CPU Undervolting 2026: 12 Steps, 45 Min, Zero Perf Loss

Undervolting is the rarest thing in PC building: a tuning knob that costs you nothing. Turn it correctly and your CPU runs cooler, draws less power, spins your fans slower, and performs exactly as well as it did before — sometimes better. Turn it too far and it crashes, harmlessly, asking to be turned back. That is the entire risk profile. Compare it to overclocking, where every reward is paid for in heat, noise, and silicon you are quietly cooking.

This guide walks the whole thing end to end for both camps: AMD's Curve Optimizer on Zen 4 and Zen 5 via Ryzen Master and BIOS, and Intel's offset undervolting via ThrottleStop, XTU, and BIOS — plus the stability testing that separates a real undervolt from a screenshot that reboots an hour later. Budget an afternoon. The active work is maybe forty-five minutes; the rest is letting stress tests run while you do something more interesting.

Why Undervolt at All

The only free lunch left in the box

Every other knob in this hobby is a trade. Overclock and you pay in heat, noise, power draw, and the slow erosion of the part. Faster RAM costs stability margin. A bigger cooler costs money and case space. Undervolting is the one adjustment that runs the ledger backwards — you spend an afternoon and the machine hands you cooler temperatures, lower power, quieter fans, and performance that stays flat or ticks upward. In a hobby built entirely on compromise, that reads like cheating.

The reason the free lunch exists is binning. Every CPU AMD and Intel sell is fed a voltage set for the worst chip that passed the same test, plus a generous guard band on top so that a sealed OEM tower running in a hot room still validates years from now. Your specific chip, in your case, with your cooler, is almost certainly better than that worst-case sample. Undervolting is you reclaiming the guard band the factory left on the table because it had to ship to the lowest common denominator.

Cooler, quieter, and — yes — sometimes faster

Here is the part people refuse to believe until they watch it happen. Modern CPUs do not run at a fixed clock; they boost opportunistically until they hit a ceiling — temperature, power, or current — and then throttle back. Feed the chip less voltage and it produces less heat and draws less power at any given frequency, so it reaches those ceilings later, so it holds higher boost clocks for longer. On a Ryzen, a well-chosen Curve Optimizer undervolt routinely nets a small multi-core gain alongside a double-digit Celsius drop. On a thermally strangled laptop the effect is bigger still, because that machine spends most of its life throttling and you have just handed it headroom it never had.

None of this is exotic. It is the direct consequence of P = CV²f — power scales with the square of voltage — meeting a boost algorithm that treats thermal and power limits as the real speed limit. Undervolting does not make the silicon faster. It stops the silicon from wasting its allowance on heat. It was not always this simple, either: in the fixed-multiplier era you set one vcore in BIOS and lived with it. AMD's move to a per-core voltage/frequency curve, and Intel's on-die voltage regulator, turned tuning into a negotiation with an algorithm rather than a static number — which is exactly what makes a small, well-placed voltage cut pay off across the whole clock range.

Who should skip it

If your desktop chip already tops out in the low 70s °C under a real load and never throttles, undervolting will make it quieter and greener but will not move your frame rate — still worth doing, but set expectations. And if you are on a locked OEM laptop where the undervolt interface has been fused shut — a genuine thing, courtesy of a 2019 security paper and Intel's lawyers, covered in the Intel section — you may discover the whole exercise is a no-op. The same physics that make a 65 W desktop chip worth undervolting make a power-hungry flagship GPU worth it even more; our breakdown of the RTX 5090's 575-watt appetite is the graphics-side version of this same argument. Know which machine you are sitting at before you spend the evening arguing with it.

What Undervolting Actually Does

P = CV²f, and why the square is the whole story

Dynamic power in a CPU follows P = C × V² × f — capacitance times voltage squared times frequency — with a leakage term stacked on top that also rises with voltage and temperature. The single most important word in that equation is squared. Frequency scales power linearly; voltage scales it quadratically. Drop core voltage by 8% and, all else equal, you shed roughly 15% of the dynamic power and the heat that goes with it. That asymmetry is why undervolting is the highest-leverage thermal move available to you, and why shaving voltage beats shaving clocks every time.

Leakage is the quiet second reason. Hotter silicon leaks more current, which makes more heat, which raises temperature further — a small positive-feedback loop. Pull voltage and temperature down and you damp that loop, which is part of why the temperature drop from a good undervolt often looks larger than the raw power math predicts.

The voltage/frequency curve and Curve Optimizer counts

A modern Ryzen does not have a voltage. It has a voltage/frequency curve: for every frequency the boost algorithm might request, the chip has a factory-calibrated voltage — the VID — it will ask for. AMD's Curve Optimizer lets you shift that curve per core, up or down, in units it calls counts. Each count is a small step — on the order of 3 to 5 mV — and a negative count means, at this frequency, ask for a little less voltage than the factory table says. A −20 all-core offset tells every core to shave about twenty steps off its request across the entire curve.

Intel expresses the same idea as a flat offset in millivolts applied to the core, and separately to the cache/ring, through the FIVR on mobile parts or the board's voltage regulator on desktop. −80 mV means: whatever you were going to request, subtract eighty. The mechanism differs; the intent is identical — move the operating point below the factory-shipped voltage while keeping the frequency the algorithm wants.

The three ceilings your boost algorithm respects

Boost stops at whichever limit it hits first. On AMD those limits are PPT (package power, in watts), TDC and EDC (sustained and peak current, in amps), and the thermal limit, TjMax — 95 °C on most Ryzen desktop parts. On Intel they are PL1/PL2 (long- and short-window power) with the Tau window, ICCMax (current), and a roughly 100 °C thermal limit. Undervolting changes which ceiling you hit: a chip that was thermal-throttling at 95 °C now runs cooler and may hold its clocks against the power limit instead, or hold them outright. Understanding that is the difference between tuning deliberately and turning knobs until something stops crashing. It is also why two chips off the same wafer behave differently — the silicon lottery is a distribution, not a guarantee, and the offset your neighbour's chip swallows is a starting hypothesis for yours, never a fact about it.

Prerequisites: Hardware, Software, Versions

Hardware — and what disqualifies you

On the AMD side, any AM4 (Zen 2/3) or AM5 (Zen 4/5) desktop chip on a board whose BIOS exposes Precision Boost Overdrive qualifies, which in practice is nearly every B- and X-series board. On the Intel side, desktop undervolting via BIOS or XTU is happiest on an unlocked K/KF/X part with a Z-series chipset (Z690/Z790/Z890), because Intel gates the tuning interface behind the overclocking-capable platform. Laptops are their own world: many Ryzen and Intel notebooks can be undervolted, but a large slice of Intel laptops have the interface locked at the firmware level — the Plundervolt story, below. The honest first step is finding out whether your machine will accept an undervolt at all before you invest the evening.

Software, with the versions that matter

Use current tools. Undervolting interfaces track new silicon closely, and old builds miss sensors or apply offsets to the wrong rail. As of mid-2026 the toolchain is:

AMD desktop
  Ryzen Master        3.1.0.5085  (May 2026) - adds Curve Shaper for Zen 5
  HWiNFO              8.50        (Jul 2026) - sensor monitoring / logging

Intel
  Intel XTU 7.14.2.71 (Mar 2026)  - 14th Gen and older
  Intel XTU 10.0.1.45 (Mar 2026)  - Core Ultra Series 2 and newer
  ThrottleStop (latest)           - mobile FIVR undervolt, AC/Battery profiles

Stress / validation (either camp)
  Cinebench 2024      - 5-minute multi-core loop (quick screen)
  OCCT (2026 build)   - 1-2 hour CPU + variable-load stability
  Prime95             - optional Small FFTs torture (worst-case heat)

Grab Ryzen Master and HWiNFO from their own sites, XTU from Intel directly, and ThrottleStop per UltrabookReview's long-running guide — the links live in the sections that use each tool. One rule up front: do not install two live tuning tools that both write the voltage curve — Ryzen Master fighting a BIOS-applied curve, or XTU fighting ThrottleStop — and then expect sane results. Pick one tool to tune with and let the others read.

The thirty minutes of prep nobody does

Update your motherboard BIOS and AGESA first; voltage-curve behavior and stability at a given offset genuinely change between AGESA revisions, and half of the internet's mystery WHEA errors are stale firmware. Enable your memory's EXPO or XMP profile and confirm the system is already stable there, because chasing a memory fault and a voltage fault at the same time is how people lose weekends. Then take a baseline: run Cinebench 2024 at stock with HWiNFO logging and write down four numbers — peak CPU Core Voltage (the SVI2 TFN sensor on Ryzen), peak package power, peak temperature, and the effective clock. You cannot claim an improvement you never measured, and the baseline row is the only thing every later change gets compared against.

AMD: Ryzen Master and the 12-Step Curve

Set the table before you touch the curve

Undervolting a Ryzen is not a standalone setting; it rides on top of Precision Boost Overdrive. Before the first offset, boot into BIOS and confirm three things: PBO is set to Auto or Enabled — not force-disabled, because disabling it amputates the very boost behavior the undervolt is meant to feed — PBO power limits are set to Motherboard so the chip can use your board's rated delivery, and the boost-clock override sits at +0 MHz for now. Leave EXPO on. You are establishing a clean, known platform so that the only variable you change from here is voltage.

Then install Ryzen Master 3.1.0 and open it standalone — not launched from inside AMD Software: Adrenalin, which is a documented way to get wrong sensor readings. AMD's own Curve Optimizer documentation describes the controls; we will drive them in order.

The twelve steps

1. Baseline stock. Run a 5-minute Cinebench 2024 multi-core loop at stock with HWiNFO logging and record voltage, power, temperature, and effective clock. Rationale: every later claim is measured against this row; without it you are guessing.
2. Open Curve Optimizer. In Ryzen Master, switch to a manual profile and find the Curve Optimizer section. Rationale: this is the per-core V/F shift; everything else in the app is a distraction for this task.
3. Set Control = All Cores. Rationale: a single all-core offset is far easier to reason about and validate than eleven different per-core numbers on day one. Per-core comes later, as refinement.
4. Set the sign to Negative. Rationale: negative shifts the curve down — less voltage per frequency. Positive would raise voltage, the exact opposite of the goal.
5. Set Magnitude = 10. Rationale: −10 is the near-universal safe starting point, cited as stable on roughly 90% of hardware, so it is the offset least likely to waste your time on a false failure.
6. Apply and hold at −10. Run the 5-minute Cinebench loop again, watching for a crash, reboot, or WHEA warning. Rationale: prove the conservative rung before climbing; if −10 already misbehaves you have a platform problem, not a voltage one.
7. Deepen to −15. Reapply, retest the same 5-minute loop. Rationale: walk the curve down in small, attributable steps so a failure points at exactly one change.
8. Continue to −20, then −25. Retest at each stop. Rationale: −15 to −25 all-core is the Zen 5 sweet spot most chips settle into; you are finding where your sample lands inside that band.
9. On any crash, reboot, or WHEA error, back off by 5. If −25 fails, drop to −20 and re-validate. Rationale: the last offset that survives testing is your all-core ceiling; the +5 margin is your stability buffer, not a suggestion.
10. Light-load test. Sit at idle, open a browser, watch a video, leave it twenty minutes and watch for a random WHEA or a desktop BSOD. Rationale: undervolts fail at low load as often as high — the single-core boost bins are the most voltage-starved and stress tests rarely touch them.
11. Transcribe the stable offset into BIOS. Enter the same Curve Optimizer values in the board's PBO menu and remove the Ryzen Master profile. Rationale: software profiles can revert on driver updates or power-plan changes; BIOS applies pre-boot and is authoritative.
12. Long-soak and game. Run OCCT or a 1-to-2-hour Cinebench loop, then play something demanding for a real session, and finally re-check the WHEA log. Rationale: five minutes catches gross instability; only hours plus real use catch the subtle kind that eats a save file at the worst moment.

Reading the result in HWiNFO (expected output)

The sensor that matters is CPU Core Voltage (SVI2 TFN) — the actual delivered vcore after droop, not the VID request. Watch it, package power (CPU PPT), and per-core Tdie. A successful −20 on a 65 W-class chip looks something like this, though your silicon will differ and these numbers are illustrative, not a promise:

Sensor                         Stock        Curve Optimizer -20
-----------------------------  -----------  -------------------
CPU Core Voltage (SVI2 TFN)    1.30 V       1.23 V
CPU Package Power (PPT)        ~142 W       ~118 W
CPU Tdie (max, Cinebench)      ~88 C        ~74 C
Effective Clock (all-core)     ~5.15 GHz    ~5.20 GHz
Cinebench 2024 (multi)         baseline     equal or +1-2%

Cooler, lower power, clocks held or nudged up, score flat-to-better. That is the signature of a correct undervolt. If your score fell, you did not undervolt — you starved the boost and it downclocked to stay stable. Back off one step.

AMD in BIOS: PBO, EXPO, and the Manual-Voltage Sledgehammer

Why BIOS wins for the permanent config

Ryzen Master is the ideal place to find your numbers — apply, test, revert, repeat, all without a reboot. But it is the wrong place to keep them. Software offsets depend on a service running, survive at the mercy of Adrenalin updates and Windows power plans, and vanish if the app does not relaunch. The BIOS applies the curve before Windows exists, cannot be clobbered by a driver, and is the configuration your machine will actually live with. Tune in software; commit in firmware. Tom's Hardware's PBO and Curve Optimizer walkthrough maps the menu names across the major board vendors if yours are hiding under different labels.

The PBO and Curve Optimizer menu, translated

Board vendors bury these under different trees — ASUS under Ai Tweaker then Precision Boost Overdrive, MSI under OC then Advanced CPU Configuration, Gigabyte and ASRock similarly — but the settings are the same four. A committed, conservative AM5 config reads:

# AMD AM5 BIOS - PBO + Curve Optimizer (committed undervolt)
Precision Boost Overdrive .......... Advanced
  PBO Limits ....................... Motherboard
  PPT / TDC / EDC .................. Auto (from Motherboard)
  Precision Boost Overdrive Scalar . Auto (1X)
  Max CPU Boost Clock Override ..... +0 MHz     ; raise later, after undervolt is proven
  Curve Optimizer
    Curve Optimizer ................ All Cores
    Sign ........................... Negative
    Magnitude ...................... 20          ; your validated number, e.g. 15-25
Memory (DRAM)
  EXPO Profile ..................... Enabled (DDR5-6000 CL30 typical)

Save, reboot, and re-run the light-load and long-soak tests from the previous section, because a BIOS-applied curve can behave a hair differently from the same numbers in software, and you validate the thing you will actually run — not its close cousin in a utility.

The manual-voltage sledgehammer

There is a blunter option, and for some chips it is the right one. Instead of shifting the curve, you can set a fixed manual vcore — the guidance for AM5 parts like the Ryzen 5 9600X or Ryzen 7 7700X is 1050 mV (1.05 V), or 1100 mV (1.1 V) at the outside — and let the chip run every core at that flat voltage. Temperatures fall off a cliff. So does opportunistic single-core boost, because you have replaced the negotiated curve with a hard ceiling; the chip can no longer briefly push past 1.4 V to hit a peak single-thread bin. It becomes, in effect, a mild all-core undervolt-and-lock.

That trade is a poor deal on a chip you want to boost high, and an excellent one on a heat-limited 3D V-Cache part or a small-form-factor build where you value silence and low temperatures over the last three percent of single-thread. Know which you are building. If you do not have a specific reason to reach for the sledgehammer, use Curve Optimizer — it keeps the boost behavior intact and gives up far less to get the same headline temperatures.

Intel Laptops: ThrottleStop, FIVR, and Plundervolt

FIVR undervolting, step by step

On mobile Intel the tool is ThrottleStop, and the undervolt lives in the FIVR window — the Fully Integrated Voltage Regulator. UltrabookReview's ThrottleStop guide, updated January 2026, is the canonical reference; the short version:

1. Open FIVR, select CPU Core, tick Unlock Adjustable Voltage, and set the offset to −80 mV. Then select CPU Cache and set the same −80 mV. Rationale: core and cache share closely related behavior on many parts, so moving them together avoids a mismatch that reads as phantom instability.
2. Apply, then run a Cinebench 2024 loop. If stable, deepen in roughly 10-to-15 mV steps. Rationale: modern mobile chips commonly take −125 to −165 mV; older 3rd- and 4th-Gen parts rarely clear −40 to −50 mV, so let the chip's generation set your expectations.
3. On any crash or WHEA, back the offset off by 10-to-15 mV and re-validate. Rationale: same ceiling logic as AMD — the last stable value, plus a margin you do not touch.

# ThrottleStop - FIVR (mobile Intel undervolt)
FIVR Control
  CPU Core    : Unlock Adjustable Voltage = ON,  Offset = -125 mV
  CPU Cache   : Unlock Adjustable Voltage = ON,  Offset = -125 mV
  Intel GPU   : Offset = -50 mV        ; optional, conservative
  iGPU Unslice: Offset = -50 mV        ; keep equal to iGPU
  [x] Apply   [x] OK - Save voltages immediately

The Plundervolt lockout (the law)

If those boxes are greyed out or the offset simply does nothing, you have met Plundervolt. In 2019 researchers showed — CVE-2019-11157 — that the same undervolting interface could be abused to corrupt data inside Intel's SGX secure enclaves. Intel's response was microcode that lets firmware disable the undervolting mailbox entirely, and a great many OEMs, Dell prominent among them, shipped BIOSes with it fused shut. The result is a generation of laptops where ThrottleStop's FIVR undervolt is a no-op. Sometimes a hidden BIOS toggle — an overclocking or undervolt-protection lock — or a specific firmware version re-opens it; UltrabookReview keeps a running guide on re-enabling it. Often it is simply gone, and the mature response is to accept it rather than chase a setting that physically cannot apply. This is the rare corner of the hobby where the law, not the physics, decides whether you get your free lunch.

Two profiles: plugged in and on battery

Laptops live two lives, so give ThrottleStop two profiles. Set one as your AC profile — deeper undervolt, higher power limits, tuned for performance while plugged in — and enable a separate Battery profile with the same voltage offset but tighter power limits for heat and runtime. In Options, name the profiles and tick Battery Profile so ThrottleStop switches automatically when the charger comes and goes. Also tick Start Minimized and Minimize on Close, and add ThrottleStop to Task Scheduler at logon — because, unlike a BIOS curve, a software undervolt evaporates on every reboot until you make it persistent.

Intel Desktop: XTU and BIOS Offsets

Core and cache offset in XTU

On unlocked desktop Intel, Intel's own Extreme Tuning Utility is the least fussy entry point. Download XTU from Intel — version 7.14.2.71 for 14th Gen and older, 10.0.1.45 for Core Ultra Series 2 — and head to Advanced Tuning. Set Core Voltage Offset to −0.050 V (−50 mV) to start, mirror it on the cache offset, apply, and stress. Increment the negative offset in 10-to-25 mV steps until you provoke a crash, then step back one increment. Rationale: −50 mV is small enough that almost any chip swallows it, giving you a stable rung to climb from rather than a first-try blue screen that tells you nothing.

# Intel XTU - desktop undervolt (unlocked K + Z-series)
Core Voltage Offset ............ -0.075 V     ; start -0.050, walk down
Cache Voltage Offset ........... -0.075 V     ; keep equal to core, or 5-10 mV shallower
Turbo Boost Power Max (PL1) .... Intel default ; do NOT chase raised limits while undervolting
Turbo Boost Short Power (PL2) .. Intel default
Core Current Limit (ICCMax) .... Intel default
# Validate each step in XTU's built-in stress test or Cinebench 2024

Moving it into BIOS

As on AMD, XTU is where you find the number and BIOS is where you keep it. In firmware, set CPU Core Voltage mode to Offset, or Adaptive + Offset, enter your validated negative offset, and do the same for the cache/ring. Adaptive mode preserves the stock voltage/frequency behavior and merely subtracts your offset from it, which is what you want — a flat manual vcore on desktop Intel throws away the same idle and single-core efficiency that a fixed vcore costs on AM5.

The 13th- and 14th-Gen asterisk

Owners of 13th- and 14th-Gen K-series parts have a specific reason to care about voltage. Those chips were the subject of a well-documented elevated-voltage instability saga, which Intel addressed with microcode — the 0x12B update — and firm guidance to run the default power and current profiles rather than the unlimited board-vendor defaults. Undervolting sits comfortably inside that guidance; you are reducing voltage stress, not adding it. But two rules follow. Do not combine your undervolt with jacked-up power limits, and make sure your BIOS already carries the latest microcode before you conclude a crash is your offset's fault rather than a platform issue Intel has already patched.

Proving It: Stress Tests and WHEA Hunting

The five-minute screen and the two-hour truth

Validation happens in two tiers. The screen is a 5-minute Cinebench 2024 multi-core loop: it loads every core hard and fast and will expose a grossly over-aggressive offset within a minute or two. Passing it means only that you are not obviously broken. The truth is a 1-to-2-hour OCCT run — its variable and extreme modes deliberately shift load to catch transient faults — or a long Cinebench soak, followed by real gaming. OCCT is free for this and its logging makes a wandering fault easy to spot.

# A sane validation ladder (run in order, stop at first failure)
1. Cinebench 2024   multi-core, 5 min loop      -> gross all-core instability
2. Idle + browse    20-30 min                   -> light-load / single-core faults
3. OCCT CPU         1-2 h, Variable + Extreme    -> transient + sustained faults
4. Prime95          Small FFTs 30-60 min (opt.)  -> worst-case heat / power
5. Real game        1-2 h session                -> the workload you actually run
# After each: check the WHEA log (next block). Zero corrected errors = pass.

Light-load is where undervolts quietly die

The counterintuitive truth of undervolting: the dangerous instability usually is not at full tilt, it is at the bottom. All-core loads pull the voltage curve into a region the factory tuned conservatively and your offset barely dents. The lightly-threaded, high-frequency boost bins — one core sprinting to 5.5 GHz while the rest idle — are where the curve is thinnest and your negative offset bites hardest. That is why a chip can ace an hour of Prime95 and then blue-screen while you are reading email. Test idle and light browsing deliberately; it is not optional, and it is the single most-skipped step in every failed undervolt.

Hunting WHEA in Event Viewer

The definitive tell for an unstable undervolt is a WHEA — Windows Hardware Error Architecture — event. A corrected error, Event ID 19 from the WHEA-Logger, means the CPU hit a fault it could paper over this time; it is a yellow warning that your offset is too deep even if nothing crashed. Uncorrected errors, IDs 18 and 17, are the crashes themselves. Check the log from an elevated PowerShell:

Get-WinEvent -FilterHashtable @{
    LogName      = 'System'
    ProviderName = 'Microsoft-Windows-WHEA-Logger'
} -MaxEvents 20 |
    Format-Table TimeCreated, Id, LevelDisplayName, Message -AutoSize

# Expected on a STABLE undervolt:
#   No events were found that match the specified selection criteria.
#
# The sound of a too-deep offset (back off by one step):
#   TimeCreated           Id  LevelDisplayName  Message
#   -----------           --  ----------------  -------
#   2026-07-10 21:14:03    19  Warning           A corrected hardware error has occurred.
#   2026-07-10 20:52:41    19  Warning           A corrected hardware error has occurred.

A clean run over hours of mixed load and a WHEA log that stays empty is the entire definition of a finished undervolt. One recurring Event ID 19 and you back off by a step — five counts on Ryzen, 10-to-15 mV on Intel — and re-validate. Chasing the deepest number that passes Cinebench exactly once while the log quietly fills with corrected errors is how people convince themselves a broken config is stable.

Five Ways People Break Their Undervolt

Chasing the number instead of the result

The leaderboard brain wants the deepest offset — −30, −35, a screenshot to post. The correct target is the deepest offset that is invisibly stable for months, which is usually several counts shallower than the one that survives a single benchmark. A −20 that never throws a WHEA beats a −27 that corrects an error once a day and eats a save file next Tuesday. Tune for the config you will trust, not the one you will brag about.

Changing two things at once

Enabling EXPO and applying a Curve Optimizer offset in the same sitting, then hitting a crash, leaves you unable to say whether the memory or the voltage is at fault. Establish EXPO stability first, as its own known-good baseline, then introduce the undervolt as the single new variable. One change, one test, one conclusion. This is the whole method, and it is the rule most often broken.

Only testing all-core

It earns a second mention because it is the number-one cause of 'my stable undervolt keeps crashing at the desktop'. A five-minute Cinebench pass certifies your all-core behavior and nothing else. If you skip the light-load and idle testing, you have validated maybe sixty percent of the operating range and declared victory over all of it.

Trusting a stranger's number

'My 9700X does −30, so yours will' is the silicon-lottery fallacy in one sentence. Two chips off the same wafer accept different offsets; that is why the factory guard band exists. A forum number is a place to start testing, never a place to stop. The same lottery is why two of the identical GPUs in our RTX 5080 versus 4080 comparison can undervolt to visibly different points — binning is a distribution, not a promise stamped on the box.

Forgetting a software undervolt is not permanent

ThrottleStop and Ryzen Master apply at the OS level and are gone the instant the service does not start — a driver update, a Windows feature update, a power-plan change. People tune a beautiful undervolt, reboot a week later, and quietly lose it without noticing the temperatures creep back. Commit the final numbers to BIOS where you can, and where you cannot — locked laptops — pin ThrottleStop to Task Scheduler at logon and verify it actually loaded before you trust it.

Troubleshooting: Symptoms, Causes, Fixes

The table

Nearly every undervolting failure is one of a dozen shapes. Match the symptom, apply the fix, re-validate — one change at a time.

When the fix is 'back off', how far?

The universal retreat is one step: +5 Curve Optimizer counts on Ryzen (−25 to −20), or +10-to-15 mV on Intel. If a single step does not clear it, you either have a genuinely weak core that wants per-core treatment, below, or a non-voltage problem — memory, microcode, temperature — masquerading as one. Do not answer a −20 crash by jumping straight to −5 and calling it fixed; you will never learn where the real edge was, and you will leave most of your free performance on the table.

When it is not the undervolt at all

Set the offset back to zero and reproduce. If the crash, the WHEA, or the throttling survives a completely stock voltage curve, the undervolt was never the culprit — you are looking at unstable memory, missing microcode, a bad contact or paste job, or a background tool fighting for the voltage rail. Undervolting gets blamed for a lot of faults it merely revealed.

Advanced: Per-Core, Curve Shaper, and Handhelds

Per-core offsets and Zen 5's Curve Shaper

An all-core offset is limited by your worst core — the whole chip can only go as deep as the weakest silicon on it tolerates. Per-core Curve Optimizer breaks that ceiling: identify the strong cores, which your board or HWiNFO flags as the two preferred or best cores, and push them deeper, while relaxing the one or two cores that throw the first WHEA. It is more testing, but on a good sample it buys several extra counts of average offset. Zen 5 adds Curve Shaper, exposed for the first time in Ryzen Master 3.1.0: where Curve Optimizer applies one flat offset per core across the whole curve, Curve Shaper adds fifteen tuning points so you can set different offsets at different temperature and frequency bands — deeper in the cool, high-frequency light-load region where the chip has margin, shallower in the hot all-core region where it does not. It is the precise answer to the light-load instability problem that ends most naive undervolts.

Undervolt and overclock in the same session

Undervolting and PBO's boost override are not opposites; they compound. Once your negative curve is proven stable, you can add Max CPU Boost Clock Override +100 to +200 MHz and the freed thermal and power headroom often lets those higher bins actually stick, where at stock voltage they would have thermal-throttled instantly. Re-validate from scratch when you do — you have changed the variable again. This is the closest thing to genuinely free performance on the platform, and it is why 'undervolting' and 'overclocking' increasingly describe the same tuning session on Ryzen rather than two opposed camps.

Handhelds, APUs, and the Linux route

The chips that benefit most from undervolting are the ones with the least cooling — mobile APUs and gaming handhelds, which throttle constantly and live on a battery. On Ryzen laptop parts like the Ryzen 5 7640HS the move is enabling Custom P-states and taking P-state 0 from Auto to Custom to pull its voltage below the stock VID; a modest −10 Curve Optimizer offset is the same near-universal safe baseline it is on desktop. On Linux — where a lot of handheld tuning lives, as our Retroid coverage keeps finding — the tools are RyzenAdj for AMD APUs, which drives the same power and, on supported parts, curve limits from userspace, and intel-undervolt for Intel — the latter subject to the exact same Plundervolt lockout as ThrottleStop. Same physics, smaller battery, bigger payoff per watt saved.

The Complete Working Configuration

AM5 desktop, committed to BIOS

A full, conservative Zen 4/5 undervolt as it should live in firmware — validated numbers, boost intact, EXPO on. Substitute your own tested Curve Optimizer magnitude for the 20 shown here:

# === AMD AM5 - final BIOS undervolt (Zen 4/5) ===
Precision Boost Overdrive .............. Advanced
  PBO Limits ........................... Motherboard
  Precision Boost Overdrive Scalar ..... 1X (Auto)
  Max CPU Boost Clock Override ......... +0 MHz      ; +100/+200 only after re-validation
  Platform Thermal Throttle Limit ...... 90 C        ; optional temp cap for quiet builds
  Curve Optimizer ...................... All Cores
    Sign ............................... Negative
    Magnitude .......................... 20          ; YOUR validated value (typ. 15-25)
Memory
  EXPO ................................. Enabled (DDR5-6000 CL30)
# Validation on this exact config:
#   Cinebench 2024 5-min .... pass
#   OCCT 2 h Variable ....... pass
#   Idle/browse 30 min ...... pass
#   WHEA-Logger ............. 0 corrected errors over 1 week daily use

Intel laptop, ThrottleStop persistent

The mobile equivalent: two profiles, matched core and cache offset, made to survive reboots via Task Scheduler.

# === Intel mobile - final ThrottleStop config ===
Options
  Profile 1 = AC       -> selected as main profile
  Profile 2 = Battery  -> [x] Battery Profile (auto-switch on unplug)
  [x] Start Minimized   [x] Minimize on Close
FIVR (both profiles)
  CPU Core  : Unlock Adjustable Voltage = ON, Offset = -125 mV   ; modern chip; older = -40..-50
  CPU Cache : Unlock Adjustable Voltage = ON, Offset = -125 mV
  Intel GPU : Offset = -50 mV
  [x] OK - Save voltages immediately
Power limits
  AC      : PL1/PL2 to taste within cooling
  Battery : PL1/PL2 lower for heat + runtime
Persistence
  Task Scheduler -> run ThrottleStop.exe at logon, highest privileges
# Validation: Cinebench 2024 5-min + 1 h real load, WHEA log clean, both AC and battery.

The one-paragraph doctrine

Measure first, change one thing, validate, repeat — and define validated as an empty WHEA log across hours of mixed load and real use, not a single benchmark that happened not to crash. Start conservative (−10 on Ryzen, −50 mV on Intel), walk down in small attributable steps, and when something breaks, back off one step and stop. Commit the survivor to BIOS wherever the platform lets you. Do that and undervolting delivers the only genuinely free upgrade in the machine: a cooler, quieter, longer-lived CPU that runs exactly as fast as the day you bought it, or a hair faster — which, next to a flagship GPU that is thirty percent quicker for twice the money, is comfortably the best value-per-hour tuning you will do all year.