error detection is the feature that sets it apart. Prime95 Small FFTs is the maximum-heat, maximum-current final exam. On AMD, CoreCycler cycles single-worker Prime95 across each physical core to catch the single-core boost failures that all-core tests structurally cannot produce, because under an all-core load every core runs cooler and slower than it does when it boosts alone. And after all of that: real workloads. Games, compiles, video exports. Synthetic stability and real-world stability are correlated, not identical, and the gap between them is where smug undervolters get humbled.

Reading WHEA, the Silent Killer

Windows logs hardware machine-check errors under the WHEA-Logger source in Event Viewer, and a system throwing corrected WHEA errors is not stable, no matter how many benchmarks it passes. An undervolt that is a hair too aggressive frequently produces exactly this: no crash, no visible artifact, just a steady drip of corrected errors as the chip computes wrong and fixes it. Check for them explicitly.

Get-WinEvent -LogName System |
  Where-Object { $_.ProviderName -eq "Microsoft-Windows-WHEA-Logger" } |
  Select-Object TimeCreated, Id, LevelDisplayName -First 20

A clean system returns nothing. A marginally-undervolted one returns something like this, and it means back off.

TimeCreated            Id  LevelDisplayName
-----------            --  ----------------
7/16/2026 21:03:11     19  Error
7/16/2026 21:03:11     17  Warning
7/16/2026 20:58:44     19  Error

Event 19 is an uncorrected error and event 17 a corrected one; either during a stress run is a verdict. Add five to ten millivolts back and retest until the log stays empty across a full session.

How Long Is Long Enough

Gross instability shows up in the first five to ten minutes — if OCCT errors out that fast, your offset is nowhere near viable. Marginal instability needs an hour or two of continuous load to surface. But the real certification is a week of ordinary use, because the two failure modes that short tests miss are cold boots and sleep-wake transitions, both of which strain the voltage regulation in ways a warm, already-running stress test never does. The mantra, which SkatterBencher states plainly, is to measure, change one thing, and validate. A configuration you can trust through months of gaming and rendering without a single WHEA error is the only configuration worth locking in.

Five Pitfalls

Most failed undervolts fail for one of a small number of reasons, and all of them are avoidable if you know the shape of the mistake in advance. Here are the ones that will cost you a weekend, each with the fix.

The Big Five

Pitfall 1: Using a fixed override instead of an offset. A fixed voltage pins the core high at all times, killing idle downclocking and sometimes raising idle temperatures above stock. Fix: always use Adaptive or Offset mode on Intel, and Curve Optimizer on AMD, so the whole curve scales and idle savings survive.

Pitfall 2: Testing only under all-core load. An all-core test runs every core cool and slow; a core that passes it can still crash the instant it boosts to its single-threaded peak in a game. Fix: on AMD, run CoreCycler; on Intel, include lightly-threaded real workloads in your validation.

Pitfall 3: Ignoring WHEA. A benchmark-stable machine quietly logging corrected machine-check errors is corrupting computation you cannot see. Fix: monitor the WHEA-Logger source in Event Viewer after every change, and treat any error as a failure.

Pitfall 4: Cutting cache or SoC voltage too hard. The cache/ring on Intel and the SoC/memory-controller rails on AMD tolerate far less undervolt than the core, and over-cutting them produces instability that looks like a core problem and wastes hours of misdirected tuning. Fix: undervolt the core first and fully, then approach cache separately and shallowly, and leave SoC voltage alone unless you know exactly why you are touching it.

Pitfall 5: Undervolting Raptor Lake without the 0x12B microcode. A pre-0x12B 13th or 14th Gen chip has a drifting voltage demand, so any offset you dial in is measured against a moving reference and will appear mysteriously unstable over time. Fix: update the BIOS to a version containing 0x12B before you begin, full stop.

Why 'Stable' Lied to You

The sixth pitfall deserves its own note because it catches experienced tuners: the cold-boot bug. A deep offset can be perfectly stable once the machine is warm and running, yet fail to POST from a cold start, because silicon needs marginally more voltage when cold. If your system runs flawlessly all day but occasionally refuses to boot in the morning, you are not haunted — you are five to ten millivolts too aggressive for cold margin. Add that margin back. "Stable" measured only from a warm state is a half-truth.

The Warranty and the Degradation Question

A word from the department of law and lore, since people ask whether undervolting damages the chip. It does not, and the reasoning is instructive. Silicon degradation is driven by excess voltage and heat — it is precisely what destroyed those Raptor Lake chips, which requested too much voltage, not too little. Undervolting moves in the opposite direction: less voltage, less heat, less electromigration stress. If anything it extends the chip's life. The honest caveat is that vendors formally disclaim damage arising from operating outside specification, and undervolting is technically out-of-spec tuning, but the realistic risk from lowering voltage is instability, which is free to reverse by clearing CMOS, not physical harm. The thing that voids goodwill and kills silicon is the overvolt, and that is the exact behavior you are here to undo.

Troubleshooting Table

When something goes wrong — and on a first run something usually does — match the symptom to the cause and apply the fix. Every entry here is reversible; nothing in undervolting is permanent, which is the entire reason it is safe to experiment with.

Symptom to Cause to Fix

Clearing CMOS When You Go Too Far

The universal undo is the CMOS clear, and you should know how to do it before you need it. Most modern boards have a Clear CMOS button on the rear I/O; failing that, there is a two-pin jumper on the board labeled CLR_CMOS or CLRTC that you short with the system powered off; failing that, pull the coin-cell battery for a minute with the power supply switched off. Any of the three resets the BIOS to defaults, wiping your offset and letting the machine boot so you can try again more conservatively. Nothing you do in the voltage menu can survive a CMOS clear, which is precisely what makes the menu safe to experiment in.

Rolling Back XTU and ThrottleStop

Software tuners are even easier to undo. In Intel XTU, hit "Revert" or set the offset back to zero and apply; the change is immediate and non-persistent unless you committed it to BIOS. In ThrottleStop, clear the offset in the FIVR window, save, and if you added a Task Scheduler entry, disable it so the tool does not reapply the old value at the next boot. On AMD, Ryzen Master reverts to defaults with a single button, and any Curve Optimizer value set in BIOS clears with — you guessed it — a CMOS reset.

Advanced Tuning

Once the basic offset is stable and soaked, there is a second tier of control for people who enjoy the diminishing-returns end of the curve. None of it is necessary. All of it is optional. Approach it the same way: one variable at a time, validated before the next.

SMU Debug and Frequency Caps

On AMD, the SMU debug tooling lets you cap the all-core frequency deliberately, which is a legitimate undervolting strategy: by holding the all-core ceiling at a known value, you let the chip run that frequency at a much lower voltage than PBO would otherwise request while chasing its maximum boost. The tested all-core caps are 5.0 GHz for Zen 5 and 4.5 GHz for Zen 4 — set the ceiling there, then push the Curve Optimizer harder underneath it, and you trade a sliver of peak single-core boost for a large drop in all-core power and heat. It is the efficiency-tuner's move, ideal for a chip that spends its life under all-core render or compile loads.

Per-Core, Cache, and the Ring

The per-core Curve Optimizer pass described earlier is the single highest-yield advanced technique on AMD, because it lets your best cores go as deep as -30 while your worst sits at -13, extracting the full potential of each rather than shackling all of them to the weakest. On Intel, the analogous refinement is a separate cache/ring offset and, on hybrid parts, distinct offsets for P-cores and E-cores. The rule is uniform: the secondary domains — cache, ring, E-cores — tolerate less undervolt than the primary cores, so tune them last and shallower, and never assume a value that worked on the core will transfer.

Load-Line Calibration and the Overshoot Trap

Load-Line Calibration is the one advanced setting that regularly does more harm than good in undervolting hands. LLC counteracts the voltage droop that occurs under sudden load; crank it too high and you get transient overshoot, voltage spikes above your target that defeat the entire purpose and stress the chip. For undervolting you generally want a moderate LLC — something in the middle of your board's range — and you let the offset, not the load-line, do the work of lowering voltage. Chasing a deeper undervolt by loosening LLC trades a stable average voltage for a spiky one, and the spikes are exactly what you were trying to eliminate. Leave the AC/DC load-lines at their sane defaults unless you have a specific, measured reason and an oscilloscope-grade understanding of what you are changing.

The Working Configuration

Here are three complete, conservative, daily-driver configurations — one per platform — presented as known-good starting points. Read that phrase carefully: starting points. The silicon lottery means these values are where you begin your own testing loop, not numbers to paste in and trust blind. A chip that will not do -75 mV is not defective, and a chip that will happily do -120 mV is not special; both are just samples from the distribution. Validate everything against your own hardware with the test stack above.

Intel Desktop (Raptor / Arrow Lake, K-series)

[ Intel daily-driver undervolt - validate on your chip ]
BIOS microcode               : 0x12B or newer (mandatory 13/14th Gen)
CPU Core Voltage Mode        : Adaptive + Offset
Core Voltage Offset          : -0.075 V
CPU Cache (Ring) Offset      : -0.050 V
Load-Line Calibration        : Level 4 (moderate)
Power Limits (PL1 / PL2)     : Intel stock for the SKU
Validation                   : OCCT AVX2 2h + Cinebench 2024 loop + WHEA clean

AMD Ryzen (Zen 4 / Zen 5)

[ AMD daily-driver undervolt - validate on your chip ]
Precision Boost Overdrive    : Advanced, Limits = Motherboard
Curve Optimizer              : Per Core (fallback: Per CCD -20 / -15)
Typical per-core spread      : best cores -30, weakest -10 to -15
Max CPU Boost Clock Override : +100 MHz (add after undervolt is stable)
PBO Scalar                   : 1X (Auto)
Global C-state Control       : Auto (Disable only on idle freeze)
Validation                   : CoreCycler all cores + OCCT + WHEA clean

Intel Laptop (ThrottleStop.ini)

; ThrottleStop daily-driver - validate on your laptop
FIVR=1
UnlockAdjustableVoltage=1
OffsetVoltageCPUCore=-100.6      ; start -50, walk toward -125
OffsetVoltageCPUCache=-100.6     ; keep matched to Core
SaveVoltagesImmediately=1
Task=1                           ; Task Scheduler at boot
; If Core voltage will not move: FIVR is Plundervolt-locked.

That is the whole craft. Update the firmware, measure the baseline, probe conservatively, deepen in steps, test until the WHEA log stays empty, and lock in the last value that survived. It is the same procedure your predecessors ran in 2018 and the same one you will run in 2030, because the guardband the factory ships is not going anywhere and neither is your right to reclaim it. The only thing that changes year to year is the errata list and the version number on the tool. The dial is the same dial. Turn it carefully.