How to Undervolt Your CPU in 12 Steps (2026, 45 Min)

There is a specific category of computing advice that sounds like a scam: do this one thing and your machine will run cooler, draw less power, throttle less, and lose no performance. The retro-gaming corner of the internet is full of such promises, most of them attached to a coupon code. Undervolting is the rare exception. It is not a trick, it is not a tweak, and it is not, contrary to the forum mythology, dangerous in the way overclocking can be. It is the act of telling your CPU to do exactly what it already does, using slightly less voltage to do it.

The reason this works at all is that silicon is conservative. Every chip leaves the fab with a voltage curve that has to keep the worst sample in a production batch stable, across a temperature range, for years, under a warranty. Your specific chip is almost certainly better than that worst-case sample. The factory voltage is a margin of safety you paid for and never use. Undervolting reclaims it. The ArchWiki documentation on undervolting, in its 2025–2026 form, puts it plainly: done correctly, undervolting reduces energy consumption and heat without affecting performance. That last clause is the entire reason this article exists. You are not trading speed for temperature. You are deleting waste.

This is a tutorial, not a sales pitch, so here is the catch up front: "done correctly" is carrying weight in that sentence. An undervolt that is too aggressive does not gently degrade. It crashes. It blue-screens mid-boss-fight, it corrupts a save, it freezes a RetroArch netplay session at the worst possible frame. The difference between a free win and an afternoon of frustration is testing methodology, and roughly ninety percent of this guide is about testing methodology. Read it before you touch a voltage slider.

Why Undervolting Works

Power dissipation in a CMOS processor is dominated by the dynamic switching term, which scales with capacitance, frequency, and the square of voltage. That squared relationship is the whole game. A modest voltage reduction produces a disproportionately large drop in power and, because power becomes heat, a disproportionately large drop in temperature. Drop voltage ten percent and you do not save ten percent of power — you save closer to twenty, because the voltage term is squared while the chip continues to clock at the same frequency it always did.

Why does temperature matter to a gamer specifically? Because modern CPUs do not run at a fixed clock. They run at the highest clock they can sustain within a thermal and power budget, and they back off — "throttle" — the instant they exceed it. A cooler chip throttles later and less. On a thermally constrained laptop running an emulator that hammers a single core, or a small-form-factor handheld running a demanding core like a Saturn or PS2 emulator, undervolting frequently raises sustained performance, not by overclocking but by keeping the chip out of the throttle ceiling longer. The clock you were promised on the box is the clock you actually get to keep.

The mechanism you adjust is the voltage offset — a negative number, expressed in millivolts (mV), subtracted from whatever voltage the chip would otherwise request at every point on its frequency-voltage curve. You do not pick a single fixed voltage. You shift the entire curve down by a constant. At idle the chip still drops to low voltage; under load it still ramps up; the whole staircase simply sits a little lower than before. This is why a well-chosen undervolt is invisible in benchmarks and obvious on a temperature graph.

One terminology note that trips up newcomers: on Intel, the two offsets that matter most are CPU Core and CPU Cache (sometimes labeled "CPU" and "CPU Cache" or "core" and "ring/LLC"). The well-worn advice from the UltrabookReview ThrottleStop guide is that these two should generally carry the same undervolt value — set Core to −80 mV, set Cache to −80 mV — because mismatched values are a common and maddening source of instability. On AMD's Ryzen platform the equivalent control is the Curve Optimizer, which expresses the offset in abstract "counts" rather than millivolts, but the principle is identical: shift the curve down, test, repeat.

Prerequisites: Hardware and Software

Undervolting is platform-specific. There is no single button. What you install and where you click depends on your CPU vendor, your generation, and your operating system. Sort yourself into the correct lane before you download anything.

Hardware requirements. You need to know your exact CPU. Not "an i7" — the specific SKU, e.g. Core i7-9750H, Core Ultra 7 265K, Ryzen 7 9800X3D, Ryzen 7 7840U. The generation determines both your method and your realistic headroom. Critically, Intel disabled voltage-offset control on most 10th-generation and many later mainstream desktop chips via the Plundervolt mitigation (a 2019 security fix for a real vulnerability). If you own one of those, software undervolting through the offset interface simply will not take, and your only path is BIOS-level control if the board exposes it. Older 6th- through 9th-generation Intel mobile chips, and the new Core Ultra / LGA 1851 desktop platform, are the most cooperative. On AMD, essentially all Zen-based Ryzen chips from Zen 2 onward support Curve Optimizer through BIOS or through Ryzen Master.

Software, by lane:

• Intel, Windows, laptop or older desktop: ThrottleStop (latest version, currently the 9.x series). It is a single portable executable — no installer. Pair it with HWiNFO64 (latest) for sensor monitoring.
• Intel, Windows, Core Ultra / LGA 1851 desktop: your motherboard BIOS, plus HWiNFO64 and an in-Windows stress tool. BIOS access varies by vendor — the 2026 Intel Core Ultra tuning guide notes F6 on ASRock and F7 on MSI to reach Advanced Mode.
• Intel, Linux: the intel-undervolt tool, configured via /etc/intel-undervolt.conf. ArchWiki documents the full workflow.
• AMD, Windows, AM5 desktop: BIOS Curve Optimizer (Precision Boost Overdrive menu), or AMD Ryzen Master (latest) for live tuning.
• AMD, Windows, Ryzen laptop: Curve Optimizer where exposed, or a vendor utility; the 2026 Ryzen laptop undervolt guide walks through it.

Stability-testing software (all lanes). This is not optional and it is not where you cut corners. ArchWiki specifically recommends stress tests that perform numerical checks — that is, tests that verify their own math and will catch a silently wrong calculation that a simple load test would miss. It names mprime (the Linux/cross-platform build of Prime95) and linpack as the preferred tools. On Windows you want Prime95 (Small FFTs torture test), and for a broader real-world net, Cinebench and 3DMark. The Ryzen laptop guide explicitly recommends testing with 3DMark, Prime95, and Cinebench together, because each stresses the chip differently and a stable system survives all three.

Time and conditions. Budget 45 minutes for a first conservative pass. Budget several hours, spread across days, for a fully validated daily-driver undervolt — because the only honest stability test is one that runs long. Do this on a machine where a crash costs you nothing: close your work, save your games, and accept that you will crash at least once. That crash is data, not failure.

The Procedure: 12 Numbered Steps

This is the platform-agnostic skeleton. The vendor-specific sections that follow fill in the exact clicks. Every step has a reason, because a step you do not understand is a step you will skip, and the one you skip is the one that bites you.

1. Identify your exact CPU and confirm it supports undervolting. Rationale: the Plundervolt mitigation silently disables offset control on a large swath of Intel chips. Five minutes spent confirming support saves an hour spent wondering why nothing changes.
2. Record a baseline. Before changing anything, run your stress test for 10 minutes at stock voltage and write down peak temperature, peak package power (watts), and sustained clock speed. Rationale: "it runs cooler" is meaningless without a before number. This is your control.
3. Set up live monitoring. Open HWiNFO64 (Windows) or watch sensors / turbostat (Linux) on a second screen or window. Rationale: you must see voltage and temperature change in real time, or you are flying blind.
4. Apply a conservative first offset. Intel: −80 mV on CPU Core, the same −80 mV on CPU Cache. AMD: Curve Optimizer −10 (negative) all-core. Rationale: −80 mV on Intel and −10 on AMD are the broadly compatible starting points recommended across the 2026 guide material. Conservative first, aggressive never.
5. Apply temporarily, not permanently. Do not save to BIOS or enable any persistence/service yet. Keep the offset live-only so a crash clears it on reboot. Rationale: a permanent bad undervolt that crashes on boot is a recovery problem; a temporary one is a non-event.
6. Run a short stability test (15–20 minutes). Prime95 Small FFTs or mprime with numerical checks enabled. Rationale: most unstable undervolts fail within the first few minutes of a numerical-check load. This pass weeds out the obviously-too-aggressive.
7. Compare against baseline. Confirm temperature and power dropped and clock speed did not. Rationale: if performance fell, you have done something other than undervolt — likely tripped a power limit — and must investigate before going further.
8. Step deeper in small increments. If stable, drop another −10 to −25 mV (Intel) or another −5 on Curve Optimizer (AMD) and retest from step 6. Rationale: the failure point is found by walking up to the edge, then stepping back, never by jumping.
9. Find the failure point, then back off. When a value crashes or errors, that is your floor. Your daily setting is one comfortable step above the floor, not at it. Rationale: a setting that barely passes a 20-minute test will fail a 6-hour gaming session. The margin is the point.
10. Run the long stability test (several hours). At your chosen daily value, run mprime/Prime95 for hours, ideally overnight, plus a real workload — the actual emulator or game you care about. Rationale: short tests catch gross instability; only long tests catch the once-an-hour rounding error that corrupts a save file.
11. Make it persistent. Only now: save to BIOS, enable intel-undervolt.service, or set ThrottleStop to start with Windows. Rationale: persistence is the reward for proven stability, not a step you take on faith.
12. Document your settings. Write down every value in a text file kept with your backups. Rationale: a BIOS reset, a Windows reinstall, or a CMOS-clear erases everything, and you do not want to re-derive your floor from scratch. The complete-configuration section at the end is your template.

Intel on Windows: ThrottleStop

ThrottleStop is the canonical Windows tool for Intel mobile and older-desktop undervolting, and despite a UI that looks like it escaped from 2009, it remains the most precise instrument in the category. The UltrabookReview guide is the reference text, and its core advice is worth tattooing somewhere visible: CPU Core and CPU Cache should generally carry the same undervolt, and you should test stability thoroughly before saving the voltages permanently.

The relevant controls live behind the FIVR button (Fully Integrated Voltage Regulator). The interface looks intimidating; you touch exactly two things. Select CPU Core in the FIVR Control box, check Unlock Adjustable Voltage, and set the Offset Voltage negative. Then select CPU Cache and set the identical offset. Here is the conceptual configuration you are building — not a file you edit, but a mental model of the two values:

; ThrottleStop FIVR — conceptual settings (set via the GUI)
; ------------------------------------------------------------
FIVR Control:
  CPU Core   -> Unlock Adjustable Voltage: ON
               Offset Voltage:  -80.1 mV     ; conservative start
  CPU Cache  -> Unlock Adjustable Voltage: ON
               Offset Voltage:  -80.1 mV     ; MUST match Core

  Intel GPU  -> Offset Voltage:   0.0 mV      ; leave for later
  System Agent -> Offset Voltage: 0.0 mV      ; leave alone

; Click 'Apply' (temporary). Do NOT check 'OK - Save voltages
; immediately when I click OK' until stability is proven.

The starting value is −80 mV on both. From there you walk deeper. The UltrabookReview guide's generational map is the most useful sizing data in this entire field: many modern mobile CPUs tolerate roughly −125 mV to −165 mV, while older 3rd- and 4th-generation Core chips may only manage about −40 mV to −50 mV before they destabilize. Do not read those as targets. Read them as the ceiling of what is plausible, so you know whether your stable result of −110 mV is a great chip or a sign you stopped too early. If you own a Haswell-era laptop and you are sitting stable at −145 mV, you have either won the silicon lottery or your test is too short. It is almost always the latter.

The single most important toggle in the whole tool is the one labeled approximately "OK – Save voltages immediately when I click OK." Leave it off during testing. With it off, an unstable value evaporates on reboot and you simply try again. Turn it on only after the long test in step 10 passes. ThrottleStop also has a per-profile structure and a "Start ThrottleStop in Windows" option in its options page; that is your persistence mechanism (step 11), enabled last.

Intel Core Ultra: BIOS Tuning

The new Intel desktop platform — Core Ultra on LGA 1851 — moves the action into the motherboard BIOS, and pairs undervolting with power-limit and ratio tuning in a way the older mobile workflow did not. The 2026 Core Ultra / LGA 1851 guide frames the upside aggressively, claiming users can extract 20–30% more performance alongside better temperatures through combined BIOS tuning — a figure that reflects power-limit headroom being unlocked as much as voltage being lowered. Treat 20–30% as a best-case marketing-adjacent number, not a guarantee, but the direction is real: on this platform, undervolting is part of a tuning package, not a standalone trick.

First, get into Advanced Mode. As noted, BIOS access varies by vendor: the guide cites F6 on ASRock and F7 on MSI. Other boards use Del to enter setup and then a key or click to reach Advanced. Once inside, two ideas matter.

Voltage offset entry, divided by 1,000. A recurring confusion on these boards is how the offset field expects its input. The Core Ultra guide's creator recommends entering a value divided by 1,000 — that is, the UI takes a volt-denominated number where a −0.080 entry means −80 mV. This is simply how many motherboard firmwares express a millivolt-style input, and getting the decimal place wrong by a factor of a thousand is how people accidentally request an offset that either does nothing or instantly kills the boot. Read your specific field's units carefully.

All-core ratio reduction for temperature. Beyond voltage, the guide suggests lowering all-core CPU ratios by about 100 MHz from stock for materially better temperatures, using concrete example values of 54 for P-cores and 46 for E-cores. Dropping one multiplier bin costs a sliver of peak clock you will never notice in an emulator and buys a meaningful thermal cushion. Here is the conceptual BIOS configuration:

# Intel Core Ultra / LGA 1851 — BIOS Advanced Mode (conceptual)
# Enter Advanced: ASRock = F6, MSI = F7

CPU Voltage Offset Mode .............. Adaptive + Offset
CPU Core Voltage Offset .............. -0.080   # = -80 mV (value / 1000)
CPU Cache (Ring) Voltage Offset ...... -0.080   # match the core

P-Core Ratio (all-core) .............. 54       # ~100 MHz under stock
E-Core Ratio (all-core) .............. 46       # ~100 MHz under stock

Long Duration Power Limit (PL1) ...... tune to cooler capability
Short Duration Power Limit (PL2) ..... tune to cooler capability

# Save & Exit, then validate in Windows before trusting it.

The BIOS path has one ergonomic disadvantage versus ThrottleStop: a bad value that crashes during POST means a trip to clear CMOS rather than a simple reboot. Locate your board's CMOS-clear jumper or button before you start, so an unbootable setting is a thirty-second fix and not a panic.

Linux: intel-undervolt

On Linux, the intel-undervolt tool is the documented path, and the ArchWiki workflow is the canonical reference. The tool is open source; the upstream project lives at github.com/kitsunyan/intel-undervolt, which is also where the configuration syntax is documented in full. Configuration lives in /etc/intel-undervolt.conf.

ArchWiki's sizing guidance for Linux mirrors the Windows numbers but states them as a range: for Intel CPU and CPU Cache, a reduction of about 100 mV to 200 mV is often stable, while going beyond 200 mV may crash the system or simply have no effect. The "no effect" outcome is the chip clamping the offset internally — a sign you have exceeded what the silicon will accept, not a sign to push harder. Edit the config to set your offsets:

# /etc/intel-undervolt.conf
# Negative offsets, expressed in millivolts.
# Start conservative; ArchWiki: 100-200 mV often stable, >200 may crash.

enable        no            # flip to 'yes' once a value is proven

# undervolt