Undervolt Your CPU 2026: 12 Steps, 45 Min, -200mV

Undervolting is the only free lunch in computing that is also a trap. The premise is genuinely sound: modern CPUs ship with a voltage curve padded for the worst silicon to leave the factory, which means yours — being merely average rather than catastrophic — runs hotter and louder than it has any reason to. Strip the padding away and you get lower temperatures, lower power draw, less fan noise, and frequently higher sustained clocks, because the chip stops slamming into thermal and power limits. The ArchWiki defines undervolting precisely this way — reducing voltage to lower energy consumption and temperature while maintaining performance, provided the result is stable.

That last clause is where the trap lives. "Provided the result is stable" is doing the work of an entire engineering discipline. An unstable undervolt does not announce itself with a polite warning. It corrupts a frame, drops a calculation, blue-screens you mid-save, or — worst of all — produces silently wrong data for weeks before you notice. This guide treats stability as the product and the voltage drop as the means, which is the opposite of how most tutorials are written.

We will cover four distinct platforms, because undervolting in 2026 is not one procedure. It is motherboard-vendor-specific on desktop Intel, fundamentally different on AMD AM5, and a separate animal again on locked-down laptops. If your interest in low temperatures comes from running emulators on a handheld rather than a tower, the same physics apply — quieter, cooler chips run a full RetroArch core stack for longer on the same battery. The principles transfer; only the knobs change.

Prerequisites: Hardware and Software

Do not skip this section to get to the exciting part. The exciting part is also the part that destroys an unsaved afternoon of work, and every prerequisite here exists because someone learned it the expensive way.

Hardware requirements

You need an unlocked or at least voltage-adjustable platform. On desktop, that means a current Intel Core Ultra chip on LGA 1851 or an AMD Ryzen on AM5, paired with a motherboard whose BIOS exposes voltage controls — virtually all Z-series and X-series boards do, most B-series boards do, and the cheapest H/A-series boards may not. On laptops, you need either a manufacturer that has not fused off voltage control (increasingly rare) or a mobile chip old enough to predate the lockdown, which we address in the ThrottleStop section. You also need adequate cooling already installed; undervolting reduces heat but cannot rescue a CPU strangled by a bad mount or a clogged heatsink.

Software requirements

The software depends on your route. For Linux Intel work you want a current kernel (6.x), the intel-undervolt utility from its GitHub repository, and a build toolchain to compile it if your distribution lacks a package. For Windows mobile undervolting you want the latest ThrottleStop release, following the canonical ThrottleStop guide. For stress testing you want mprime (the Linux build of Prime95, from mersenne.org) or linpack, which the ArchWiki recommends explicitly for validating an undervolt. For desktop BIOS work the software is the BIOS itself plus a hardware monitor like HWiNFO to watch the results.

Backups, expectations, and the recovery plan

Before you change a single value: know how to clear CMOS on your board. Find the jumper or the button now, while the machine still boots, not at 2 a.m. when it does not. Save any work and close anything you care about. Set the expectation correctly — a 2026 Intel Core Ultra tuning guide claims you can extract 20–30% more performance from BIOS changes plus better temperatures, but treat that as a creator's enthusiastic estimate rather than a benchmark you are entitled to. Cooler and quieter is the reliable outcome. Faster is a frequent bonus, not a guarantee. The same discipline applies whether you are tuning a CPU or overclocking a GPU safely over a few hours — small steps, test, repeat.

The 12-Step Process at a Glance

Every platform below is a specialization of one workflow. Read these twelve steps once so the platform sections feel like instances of a pattern rather than twelve unrelated rituals. Each step carries its rationale, because a step you do not understand is a step you will skip.

1. Step 1 — Update firmware and tools. Run the latest stable BIOS and the latest tuning utility. Rationale: vendors fix voltage-table bugs in BIOS updates; an old BIOS may apply your offset to the wrong rail or ignore it entirely.
2. Step 2 — Record a clean baseline. Note idle and load temperatures, clocks, package power, and a benchmark score at stock. Rationale: you cannot tell whether an undervolt helped if you never measured the before.
3. Step 3 — Enable XMP/EXPO first. Apply your memory profile and confirm it boots and passes a quick test. Rationale: the 2026 Intel guide treats memory tuning as a prerequisite; doing it before the undervolt means a later crash points at voltage, not RAM.
4. Step 4 — Choose offset over fixed voltage. Pick an adaptive/offset mode rather than a flat fixed voltage where the platform allows it. Rationale: offset preserves the chip's ability to scale voltage with load; fixed voltage often raises idle draw and removes the safety the curve provides.
5. Step 5 — Apply a conservative first cut. Start small: roughly -80 mV on a laptop core, or a modest negative Curve Optimizer value on Ryzen. Rationale: a conservative starting undervolt that boots is worth more than an aggressive one that does not.
6. Step 6 — Match core and cache. On Intel, set CPU Core and CPU Cache to the same offset. Rationale: the ThrottleStop guide notes these two rails are typically electrically linked; a mismatch causes instability that looks random.
7. Step 7 — Boot and reach the desktop. Confirm a clean boot to a usable OS before anything else. Rationale: if it will not boot, no amount of stress testing matters; reduce the offset and retry.
8. Step 8 — Idle-test for ten minutes. Sit at the desktop and watch for instant crashes or WHEA errors. Rationale: low-load conditions, counterintuitively, expose many undervolt failures because the chip runs at high voltage-sensitive boost clocks while nearly idle.
9. Step 9 — Run a sustained stress test. Hit it with mprime or linpack for at least 30–60 minutes. Rationale: heat and current change the voltage a chip actually needs; a load test is the only honest stability check.
10. Step 10 — Step deeper, one notch at a time. If stable, increase the undervolt by a small increment and repeat the test. Rationale: the stability cliff is sharp; you want to find the edge and then back off, not leap over it.
11. Step 11 — Back off and add margin. Once you find the failing point, return to the last stable value and add a safety margin on top. Rationale: a value that barely survives a stress test will fail on a hot day or an unusual workload.
12. Step 12 — Make it persistent. Commit the value — save the BIOS profile, enable a service like intel-undervolt.service, or set ThrottleStop to start with Windows. Rationale: an undervolt that evaporates on reboot is a hobby, not a configuration.

Intel Core Ultra in the BIOS

Desktop Intel undervolting in 2026 happens in the BIOS, and the BIOS is a different country depending on who made your board. This is the single biggest source of confusion, so we name names.

Getting into the right menu by vendor

Per the 2026 Intel Core Ultra guide, the entry point differs by manufacturer. On ASUS boards you typically begin by switching into Advanced Mode (the default EZ Mode hides the tuning menus). On MSI boards the equivalent is reached by pressing F7 to flip out of the simplified view. Gigabyte and ASRock have their own toggles. There is no universal path — anyone who tells you "go to the OC menu" without asking your board vendor is reciting from memory and will send you in circles. Find the advanced view for your board first, then proceed.

Enable XMP, then set core input voltage

The guide is emphatic about order: enable XMP first, confirm the system is stable with the memory profile applied, and only then touch the CPU. With memory settled, locate the core input voltage control. Choose either offset or fixed mode — offset is the recommended default for the reasons in Step 4 — and enter your value in the BIOS's expected units, which on these platforms means dividing your intended millivolt figure by 1,000. A 100 mV reduction is entered as 0.100 in the offset field, not as 100. Getting this units conversion wrong is how people accidentally request a one-volt change and brick a boot.

Ratios, DLVR, and power limits

Voltage is not the only lever. If your goal is purely lower temperatures rather than maximum clocks, the guide suggests dropping CPU ratios about 100 MHz below stock — a tiny clock concession that buys a disproportionate thermal saving because voltage scales with the square of frequency-driven demand. The guide also points to a CPU DLVR value of 1.15 as a tuning reference, and directs attention to the power-limit controls: Long Duration Power Limit, Short Duration Power Limit, and Current Limit. These power limits define how long and how hard the chip is allowed to boost; pairing a sensible undervolt with deliberate power limits is what produces the "cooler and not slower" result everyone is actually after. A rough planning sheet looks like this:

; Intel Core Ultra (LGA 1851) BIOS planning sheet
; Enter offsets as volts: 100 mV -> 0.100

Memory profile        : XMP enabled (verify stable first)
Core input voltage    : OFFSET mode
Core voltage offset   : -0.100 V   (start here, -100 mV)
CPU ratio             : stock minus 1 (approx -100 MHz, optional)
CPU DLVR              : 1.15
Long Duration PL (PL1): set to TDP target
Short Duration PL(PL2): set to boost ceiling
Current Limit (IccMax): keep at safe board default initially

Apply, save, reboot, and proceed straight to stability testing. Do not change five things and a power limit in one sitting — if it crashes you will not know which one did it.

Intel on Linux: intel-undervolt

If you run Linux, you do not have to live in the BIOS. The intel-undervolt utility writes offsets directly through the CPU's MSR interface, which means you can tune, test, and revert without rebooting — a far faster loop than the save-reboot-pray cycle of BIOS work.

Install and apply an offset

Install from your distribution's package or build it from the repository. The configuration file lives at /etc/intel-undervolt.conf. The two rails that matter most are CPU and CPU Cache, and per the ThrottleStop guide's logic they should carry the same value. A typical first configuration:

# /etc/intel-undervolt.conf
# Values are millivolts, negative = undervolt

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 (watts) and time window (seconds)
# power package 35 45
# tjoffset -10

Apply and read back the result with the command-line tool:

$ sudo intel-undervolt apply
$ sudo intel-undervolt read

CPU         : -100.00 mV
GPU         : -50.00 mV
CPU Cache   : -100.00 mV
System Agent: 0.00 mV
Analog I/O  : 0.00 mV

That read-back is your expected output — the numbers you wrote should reflect back unchanged. If read shows 0.00 where you set -100, your kernel or firmware is blocking MSR writes and you have a different problem to solve before tuning.

How far to push, per ArchWiki

The ArchWiki gives the single most useful pair of numbers in this entire field: for Intel CPU and CPU Cache, reductions of 100 to 200 mV are usually stable, while going beyond 200 mV tends to cause crashes or simply has no effect because the platform clamps it. That is your search space. Start at -100, validate, step toward -150, validate, and approach -200 with suspicion. Anything past -200 that appears to "work" is probably being ignored, and anything that crashes there is telling you the truth about your silicon.

Make it persistent with a service

An MSR write does not survive a reboot. The ArchWiki documents the fix: once you have a stable configuration, enable intel-undervolt.service so the offset is reapplied automatically at boot (and, with the loop variant, reapplied periodically in case the firmware resets it after a power event).

$ sudo systemctl enable --now intel-undervolt.service
$ systemctl status intel-undervolt.service

* intel-undervolt.service - Intel Undervolting Service
     Loaded: loaded (/usr/lib/systemd/system/intel-undervolt.service; enabled)
     Active: active (exited) since ...
      Tasks: 0 (limit: 9216)

AMD AM5 and Curve Optimizer

AMD does not undervolt the way Intel does. There is no single "core voltage offset" you dial in; instead the chip runs a per-core voltage-frequency curve, and you shift that curve down with Curve Optimizer. The mental model is different and worth internalizing: you are not lowering a number, you are telling each core it can hit its target frequency at a lower voltage than AMD's conservative factory assumption.

Voltage targets for AM5

A 2026 AM5 tuning guide gives concrete starting points depending on intent. For performance-oriented tuning, a practical starting voltage is around 1.200 V. If your priority is cool and quiet rather than maximum clocks, aim lower — in the 1.05 V to 1.10 V range. These are starting points, not destinations; the silicon decides where you actually land. Choose your intent before you touch anything, because tuning for temperature and tuning for performance pull the curve in different directions and you cannot maximize both at once.

Using Curve Optimizer correctly

The procedure: set Curve Optimizer to a negative value (negative means undervolt — a positive value would add voltage), boot into Windows, stress-test, and adjust if the system crashes. This is the standard undervolting workflow expressed in AMD's idiom. The catch with Ryzen is that low-load instability is common, so do not declare victory after a stress test alone — idle and lightly-threaded crashes are the signature failure of an over-aggressive AM5 curve, exactly as Step 8 warns.

; AMD AM5 Ryzen — BIOS (PBO / Curve Optimizer)
; Negative = undervolt. Per-core or all-core.

Precision Boost Overdrive : Advanced
Curve Optimizer mode      : All Core (start here)
Curve Optimizer sign      : Negative
Curve Optimizer magnitude : 15        ; conservative start

; Intent-based VID guidance from 2026 AM5 guide:
;   performance focus : ~1.200 V
;   thermal focus     : 1.05 - 1.10 V

EXPO / DOCP             : enabled (set BEFORE curve tuning)

Per-core versus all-core

Start with an all-core negative value because it is simpler and faster to validate. Once you find an all-core figure that survives testing, the advanced move is per-core tuning: AMD's best cores (the ones the chip boosts highest) often tolerate a larger negative offset than the weak cores, so an all-core value is always limited by your worst core. Per-core tuning recovers that headroom but multiplies your testing time. For most people the all-core result is the sane stopping point.

Ryzen Laptops and Mobile P-States

Laptops are where undervolting earns its keep — thermally constrained, fan-noisy, battery-bound — and also where vendors have done the most to lock you out. On mobile AM5 and Ryzen laptop CPUs, the door that remains open is custom core P-states.

Entering custom P-states

Per a 2026 Ryzen laptop guide, entering custom core P-states gives you access to frequency and voltage controls, and undervolting happens by lowering the VID values within those states. A P-state is a defined operating point — a frequency paired with a voltage — and by editing the VID downward you tell the chip to hold that frequency at less voltage. This is undervolting by another name, exposed through whatever utility your platform supports for P-state editing.

The -10 rule and how far to go

The guide gives a refreshingly honest distribution: a P-state offset of -10 works for roughly 90% of hardware, while -20 may work only on some systems. Its recommendation for the average person is to target around 15 or 20 as an upper bound rather than chasing the absolute edge. Translation: -10 is the value you should expect to keep; anything beyond it is a per-unit lottery. Do not read someone else's "-20 works great" as a target — their chip is not your chip, and the 90% figure exists precisely because the tail is unreliable.

; Mobile Ryzen — custom core P-state (VID undervolt)
; Lower VID = undervolt. Offset notation varies by tool.

Pstate0 (boost)  : FID = stock,  VID offset = -10   ; ~90% of chips OK
Pstate1          : FID = stock,  VID offset = -10
Pstate2          : FID = stock,  VID offset = -10

; -20 = aggressive, system-dependent, validate hard
; average target ceiling: 15 to 20

Why laptops reward this most

A desktop with good cooling barely notices a 100 mV cut beyond a quieter fan. A laptop notices everything: lower VID means lower heat, which means the chip throttles less, which means it holds boost clocks longer, which on a thermally-limited machine can read as a real performance gain even though you only ever reduced voltage. If you are shopping rather than tuning, this is also why thermal design dominates our 2026 gaming laptop rankings — a chip you can undervolt into a quiet, cool envelope beats a faster chip pinned at its thermal ceiling.

ThrottleStop: The Windows Mobile Route

For Intel mobile chips on Windows — especially the generations before voltage control got fused off — ThrottleStop remains the standard tool, and the canonical ThrottleStop guide is the reference everyone cites.

The FIVR window and starting values

Undervolting in ThrottleStop happens in the FIVR (Fully Integrated Voltage Regulator) window. The guide's conservative starting point is -80 mV for CPU Core, and — critically — it states that CPU Core and CPU Cache should generally use the same value. This is the same core-equals-cache rule from Step 6, and it exists because on these chips the two rails are physically tied; setting them differently produces instability that masquerades as randomness and will waste hours of your debugging.

How much your chip will take

The guide's experience-based ranges are the most useful calibration you will find. Modern mobile CPUs often undervolt well in the -125 to -165 mV range. Older 3rd- and 4th-generation Core chips, by contrast, may only tolerate 40–50 mV before they fall over. So the age of your laptop sets your expectations: a recent machine has real headroom, a decade-old one has almost none, and trying to force a modern offset onto ancient silicon just buys you crashes.

ThrottleStop FIVR — expected stable example (modern mobile)

FIVR Control
  CPU Core     Offset Voltage : -125.0 mV
  CPU Cache    Offset Voltage : -125.0 mV   (match Core)
  Intel GPU    Offset Voltage :  -50.0 mV
  Uncore       Offset Voltage :  -50.0 mV

Apply -> uncheck "OK - Save voltages immediately" until tested
Start with -80 mV, step toward -125/-165 only if stable

Making it stick

Once validated, set ThrottleStop to start with Windows (Task Scheduler, run at logon, highest privileges) and check "OK - Save voltages immediately" so the profile reapplies on boot. Without this, your undervolt vanishes the moment the laptop sleeps deeply or restarts — the Windows equivalent of forgetting to enable intel-undervolt.service.

Stability Testing and Validation

This is the section that separates a configuration from a liability. An undervolt is a hypothesis — "this chip is stable at this voltage" — and a hypothesis you have not tested is just optimism.

The right tools

The ArchWiki names two: mprime and linpack. mprime (Prime95's Linux build) hammers the CPU with floating-point work that surfaces voltage instability fast; linpack stresses the chip differently and catches failures mprime misses. Running both is better than running either twice. On Linux you can also lean on stress-ng for a quick smoke test before committing to a long run.

# Quick smoke test (Linux), then the real thing
$ stress-ng --cpu $(nproc) --timeout 120s --metrics-brief

# mprime / Prime95 — small FFTs stress the core hardest
$ ./mprime -t          # torture test, run 30-60+ minutes

# Watch temps and errors in another terminal
$ watch -n2 'sensors | grep -E "Package|Core"'

What "stable" actually means

Stable is not "booted once." Stable is: survives a 30–60 minute stress test without errors, survives idle and light-load sitting (where many undervolts die), survives your actual workload — the game, the compile, the emulator — and survives a cold boot the next morning. On Linux, watch dmesg and journalctl for MCE or WHEA entries; on Windows, watch the Event Viewer for WHEA-Logger warnings. A machine logging corrected machine-check errors is not stable even if it has not crashed yet — it is telling you the voltage is too low and the hardware is papering over it. Treat the first WHEA warning as a failure and back off.

The expected good result

A successful run looks boring, which is the point. mprime completes its torture test with zero errors. Package temperature is meaningfully lower than your stock baseline at the same load. Sustained clocks are equal or higher because the chip is no longer throttling. Package power is down by a measurable margin. If all four are true and the machine cold-boots clean the next day, you have a configuration, not a gamble.

Common Pitfalls and Fixes

Five mistakes account for the overwhelming majority of failed undervolts. None of them are exotic.

Voltage and units mistakes

Pitfall 1 — the units error. Entering 100 where the BIOS wants 0.100 requests a thousandfold larger change. Fix: on Intel Core Ultra BIOS, divide your millivolt target by 1,000 before entering it; -100 mV is -0.100 V. Pitfall 2 — mismatched core and cache. Setting CPU Core to one value and CPU Cache to another, on platforms where they are linked, causes instability that looks random. Fix: set them to the same value, per the ThrottleStop guide.

Process and ordering mistakes

Pitfall 3 — skipping XMP/EXPO. Tuning the CPU while memory is at stock means a later memory profile can destabilize a "validated" undervolt, and you will blame the wrong subsystem. Fix: enable and verify XMP/EXPO first, exactly as Step 3 dictates. Pitfall 4 — changing five things at once. Offset plus ratio plus DLVR plus two power limits in one reboot means a crash tells you nothing. Fix: change one variable per test cycle. Pitfall 5 — declaring victory on a stress test alone. Especially on Ryzen, idle and light-load crashes are common even when the chip survives full load. Fix: test idle, light, full, and your real workload before committing.

Persistence and recovery mistakes

Bonus pitfall — the disappearing undervolt. An MSR write or a ThrottleStop session that is never made persistent silently reverts on reboot or sleep, and you spend a week confused about why temperatures crept back up. Fix: enable intel-undervolt.service on Linux, or run ThrottleStop at logon with voltages saved on Windows. And always — always — know your CMOS-clear procedure before you start, so a no-boot offset costs you two minutes, not an evening.

Troubleshooting Table

When something goes wrong, it usually goes wrong in one of these recognizable ways. Match the symptom, apply the fix, retest.

Symptom	Likely Cause	Fix
System won't POST after offset	Offset too aggressive for boot voltage	Clear CMOS, restart at a smaller offset (e.g. -50 mV)
Boots, then crashes at idle/desktop	High boost clock at low load needs more voltage	Reduce offset magnitude; idle is voltage-sensitive (Step 8)
Crashes only under full load	Undervolt fails at thermal/current peak	Back off one notch; add margin per Step 11
WHEA / MCE errors in logs, no crash	Silent instability — voltage too low	Treat as failure; raise voltage until logs are clean
intel-undervolt read shows 0.00 mV	MSR writes blocked by kernel/firmware	Check Secure Boot / kernel lockdown; enable MSR access
Undervolt gone after reboot	Not made persistent	Enable intel-undervolt.service or ThrottleStop-at-logon (Step 12)
No temperature change at all	Offset beyond -200 mV being clamped, or ignored	Stay within 100–200 mV per ArchWiki; verify it applied
Ryzen unstable despite small CO value	One weak core limits the all-core offset	Switch to per-core Curve Optimizer; relax the weak core
Old laptop crashes at -125 mV	3rd/4th-gen Core tolerates far less	Drop to 40–50 mV range per ThrottleStop guide
Memory errors after XMP + undervolt	SoC/SA rail starved by aggressive undervolt	Leave System Agent at 0; undervolt core/cache only

Advanced Tips

Once the basics are stable, there is a second tier of refinement that separates a working undervolt from an optimized one.

Pair undervolting with power limits

The real magic of the Intel Core Ultra approach is combining a voltage offset with deliberate Long Duration and Short Duration Power Limits plus a sensible Current Limit. Undervolting lets the chip do the same work at lower wattage; power limits then let you decide whether to spend that savings on lower temperatures, on holding boost longer, or on a quieter fan curve. The 100 MHz-below-stock ratio trick is the most efficient single lever for pure thermal reduction, because demand scales steeply with frequency.

Per-core tuning and the silicon lottery

On AMD, per-core Curve Optimizer recovers the headroom that an all-core value wastes on your weakest core. The same principle explains why no two chips undervolt identically and why borrowed settings are a starting hint, not a target — the -10 P-state rule works for 90% of hardware precisely because the other 10% lost the lottery. Tune your chip, not someone's screenshot.

Undervolting for battery and handhelds

The emulation crowd has quietly turned undervolting into a battery-life technique. A cooler, lower-power chip in a handheld runs a stack of RetroArch cores longer per charge, and the official libretro documentation is worth reading alongside any power-tuning work, because core choice and run-ahead settings interact with how hard the CPU is driven. If you are weighing a device specifically for tunable thermals and endurance, our Retroid Pocket 6 verdict covers exactly that trade-off.

Complete Working Configuration

Here is a complete, conservative, validated-shape configuration for each route. These are starting points calibrated to the numbers in the research, not maximums — apply, then test and tighten per the 12-step process. None of these will be optimal for your exact silicon, and that is correct; they are designed to boot, survive testing, and give you a safe base to step deeper from.

Linux Intel — intel-undervolt.conf

# /etc/intel-undervolt.conf  — conservative, ArchWiki-aligned
# Stable range for CPU/Cache is 100-200 mV; start at 100.

undervolt 0 'CPU'         -100
undervolt 1 'GPU'          -50
undervolt 2 'CPU Cache'   -100   # match CPU
undervolt 3 'System Agent'   0   # leave at 0 to protect memory
undervolt 4 'Analog I/O'     0

# Apply once, validate, then deepen toward -150.
# Persist with: systemctl enable --now intel-undervolt.service

Desktop Intel BIOS and AMD AM5 — planning values

; ---- Intel Core Ultra (LGA 1851), enter offsets as volts ----
XMP                    : enabled (verify first)
Core input voltage     : OFFSET, -0.100 V        ; = -100 mV
CPU ratio              : stock -1 (~-100 MHz, optional, thermal)
CPU DLVR               : 1.15
Long/Short Power Limit : set to TDP / boost targets
Current Limit          : safe board default

; ---- AMD AM5 Ryzen ----
EXPO                   : enabled (verify first)
PBO                    : Advanced
Curve Optimizer        : All Core, Negative, magnitude 15
VID intent             : ~1.200 V perf  |  1.05-1.10 V thermal

Windows mobile — ThrottleStop FIVR

; ThrottleStop FIVR — start conservative, deepen if stable
CPU Core   : -80 mV   (start) -> -125 to -165 mV (modern, if stable)
CPU Cache  : match CPU Core exactly
Intel GPU  : -50 mV
Uncore     : -50 mV

; Old 3rd/4th-gen Core: expect only -40 to -50 mV total.
; Persist: run at logon (Task Scheduler) + save voltages immediately.

Apply whichever block matches your hardware, then walk the 12 steps: baseline, XMP/EXPO, conservative offset, boot, idle test, stress test, step deeper, back off with margin, persist. If you internalize one thing from this entire guide, let it be that the voltage number is not the goal — the goal is a machine that is cooler, quieter, and exactly as correct as it was at stock. For the condensed, time-boxed version of this same procedure, see our companion 12-step, 45-minute undervolting walkthrough. The Machine has watched too many people chase -200 mV bragging rights into a corrupted save file. Don't be a data point.