CPU Undervolting 2026: -155mV in 13 Steps, 45 Min

Your CPU is a coward, and the factory made it that way on purpose. Not in operation — in provisioning. The voltage your chip requests at any given frequency was never measured against your specific die. It was stamped on by a binning process that assumes you received the worst sample that still passed QA, running in a poorly ventilated case, on a hot afternoon, after the paste pumped out. The manufacturer pads the voltage so the bottom fraction of a percent of silicon does not generate warranty returns. You are almost certainly not the bottom fraction of a percent. Undervolting is the legally and electrically uneventful act of repossessing that padding.

This is a tutorial, not a sales pitch, so here is the honest framing up front: undervolting lowers the voltage your CPU consumes at a given clock speed, which drops heat and power without touching performance. A November 2025 owner of an Intel Core Ultra 7 265K pulled -155mV and reported 20°C cooler temperatures and 50W less package power on a humble air cooler. That is the upside. The downside is that an over-aggressive offset produces instability ranging from a clean blue screen to silent arithmetic corruption that quietly poisons your data while everything appears fine. We will spend most of our time making sure you land on the first side of that line. Plan on roughly 45 minutes for the initial offset and several hours — ideally overnight — for the stress validation that actually certifies it.

What Undervolting Actually Is

Before you touch a single register, understand the physics, because the physics is the entire reason this works and the entire reason it has limits. People who skip this part are the people who later post screenshots of WHEA errors asking why their "stable" undervolt corrupted a 40GB archive.

Voltage, Heat, and the Silicon Lottery

Dynamic power dissipation in a CMOS processor follows P = C × V² × f, where C is the switched capacitance, V is the core voltage, and f is the frequency. The term that matters is V². Power scales with the square of voltage but only linearly with frequency. Shave voltage by ten percent and, holding clocks constant, you drop dynamic power by roughly twenty percent. Less power in means less heat out, and since the frequency f never changed, your benchmark scores do not move. That is the whole trick. There is no free lunch elsewhere in the system, but here, for once, the vendor left a genuinely free lunch on the table and walked away.

The reason the lunch exists is the silicon lottery. Two CPUs with the same model number, cut from adjacent positions on the same wafer, can require meaningfully different voltages to hit the same boost clock. Manufacturing variation is physical reality. The vendor cannot ship a different voltage table to every die — it ships one table, validated so that even an unusually leaky, unusually thirsty sample stays stable. If you drew an average or good die, the factory voltage is wildly generous, and that generosity is what you are clawing back.

Why Your Chip Ships Over-Volted

The ArchWiki's 2025 undervolting guide frames undervolting precisely: lowering voltage to reduce heat and power without sacrificing performance. The padding the factory adds is called guard-band, and it covers worst-case silicon, worst-case temperature, worst-case aging over a multi-year warranty period, and worst-case workload transients. Every one of those is a margin you, the individual owner of one specific chip in one specific room, can measure empirically rather than assume pessimistically. The vendor optimizes for zero returns across millions of units. You optimize for one unit you can actually test.

A brief note for the warranty-anxious, because The Machine knows the law as well as the lore: in the United States, the Magnuson-Moss Warranty Act bars a manufacturer from voiding your warranty merely because you ran software that adjusts a documented, user-accessible voltage offset — they must show your modification actually caused the failure. Undervolting writes to the same offset registers Intel and AMD expose to their own tuning utilities. It is categorically different from delidding or shunt-modding. That said, OEM laptop vendors are a separate and pettier animal, and we will get to why some of them lock the door entirely.

What Undervolting Won't Do

Undervolting is not overclocking. If you want the opposite ritual — pushing clocks up and feeding more voltage to get there — that is a different and longer afternoon, and we have a separate 14-step GPU overclocking walkthrough for the silicon on your graphics card. Undervolting holds frequency constant. It will not raise your single-core score. What it can do indirectly is let a thermally or power-limited chip sustain its boost clock longer, because you have removed the heat that was forcing it to throttle. On a desktop with a 360mm radiator and a 250W power budget, an undervolt may show up purely as lower temperatures and a quieter fan curve. On a thin-and-light laptop slamming into a 28W package limit, the same undervolt can translate directly into higher sustained clocks, because the chip is no longer voltage-starved into throttling. Know which machine you have before you decide what "success" looks like.

Prerequisites & Versions

Gather everything first. Half the failed undervolts in the wild are not unstable silicon — they are someone discovering mid-process that their BIOS hid the relevant menu, or that a security mitigation silently neutered the offset they thought they applied.

Hardware Requirements

For Intel, you need a Haswell-generation chip or newer — that is 4th Gen Core and up — because that is where the offset-voltage interface the tools rely on became accessible. The current sweet spot is 13th and 14th Gen Core mobile and the Core Ultra desktop parts. Be warned that on many 12th/13th/14th Gen platforms the voltage offset is locked by default and only re-openable on an unlocked (K-series) CPU paired with a Z-series board, or via a specific BIOS toggle — more on that lock below. For AMD, an AM5 desktop Ryzen (the 9700X, 9900X3D, and siblings) gives you the BIOS Curve Optimizer; a Ryzen mobile chip (7640HS, 7640U, and relatives) gives you custom P-states. You also need adequate cooling already installed and seated — undervolting reduces thermal load but it is not a substitute for a heatsink, and you cannot validate stability on a chip that is thermal-throttling for unrelated reasons.

Software and Versions

On Windows/Intel, install ThrottleStop 9.7, released January 2025, which exposes undervolting through its Turbo FIVR Control (FIVR) menu. Pair it with HWiNFO for monitoring core VID, package power, and — critically — the WHEA error counter. For stress validation you want Prime95 / mprime and ideally a Linpack-based tool, both of which run numerical correctness checks rather than merely heating the die. On AMD AM5 you need a recent BIOS — check that your board's AGESA is current, because Curve Optimizer behavior and per-core support improved materially across 2024-2026 AGESA revisions. On Ryzen laptops, the Universal x86 Tuning Utility (UXTU) is the pragmatic path to the same SMU that the P-state editor touches. On Linux, install intel-undervolt for Intel and amdctl for AMD, plus the msr kernel module and turbostat for measurement.

Back Up Your Sanity First

Create a Windows restore point or note your exact BIOS defaults before you change anything. Write down your current idle and load temperatures and package power from HWiNFO so you have a baseline to compare against — "it feels cooler" is not data. If your machine is a laptop you rely on for work, do this on a day you can afford a few forced reboots. And read the next section's warning about locked offsets before you spend twenty minutes applying a -100mV offset that the platform is silently discarding. The single most common "my undervolt does nothing" report is not subtle silicon behavior; it is a security mitigation eating the write.

Intel: ThrottleStop in 13 Steps

This is the headline path, because ThrottleStop on a supported Intel chip is the fastest route from stock to a validated offset. The 2026 ThrottleStop guidance recommends a conservative starting point of -80mV applied to both CPU Core and CPU Cache, with well-behaved modern dies often landing stable somewhere between -125mV and -165mV. We will start safe and walk down.

Steps 1–5: Setup and the FIVR Menu

Each step has a reason. Do not skip the reasons.

1. Install and launch ThrottleStop 9.7 as administrator. Rationale: writing to the voltage MSRs requires elevation; without it the offset fields are inert.
2. Open HWiNFO in Sensors-only mode alongside it. Rationale: you need live readings of Core VID, CPU Package Power, package temperature, and the WHEA error counter to know whether an offset took effect and whether it is generating correctable errors.
3. Record your stock baseline. Run a quick load and note VID, watts, and temps. Rationale: every later claim of improvement is measured against this number, not against memory.
4. Click the FIVR button to open Turbo FIVR Control. Rationale: this is where Intel's Fully Integrated Voltage Regulator offsets live; everything voltage-related happens in this dialog.
5. Select "CPU Core" in the voltage plane list and tick "Unlock Adjustable Voltage." Rationale: until this box is ticked, the offset slider is decorative. If ticking it produces no effect — the slider snaps back, or HWiNFO shows no VID change later — your platform has the offset locked, and you should jump to the troubleshooting table before going further.

Steps 6–9: Applying the Offset

6. Set the CPU Core "Offset Voltage" to -80.1 mV. Rationale: -80mV is the conservative anchor from the 2026 guide — large enough to measure, small enough that almost no functioning die rejects it. ThrottleStop's slider lands on values like -80.1 due to the underlying voltage step granularity; that is normal.
7. Select "CPU Cache" and set the same -80.1 mV offset. Rationale: on the majority of Intel dies the Core and Cache planes are most stable when matched. Mismatching them is an advanced technique, not a starting move.
8. Leave Intel GPU and System Agent at 0 mV for now. Rationale: the iGPU and SA offsets buy little on a desktop with a discrete card and complicate fault isolation. Add them later, separately, if at all.
9. Click "Apply" and confirm in HWiNFO that Core VID dropped. Rationale: this is your proof the write landed. If VID does not move, the offset is being rejected — stop and diagnose rather than stacking more offset on a dead write.

[ThrottleStop FIVR — Turbo FIVR Control]
FID Multiplier      : default (do not touch)
CPU Core Voltage    : Adjustable / Offset
  Offset Voltage    : -80.1 mV       # conservative anchor
CPU Cache Voltage   : Adjustable / Offset
  Offset Voltage    : -80.1 mV       # match Core on most dies
Intel GPU           : Offset  0.0 mV # leave for now
System Agent (SA)   : Offset  0.0 mV # rarely worth the trouble
[x] Unlock Adjustable Voltage
[x] OK - Save Voltages Immediately

Steps 10–13: Locking It In

10. Tick "OK - Save Voltages Immediately" in the FIVR dialog. Rationale: the 2026 guidance is blunt about this — save voltages the instant they apply, so a later crash does not force you to re-type everything from memory. This is the single most-skipped step and the source of the most rage.
11. Run a short stability pass (10-15 min of Prime95 Small FFTs) at -80mV. Rationale: confirm the safe anchor is solid before you push deeper, so any instability at a lower voltage is unambiguously attributable to the step you just took.
12. If clean, walk the offset down in 5-10 mV increments — toward -125mV, then -145mV, then -165mV — re-testing at each rung. Rationale: stable modern chips live in the -125mV to -165mV band, but you find your die's floor by stepping, not leaping. The November 2025 265K owner who got -155mV / 20°C cooler / 50W less did it by stepping to that number, not by guessing it.
13. Create two ThrottleStop profiles and set ThrottleStop to start with Windows. Rationale: the 2026 guide recommends at least two profiles — one for AC, one for battery — and offsets do not persist across reboot unless ThrottleStop launches at login (via its own option plus a Task Scheduler entry running as admin). An undervolt that evaporates every restart is not an undervolt; it is a daily chore.

Here is roughly what a successful -80mV-to-validated run looks like in your monitoring, measured against the baseline you recorded in Step 3:

HWiNFO — before → after (CPU Core -125 mV, Cache -125 mV)
Core VID (avg)        1.060 V   →  0.965 V
CPU Package Power     81.0 W    →  63.4 W
CPU Package Temp      93 °C     →  76 °C
Max Boost Clock       unchanged (5.4 GHz)
WHEA Errors           0
Rounding/MCE errors   0

AMD AM5: Curve Optimizer

AMD desktop owners do not use a Windows tray app for this; they use the BIOS, and the mechanism is elegant. Instead of a flat voltage offset, AMD's Curve Optimizer shifts the entire voltage/frequency curve, so the undervolt adapts across the operating range rather than subtracting a fixed number everywhere.

Entering the Curve Optimizer

Reboot into BIOS and navigate to AMD Overclocking, then Precision Boost Overdrive (PBO). Set PBO to Advanced so the sub-options appear, then open Curve Optimizer. The relevant June 2026 walkthrough sets the Curve Optimizer to Negative sign with All Core mode, using counts like -20 — which the walkthrough pegs at roughly -200mV effective — for meaningful temperature drops, and counts as deep as -25 (≈ -250mV) for sustained CPU-intensive scenarios. Treat the millivolt equivalences as approximate: the per-count voltage shift varies by die and by where you sit on the V/F curve, which is precisely why the curve-based approach behaves better than a flat offset.

Advanced / AMD Overclocking / Precision Boost Overdrive
  PBO Mode ................... Advanced
  PBO Limits ................. Motherboard
  Curve Optimizer
    Curve Optimizer Mode ..... All Core
    Magnitude / Sign ......... Negative
    All Core Count ........... 20        # ≈ -200 mV effective
  Max CPU Boost Clock Override . 0 MHz    # leave at stock until validated

All-Core vs Per-Core

All-Core applies one count to every core and is the correct starting mode — it is simpler to reason about and faster to validate. Per-core lets you assign a different (usually deeper) count to your stronger cores and a shallower count to the weak one that crashes first, squeezing out more total undervolt. Per-core is an optimization you earn after All-Core is rock solid, not a place to begin. If you go per-core, your weakest core sets the practical floor for the All-Core value you could have used, and finding it is a process of elimination that costs hours.

Reading the Counts

After saving and rebooting, the proof is the same as everywhere else: lower load temperatures at unchanged or higher sustained clocks. AM5 chips, especially the X3D parts with their thermally sensitive stacked cache, often show the most dramatic improvement here — the cache die hates heat, and pulling voltage pulls heat directly off it. If you are also feeding that CPU to a power-hungry graphics card, note that the RTX 5090's draw means every watt you reclaim on the CPU side is a watt your PSU and case thermals do not have to fight twice. Validate the AM5 undervolt with the same stress regimen we cover below; the Curve Optimizer can pass a quick test and fail a long one, exactly like an Intel offset.

Ryzen Laptops: P-States

Ryzen mobile is the fiddliest of the four paths because the mechanism is buried and the tooling varies by OEM. The research-grade method adjusts custom Core P-states, lowering the P0 voltage below stock VID. In practice most people reach the same SMU through UXTU rather than raw BIOS P-state editing, because few laptop BIOSes expose the P-state table at all.

Finding the P-State Menu

If your laptop BIOS does expose custom Core P-states (some do, most do not), you are looking for the P0 entry — P0 is the highest-performance state, defined by an FID (frequency), DID (divider), and VID (voltage ID). Undervolting means dropping the P0 VID below its stock value. If the menu is absent — the common case — install UXTU, which talks to the same SMU and exposes a Curve Optimizer-style control on mobile silicon. Either way the conservative starting point is the same modest number.

# Universal x86 Tuning Utility (UXTU) — mobile Ryzen
Preset .................. Custom
Curve Optimizer (All) ... -10        # safe start (7640HS / 7640U)
Tctl Temp Limit ......... 90 °C
PPT / Fast / Slow ....... stock for now

# Or, if your BIOS exposes raw Core P-States:
P0  FID=0xA8  DID=0x08  VID=0x20      # lower VID index = lower voltage

Editing P0 VID

The research is explicit on values: a safe starting point is -10, with -20 achievable on lucky hardware. Lower the P0 VID — or set the UXTU count — to -10 first and validate fully before considering -20. Mobile silicon has far less thermal headroom and far more aggressive boost behavior than desktop, so an offset that looks stable in a five-minute test can fail when the chip sustains boost against a low package limit on battery. Be conservative; mobile crashes are more disruptive because they often happen mid-task on a machine you are actively using.

Unlocking Power Limits

Once the undervolt is genuinely stable, the research notes you can unlock power limits — using Prime95, Cinebench, or 3DMark as the load — to convert your new thermal headroom into sustained performance. This is the laptop-specific payoff: a stable undervolt plus a raised PPT can mean the chip holds a higher clock instead of throttling. But cooling must be managed carefully, because raising the power limit raises heat, partially spending the headroom the undervolt just bought you. If you are shopping for a chassis that tolerates this kind of tuning in the first place, the thermal design matters more than the sticker clock — our 2026 gaming laptop teardown gets into which coolers actually have margin to give. As Engadget observed in its 2025 coverage of laptop thermal tuning, undervolting is among the most effective non-invasive methods to reduce heat in thin-and-light machines, frequently dropping temperatures 10-15°C with no hardware change at all.

Linux: intel-undervolt & amdctl

Linux has no ThrottleStop, but it has two purpose-built command-line tools that hit the same registers. The ArchWiki guide documents both. They are less hand-holding than the Windows path and more honest about what they touch.

intel-undervolt for Haswell and Newer

intel-undervolt is a tool built on MSR and MCHBAR registers that undervolts Haswell and newer Intel CPUs. Its config lives at /etc/intel-undervolt.conf, and a typical setup the ArchWiki cites pulls CPU Cache to -100mV. Negative values are the undervolt; positive would be an overvolt you almost never want. The tool also exposes package power limits and a Tjmax offset, but start with just Core and Cache.

# /etc/intel-undervolt.conf
# Negative values = undervolt (millivolts)
undervolt 0 'CPU'          -100
undervolt 1 'GPU'           -50
undervolt 2 'CPU Cache'    -100
undervolt 3 'System Agent'    0
undervolt 4 'Analog I/O'      0

# Optional: power limits (W) and tjmax offset — leave commented to start
# power package 45 35
# tjoffset -5

Apply and read back to confirm the write landed:

$ sudo modprobe msr
$ sudo intel-undervolt apply
CPU         : -100.00 mV
GPU         :  -50.00 mV
CPU Cache   : -100.00 mV
System Agent:    0.00 mV
Applied successfully.

$ sudo intel-undervolt read     # verify values stuck

amdctl for K10 and Newer

amdctl is the AMD counterpart, enabling undervolting on K10 and newer AMD CPUs through kernel-level MSR access. It reads and writes the per-P-state voltage/frequency registers directly. One honest caveat the README and community both stress: amdctl's most reliable territory is pre-Zen and early-Zen; modern Zen 4/5 desktop parts are better served by Curve Optimizer in BIOS because their boost is governed by the SMU rather than the classic P-state VIDs. Check amdctl -h for your build's exact flag forms — they have shifted across versions.

# Read current P-states (root + msr module required)
$ sudo modprobe msr
$ sudo amdctl -m

# Lower P0 core voltage VID to undervolt (representative syntax)
$ sudo amdctl -p0 -v36        # set P-state 0 to a lower VID index

# Verify the new voltage
$ sudo amdctl -g -c0
P0: 4400 MHz @ 1.225 V   (was 1.300 V)

Making It Persist

Neither tool's changes survive reboot or, importantly, resume-from-suspend unless you make them. intel-undervolt ships a systemd service; enable it, and add a resume hook so a wake from sleep does not silently restore stock voltage. The SMU and MSRs reset on resume on many platforms, which is why a laptop "loses" its undervolt after closing the lid.

# Enable at boot
$ sudo systemctl enable --now intel-undervolt.service

# /etc/systemd/system/intel-undervolt-resume.service — reapply on wake
[Unit]
Description=Reapply undervolt after resume
After=suspend.target hibernate.target

[Service]
Type=oneshot
ExecStart=/usr/bin/intel-undervolt apply

[Install]
WantedBy=suspend.target hibernate.target

Confirm the result with turbostat, which gives you watts, MHz, temperature, and VID in one place — the Linux equivalent of the HWiNFO readout:

$ sudo turbostat --interval 5 --quiet
Busy%   Bzy_MHz   PkgWatt   PkgTmp   CorWatt   VID
99.1    4992      62.3      74        58.9     0.978
# Baseline before undervolt: PkgWatt 81.0, PkgTmp 93, VID 1.060

Stress Testing & Validation

An unvalidated undervolt is a rumor. The entire difference between "I undervolted" and "I have a stable undervolt" is the testing, and the testing must be the right kind.

Why mprime and Linpack

The ArchWiki guide strongly recommends mprime or Linpack specifically because both run numerical correctness checks — they do arithmetic with known answers and verify the results. This matters more than it sounds. A merely-hot stress test can heat your die without ever catching the failure mode that should terrify you: a marginally-undervolted CPU that does not crash but computes wrong answers. That is silent data corruption — checksums that pass when they should fail, a compile that produces a subtly broken binary, an archive that extracts with one flipped bit. mprime's Small FFTs and Linpack's residual checks are designed to flag exactly this. A test that only watches for a blue screen will happily certify a die that is quietly lying to you.

A Validation Protocol

Run mprime Small FFTs — maximum heat and the tightest numerical check — for at least 60 minutes per offset rung, and run it overnight before you trust an offset for real work. The ArchWiki notes the empirical guardrails plainly: for Intel, dropping CPU Core and Cache by 100-200mV is usually stable, while exceeding 200mV risks crashes or simply produces no further benefit, and crashes above -200mV are common. Stay inside that envelope. If you are stress-testing a small-form-factor emulation box — the kind of mini PC people undervolt specifically to run a wall of RetroArch cores silently in a living room — remember that a passively-cooled or low-airflow chassis changes what "stable under load" means, because the thermal ceiling arrives sooner.

# mprime (Prime95 for Linux) — Small FFTs = max heat + numerical check
$ ./mprime -t
Choose:  16) Torture Test  →  Small FFTs
Run time: 60 min minimum per rung; overnight to certify

# A FAILED undervolt — note this is NOT a crash:
[Worker #3] FATAL ERROR: Rounding was 0.5, expected less than 0.4
[Worker #3] Hardware failure detected, consult stress.txt
[Worker #3] Worker stopped.

Reading the Results

Three outcomes. Clean run to completion with zero worker failures and zero WHEA/MCE entries: that rung is a candidate, and you may step deeper. A worker stops with a rounding or correctness error: the offset is unstable — back off 5-10mV and retest. A clean run but the system feels off, with correctable machine-check events appearing in your logs: back off anyway, because correctable errors are the canary. The test "failing" is the test working — it caught instability on the bench instead of in your data three weeks from now.

Common Pitfalls

Five ways this goes wrong, and how to keep them from going wrong for you. Each of these is a real, recurring failure pattern, not a hypothetical.

The Plundervolt Lock and the Silent Discard

Here is the lore that explains the single most baffling failure. In 2019, researchers demonstrated Plundervolt (CVE-2019-11157), an attack that used the very undervolting interface we are discussing to corrupt computations inside Intel SGX secure enclaves. Intel's mitigation, delivered by microcode, was to let firmware disable the voltage-offset interface entirely. OEMs — especially laptop vendors — frequently ship with that interface locked. The symptom: you apply -100mV in ThrottleStop, it appears to accept it, and nothing changes. No temperature drop, no VID movement, no error. Fix: confirm Core VID actually moves in HWiNFO after Apply. If it does not, look for a BIOS option such as "Overclocking Lock," "CFG Lock," or "Undervolt Protection" and, where present, disable it. On fully locked OEM laptops there may be no legitimate fix short of a vendor BIOS update that re-exposes the toggle. Do not stack more offset on a write that is being silently discarded.

Forgetting to Save Before Testing

The 2026 ThrottleStop guidance hammers this for a reason: save voltages immediately after setting them, before you stress test. If you test an aggressive offset that crashes hard, the machine reboots, and an unsaved offset is gone — but so is your record of exactly what you tried. Fix: tick "OK - Save Voltages Immediately," and keep a plain-text log of each rung and its result. The crash should cost you the offset, not your memory of the experiment.

One Profile for Every Power State

A single profile that is stable plugged in can misbehave on battery, where the chip's boost and power behavior shift. Fix: follow the 2026 guidance and build at least two profiles — one for AC, one for battery — and let ThrottleStop switch automatically. On Linux, mirror this with separate apply-on-AC and apply-on-battery hooks.

Trusting a Five-Minute Test

An offset that survives five minutes is not validated; it is merely not-immediately-fatal. Many marginal undervolts fail only after the die soaks at temperature or hits a specific FFT size. Fix: 60 minutes minimum per rung, overnight before you trust it with anything you cannot afford to recompute.

Chasing the Last Millivolt

The difference between -145mV and -155mV is often a couple of degrees and a real increase in crash risk. People chase the leaderboard number and end up running an offset that is stable 95% of the time, which is the worst possible place to be. Fix: when you find your floor, step back up one rung and live there. A daily-stable -135mV beats a mostly-stable -155mV every single day. The BSOD-recovery procedure below codifies exactly this retreat.

Troubleshooting Table

Match your symptom to the row. Most undervolting problems are one of these ten, and most resolve by backing off, not by pushing harder.

The BSOD-Recovery Procedure

If a deep offset causes a blue screen, do not abandon undervolting — retreat methodically. The Reddit ThrottleStop guidance gives a concrete recipe: if a -155mV undervolt BSODs, drop to -145mV, and if it still falls over, lower P-cache and E-cache by 10mV each until stable. The principle generalizes to every platform here: reduce magnitude in 10mV steps, and when Core and Cache are separable, walk the Cache down before the Core, because cache instability tends to surface first and at lighter loads.

When to Walk It Away Entirely

If your platform turns out to be hard-locked (the FIVR row), or every offset beyond a trivial -20mV produces correctable errors, the honest answer is that your specific die or your specific OEM did not leave usable margin. That is a legitimate outcome, not a failure of method. Pocketing a safe, dull -50mV is a fine result; insisting on a number you saw on a forum is how stable machines become flaky ones.

Advanced Tips

Once you have a validated baseline, these are the refinements that separate a competent undervolt from a tuned one. None of them are starting moves.

Per-Core and Asymmetric Offsets

On both Intel (via per-core offset where supported) and AMD (per-core Curve Optimizer), your cores are not equal. The strongest cores tolerate deeper offsets; your single weakest core sets the All-Core floor. By assigning each core its own count — deeper on the strong, shallower on the weak — you can lower average voltage further than any single All-Core value would allow. The cost is hours of elimination testing to find which core fails first, and a more fragile configuration to reason about. Do this only after All-Core is certified, and keep notes, because a year from now you will not remember why core 6 is at -18 and the rest are at -25.

Temperature-Aware and Workload-Aware Profiles

The two-profile minimum (AC and battery) is the floor, not the ceiling. ThrottleStop and UXTU both support multiple profiles you can bind to conditions. A common advanced setup: an aggressive offset for sustained productivity loads where the die soaks and you want maximum thermal relief, and a slightly conservative offset for gaming, where bursty, mixed loads and transient voltage droop expose instability that a steady all-core test never triggers. Validate each profile separately — a profile is only as trustworthy as its own overnight run.

Pairing Undervolt With Power Limits

The deepest gains come from combining a stable undervolt with deliberate power-limit tuning, in the right order. On a laptop, the research path is undervolt-first, then raise PPT to convert the reclaimed thermal headroom into sustained clocks. On a desktop you might instead lower the package power limit, riding the undervolt to hold the same clocks at a far lower wattage and a near-silent fan curve. Either way, change one variable at a time and re-validate. Stacking an aggressive undervolt and an aggressive power-limit change simultaneously means that when it crashes — and it will — you have no idea which knob did it. Discipline here is the difference between an afternoon and a week.

The Complete Config

Here is a complete, validated reference configuration across the platforms, representing values that the research supports as commonly stable. These are starting templates you confirm against your own silicon, not guarantees — the silicon lottery means your floor is yours alone. If you want the same destination in a tighter, condensed format, our companion 125mV-in-12-steps quick reference covers the short version. And if all this voltage tuning is in service of a quieter emulation handheld rather than a desktop, the thermal math scales down too — see how a tuned chip behaves in our Retroid Pocket 6 review, where every reclaimed watt is battery life.

ThrottleStop — Validated Intel Profile

[ThrottleStop 9.7 — FIVR, profile: "AC - Tuned"]
CPU Core  Offset  : -125.0 mV     # validated floor for this die; -80 if unsure
CPU Cache Offset  : -125.0 mV     # matched to Core
Intel GPU Offset  :    0.0 mV
System Agent      :    0.0 mV
[x] Unlock Adjustable Voltage
[x] OK - Save Voltages Immediately
Start with Windows: yes (Task Scheduler, admin)
Profiles          : "AC - Tuned", "Battery - Conservative"
Validation        : 8 h Prime95 Small FFTs, 0 WHEA, 0 rounding errors

intel-undervolt.conf — Validated Linux Profile

# /etc/intel-undervolt.conf  (validated, Haswell+)
undervolt 0 'CPU'          -125
undervolt 1 'GPU'           -50
undervolt 2 'CPU Cache'    -125
undervolt 3 'System Agent'    0
undervolt 4 'Analog I/O'      0
# power package 45 35   # uncomment only after separate validation

# Persist:  systemctl enable --now intel-undervolt.service
# Resume:   intel-undervolt-resume.service (see Linux section)
# Verify:   sudo turbostat --interval 5 --quiet

AMD Curve Optimizer — Validated AM5 Summary

AMD Overclocking / PBO  (validated, AM5)
  PBO Mode ............ Advanced
  Curve Optimizer ..... All Core, Negative
  All Core Count ...... 20        # ≈ -200 mV; step toward 25 only if certified
  Max Boost Override .. 0 MHz     # raise later, separately
Mobile Ryzen (UXTU):  Curve Optimizer All = -10 (safe) / -20 (lucky die)
Validation:  overnight Prime95 Small FFTs, watch for worker errors

That is the whole method: understand the V² physics, confirm the offset interface is actually open, start conservative, step down with numerical-checking stress tests at every rung, save the instant it applies, and retreat one rung from your floor for daily use. Do that and you land where the November 2025 Core Ultra 7 265K owner landed — meaningfully cooler, meaningfully lower power, identical clocks — without ever opening the case. The factory left the margin sitting there. Go take it back, then stop one rung short of greedy.