GPU Overclocking 2026: 12 Steps, +15% Power, 60 Min

Here is the sentence every other overclocking guide buries under a stock photo of RGB fans: on a 2026 graphics card, the factory already did the overclocking for you. GPU Boost on the NVIDIA side and the equivalent clock-stretching on AMD's RDNA silicon spend every last millivolt and every degree of thermal headroom the moment you load a game. The card is already running as fast as its power limit, its temperature limit, and its voltage table will allow. What you are about to do is not unlock hidden speed. It is widen the ceilings the card is already pressed against, and then coax another few percent out of the gap.

That is the honest framing, and it is the one The Machine will hold to for the next several thousand words. If you came here expecting a free 40% and a new high score to lord over your friends, close the tab. If you came here to extract a real, stable, measurable gain from hardware you already own, using tools that are free and a method that has not meaningfully changed since Kepler, read on. We will do it in twelve numbered steps, push the power limit by roughly ten to twenty-five percent, and be done inside an hour of active work plus a couple of hours of validation you can run while you sleep.

Why Overclock at All in 2026

The uncomfortable truth is that overclocking a modern GPU is a low-yield activity dressed up as a high-yield one. It is worth doing anyway, for reasons that have more to do with headroom and heat than with hero benchmark numbers. Let us set expectations before you touch a single slider, because a tutorial that promises the moon is a tutorial that gets you a black screen at 2am.

The factory already ate your headroom

Every current-generation card ships with a boost algorithm that continuously auto-overclocks based on real-time power draw, temperature, and voltage. When you see a card rated at a 2,600 MHz "boost clock" quietly running at 2,800 MHz in a cool case, that is the algorithm doing its job. It has already taken the easy speed. This is why a manual overclock in 2026 typically buys you low single digits to low double digits in percent, not the 30% folklore from the era of fixed clocks and unlocked multipliers. The generational jump between two GPU tiers dwarfs anything you will squeeze out of a tune — if you want a real leap, that is a hardware conversation, not a slider conversation. Our RTX 5080 versus 4080 breakdown lays out exactly how small a generational delta can be, which puts a five-percent overclock in its proper, humble place.

What overclocking actually buys you now

Three things, in rough order of value. First, memory bandwidth: GDDR overclocks often scale better than core clocks on bandwidth-starved cards, and a memory-bound title can pick up more from a +500 MHz memory offset than from a +100 MHz core offset. Second, sustained clocks under a widened power and temperature limit, which is a genuine and repeatable win on cards that were throttling before you touched them. Third, and least glamorous, the education: you will come out of this understanding your specific card's thermal behavior, which is worth more than the frames. The physics is unforgiving and worth memorizing — dynamic power scales as P = C × V² × f, capacitance times voltage squared times frequency. Frequency buys you linear performance at linear power; voltage buys you frequency at quadratic power. That squared term is why the last 50 MHz costs so much heat, and why chasing maximum frequency is a fool's errand compared to a modest, efficient tune.

When it is worth it, and when it is not

Worth it: a well-cooled desktop card with thermal headroom to spare, a memory-bound workload, or a card that visibly throttles under load and would benefit from a raised power and temperature ceiling. Not worth it: any laptop or handheld (those are thermally and power constrained by design — you want undervolting there, not overclocking), a card already sitting at its thermal limit with the fans screaming, or a flagship that is already melting its own power connector. Speaking of which, if you are running a top-tier Blackwell card, read our RTX 5090 review before you widen its power limit — the phrase "still melts" is in the headline for a reason, and adding fifteen percent to an already-strained 12VHPWR connector is not a decision to make casually. Finally: raw frames are not the same as smooth frames. If your goal is a better-feeling game, a properly configured variable refresh display does more than an overclock ever will, as our G-Sync versus FreeSync piece argues at length.

Prerequisites: Versions and Hardware

Overclocking is a measurement discipline before it is a tuning discipline, and measurement requires the right instruments at the right versions. Get this section wrong and every number you record afterward is noise. Get it right and the rest is procedure.

The software stack, with correct 2026 versions

The tool everyone still reaches for is MSI Afterburner, and there is a version trap you must not step in. As of 2026, 4.6.6 is the stable build — released in October 2025, it was the first stable release in roughly two years and it added full support for the RTX 50 (Blackwell) series and unofficial support for AMD's Radeon RX 9000 (RDNA 4) cards, bundling RivaTuner Statistics Server 7.3.7. There is a 4.6.7 floating around, but it is Beta 3 (build 17352) — a beta with a genuinely useful Voltage/Frequency curve editor and thermal-analysis tooling, but a beta nonetheless, and betas carry expiration timers that stable builds do not. Use 4.6.6 for daily driving; dip into 4.6.7 only if you specifically want the V/F curve tools. Critically: download Afterburner only from MSI.com or Guru3D. Every other result — the SEO-farmed "MSI Afterburner: Overclock GPU in 12 Steps" clones in particular — is a re-hosted, ad-wrapped, or outright trojaned copy. There is no legitimate mirror.

You also need a benchmark, a stress tester, and a sensor logger. The version-correct 2026 stack:

# 2026 GPU-overclocking toolkit -- verified versions
MSI Afterburner       4.6.6   (STABLE, Oct 2025)   # 4.6.7 = Beta 3, not daily
RTSS (bundled)        7.3.7
GPU-Z                 latest  (techpowerup.com/gpuz)
HWiNFO64              8.x     (sensor log: VRAM + hotspot temps)
Unigine Superposition 1.1     (Basic edition = free)
3DMark Time Spy       latest  (optional, paid tiers exist)
FurMark 2             2.10.2  (2.6+ added RTX 50 support)
OCCT                  latest  (VRAM error-detection test)
NVIDIA driver         610.88 WHQL  (610 branch)     # or AMD Adrenalin (latest)

# Download Afterburner ONLY from:
#   https://www.msi.com/Landing/afterburner/graphics-cards
#   https://www.guru3d.com/download/msi-afterburner-beta-download/
# Anything else is re-hosted, ad-wrapped, or trojaned.

On the driver: keep to the current branch. The NVIDIA 610 series is current in mid-2026 (610.88 WHQL as of this writing), and note that the 610.47 release earlier in the year retired the classic Control Panel in favor of the NVIDIA App, which is now where the built-in Automatic Tuning lives. AMD users want the latest Adrenalin. Update the driver before you benchmark anything, so your baseline and your overclock share the same driver — otherwise you are measuring the driver, not your tune.

Hardware and thermal prerequisites

An overclock is a request for more power and more heat. The card can only honor it if the rest of the system can supply and dissipate both. Concretely: a power supply with genuine headroom above the card's rated board power (widening the power limit by twenty percent means the card will actually draw that), case airflow that moves intake to exhaust rather than recirculating hot air, and heatsinks that are not choked with three years of dust and cat hair. Physically inspect the card. If it sags in the slot — and every heavy triple-fan card does — the die-to-cooler contact and the fan bearings both suffer; our GPU support bracket guide is a twenty-minute fix worth doing before you add thermal load. Know your memory type, too, because it changes how hard you can push: RTX 50 cards run GDDR7 (the 5090 carries 32GB, the 5080 16GB), the RTX 40 generation ran GDDR6X, and AMD's RX 9000 cards run GDDR6. GDDR7 and GDDR6X both run hot and both report a dedicated memory junction temperature you must watch.

Ambient reality and a rollback plan

Overclock stability is measured in your room, not in a lab. A widely repeated 2026 prep rule is to keep ambient room temperature under 28°C, with clean heatsinks and functioning intake and exhaust airflow, before you begin — an overclock validated on a cold February night can and will crash during an August heatwave, so build your margin against the worst case, not the best. Finally, know your undo. In Afterburner, the reset button reverts every offset to zero instantly; the profile is not written to the card's firmware, so a reboot clears anything you did not explicitly save. The nuclear option, for a driver that will not behave, is Display Driver Uninstaller in safe mode followed by a clean driver install. On the warranty question: the NVIDIA App's Automatic Tuning is a first-party feature and does not void your warranty (it scans for 10 to 20 minutes and nets roughly +3 to +5% on average); manual offsetting occupies the usual grey area, though a card killed by a modest core offset is essentially unheard of because the voltage and temperature clamps stay active the entire time.

How GPU Overclocking Works

You cannot tune what you do not understand, and the reason most overclocks are unstable is that the person driving the sliders never learned what each one does. There are four of them. Learn them in order.

The four sliders

The table below is the entire mental model. Everything else is procedure built on top of it.

Why the power limit comes first

Because the boost algorithm is gated by power and temperature before it is gated by anything else. If the card keeps hitting its wattage ceiling, it will down-clock to stay under it, and any core offset you dial in simply cannot be realized — the card throttles right back through your "overclock." This is why every credible 2026 method, PCMag's and Tom's Hardware's included, opens by pushing the Power Limit slider to its maximum, described as roughly +10% to +25% depending on the card and vendor BIOS. On AMD's RX 9000 cards the Adrenalin power limit is hard-capped at +10%, so do not go looking for more; on many NVIDIA partner cards you will find +15% to +20% available. Raise the Temperature Limit to its maximum at the same time and, where Afterburner offers it, link the two so they move together. This single step, before you touch the core clock at all, often nets a measurable gain on a card that was previously throttling — free performance for one slider drag.

GDDR error correction: the score that drops before the artifact

This is the one piece of mechanism that separates people who understand memory overclocking from people who guess. Modern GDDR does not simply corrupt data when pushed too hard — it runs an error-detection-and-retransmission scheme, conceptually like the retransmit logic in Ethernet or Wi-Fi. When you push the memory clock past its real stability point, the memory controller starts silently catching bad transfers and re-sending them. You will not see an artifact yet. What you will see is your benchmark score climb, plateau, and then drop — because all that re-transmission is costing you effective bandwidth. That score drop, occurring before any visible corruption, is the real tell that you have passed the memory sweet spot. It is the single most useful fact in this entire tutorial: when tuning memory, watch the score, not just the screen. The moment the score stops rising, or dips, you have gone one step too far. This is straight from Tom's Hardware's method, and it is why memory tuning is a benchmark exercise, not a visual one.

Step 0: The Stock Baseline

Numbered step zero, because it happens before the overclock and because skipping it makes everything after it meaningless. You cannot claim a gain you never measured against a stock reference on the same driver, the same day, in the same room.

Why you benchmark before you touch anything

Every 2026 guide worth citing opens the same way: run a stock benchmark first. The reason is not ceremony. Silicon varies, ambient varies, driver versions vary, and case airflow varies; the only way to know your tune did something is to hold all of that constant and change one thing. Run the baseline immediately before you start tuning, on the driver you will keep, with the case closed and the side panel on — the way you actually run the machine. A baseline taken with the panel off and a box fan pointed at the card is a lie you will tell yourself later.

Running the baseline

The two standard choices are 3DMark Time Spy and Unigine Superposition. Superposition's Basic edition is free and perfectly sufficient. Configure a repeatable run and record everything:

Unigine Superposition 1.1 -- baseline run
  Preset:      1080p Extreme    (use 4K Optimized on a 5080/5090-tier card)
  API:         DirectX
  Fullscreen:  Yes
  Loops:       1  (for scoring)  /  Endless  (for stability later)

Record for the run:
  - Score
  - Average FPS
  - Sustained core clock (MHz, under load, not the idle spike)
  - GPU edge temp AND hotspot temp (C)
  - VRAM / memory junction temp (C)
  - Board power draw (W)

What to record, and what it looks like

Log the numbers somewhere you will not lose them. HWiNFO64 can write a CSV of every sensor across the run, which is the disciplined approach; a phone photo of the Afterburner OSD is the lazy one. Your output will look something like the block below — and this is the last time we will say it, so internalize it: these numbers are illustrative. They are the shape of the data, not a claim about your card. Your silicon, your cooler, and your room will produce different figures, and that is exactly the point of taking your own baseline.

# ILLUSTRATIVE ONLY -- your card, your silicon, your numbers
Baseline (stock):
  Score .......... 12,840
  Avg FPS ........ 96.1
  Core clock ..... 2,745 MHz (sustained under load)
  Edge temp ...... 67 C
  Hotspot ........ 79 C
  VRAM temp ...... 74 C
  Board power .... 285 W

The Overclock in 12 Steps

This is the procedure. It is deliberately conservative, deliberately incremental, and deliberately boring, because boring is what stable looks like. Each step states the action and the reason for it. Do them in order. Do not skip ahead to memory before the core is locked in, and do not touch voltage until you have read the advanced section and decided you actually need it.

1. Update the driver, then install Afterburner 4.6.6 stable. A current driver on the correct branch (NVIDIA 610-series, or the latest AMD Adrenalin) and the stable Afterburner build give you a known-good, RTX 50 / RX 9000-aware starting point. Reboot after the driver install so you are not tuning on a half-loaded driver.
2. Run and log a stock baseline. One clean pass of Superposition or Time Spy with the case closed, recording score, clocks, all three temperatures, and board power. This is your reference; every later number is meaningless without it.
3. Raise the Power Limit to maximum, and the Temp Limit with it. Drag Power Limit to its ceiling — typically +10% to +25%, or the hard +10% cap on AMD RX 9000 — and raise the Temperature Limit to max, linking them if Afterburner lets you. Boost is power- and thermal-gated; without this, later core offsets simply throttle away.
4. Apply, then re-run the benchmark. Measure the "free" gain from the power and temperature ceilings alone. On a previously-throttling card this step alone can be most of your total uplift, and it costs you nothing in stability risk.
5. Raise the Core Clock by +15 MHz. Apply. Small increments are the entire game. A +15 MHz step (inside the widely-recommended +10 to +25 MHz range) keeps each change small enough that when something breaks, you know precisely what broke it. Large jumps hide the ceiling and waste your afternoon.
6. Retest after every single step. Run a short benchmark loop — PCMag's method is to apply, then retest with Superposition each step — watching for artifacts, a hang, a driver reset, or a black screen. If it passes clean, log the new score and continue. This is tedious. Tedious is correct.
7. Repeat +15 MHz until the first instability. Keep stepping the core offset up, retesting each time, until the card throws its first symptom: a crash to desktop, a driver timeout, a flicker, or a hard hang. That symptom marks your core ceiling. Note the offset value at which it failed.
8. Back off 15 to 25 MHz from the failure point. Once the first instability appears, step back — the standard 2026 advice is 10 to 25 MHz below the crash — to land comfortably under the wall rather than right against it. "Passes once" is not "stable"; you want daily-driver margin, not a benchmark stunt.
9. Now tune Memory Clock in +50 MHz steps, retesting each. With the core locked in, move to memory. Step the memory offset up by +50 MHz at a time toward the +300 to +500 MHz range, retesting after each. Watch the score, not just the screen — remember the error-correction mechanism.
10. Back memory off to the last score peak. The instant the benchmark score stops rising or dips — which happens before visible artifacts, because GDDR silently retransmits bad data — back the memory off by roughly 50 to 100 MHz to the last offset where the score was still climbing. That peak is your true memory sweet spot; anything past it is negative throughput dressed up as a bigger number.
11. Build and apply a custom fan curve. Sustained clocks demand sustained cooling, and the stock curve is tuned for quiet, not for heat. Configure the breakpoints in the fan-curve section below so the card actually holds its new clocks under load instead of thermal-throttling right back through your work.
12. Validate, then save the profile and enable apply-at-startup. Run 30 to 60 minutes of synthetic stress, then 1 to 2 hours of real games, before you trust it. Only after it survives both should you save the profile and enable "apply at startup." Synthetic stability does not guarantee real-world stability, and a profile that auto-applies at boot before you have validated it is how you get a crash-loop on the login screen.

Stress Testing and Validation

An overclock that passes a single benchmark loop is a hypothesis, not a result. Validation is where you try to disprove it, and if you are not actively trying to make the card crash, you are not testing — you are hoping.

Synthetic first: quick check, then a real soak

Work in two tiers. The quick check is 10 to 15 minutes of a looping load — enough to catch the obvious failures, the offsets that were never going to hold. If it survives that, escalate to the real soak: 30 to 60 minutes of sustained synthetic load before you are allowed to use the word "stable." The longer window matters because thermal saturation is slow; a card that is rock-solid at minute five can fail at minute forty once the VRAM and the VRM have fully heat-soaked and the boost clocks have settled to their true sustained values. Do not shortcut this. The soak is the difference between an overclock and a time bomb.

The tools, and what each one catches

Use the right instrument for the right failure mode. Superposition on an Endless loop is your general-purpose core and thermal soak. OCCT's dedicated VRAM test is the one that catches memory errors specifically — it reports an error count that must read exactly zero, and it will find memory instability that a graphics benchmark's score-drop only hints at. FurMark 2 (2.10.2, with RTX 50 support since 2.6) is a deliberate power-virus: it drives the card to a thermal worst-case that no real game will ever produce. That makes it excellent for validating your cooling and useless for validating gaming stability — use it to check that your fan curve holds the line under maximum heat, not to certify your overclock. Log the whole session with HWiNFO64 so you can go back and see exactly where and when anything went wrong.

# Quick check (10-15 min):  Superposition -> Endless loop.
#   Watch for: artifacts, flicker, hang, or a driver reset.
#
# Memory validation:        OCCT -> GPU -> VRAM test.
#   Error count MUST be 0. Any errors = memory offset too high.
#
# Thermal/power soak (30-60 min): OCCT 3D Adaptive OR Superposition Endless.
#   FurMark 2 = power-virus: use it to test COOLING, not stability.

# ILLUSTRATIVE PASS (OCCT VRAM test):
  Test duration .. 00:45:00
  VRAM errors .... 0
  Hotspot max .... 83 C
  Throttling ..... none detected

Then the games: 1 to 2 hours of real load

Synthetic stability does not guarantee real-world stability — the two stress the card differently, and plenty of overclocks that survive an hour of OCCT fall over twenty minutes into an actual game with its irregular, bursty load and its own shader-compilation spikes. The final validation pass is 1 to 2 hours of the games you actually play, ideally the most demanding ones you own. Only after the card survives both the synthetic soak and the real-game session should you consider the overclock done and save it as your startup profile. A tune that passes synthetics but crashes in your favorite game is not a stable tune; it is a tune that has not met your game yet.

Building a Fan Curve

The stock fan curve is a compromise the vendor made on your behalf, and the priority they chose was acoustics, not sustained clocks. When you widen the power and temperature limits, you change the deal: the card now wants to run hotter for longer, and the quiet stock curve will let it thermal-throttle right back through your overclock. A custom curve is not optional on a tuned card.

Why the stock curve fights your overclock

Vendors tune the default curve so the card is quiet in a showroom and in reviews, which means it lets the temperature climb before it spins the fans up, and it prioritizes low noise over holding a clock. That is a reasonable default for a stock card and a bad one for an overclocked card, because every degree over your temperature limit is a down-clock. You want the fans to respond earlier and more aggressively, trading some noise for sustained frequency. The trade is yours to make, not the vendor's.

The breakpoints

A sensible, widely-cited 2026 curve uses five breakpoints that keep the fans calm at idle and ramp them hard as the card heats up. Set it in Afterburner under Settings, Fan, then enable user-defined software automatic fan control:

# MSI Afterburner -> Settings -> Fan
#   -> Enable user defined software automatic fan control
# Breakpoints (temp C -> fan %):
30 -> 20
50 -> 35
70 -> 65
80 -> 85
85 -> 100

# Set "Update period" to ~1000 ms and enable a small hysteresis
# so the fans do not oscillate/hunt around a breakpoint.

The acoustic-versus-thermal trade

This curve is intentionally aggressive in the 70 to 85°C band, which is where an overclocked card lives under sustained load, and it will be audibly louder there than the stock profile. That is the correct trade for a tuned card. If the noise bothers you, the answer is not a lazier curve — it is better case airflow and a lower ambient, which lets the whole curve sit at lower temperatures. Do not solve a thermal problem by telling the fans to ignore it. Set the update period to around a second and add a little hysteresis so the fans do not hunt annoyingly up and down every time the temperature brushes a breakpoint.

Common Pitfalls and Fixes

The failure modes here are predictable, which is good news: predictable means avoidable. Here are the eight that catch the most people, grouped by where in the process they bite.

Software and source mistakes

Pitfall 1: downloading Afterburner from the wrong place. The search results are polluted with re-hosted, ad-wrapped, and trojaned copies, including SEO-farm clickbait promising to "overclock your GPU in 12 steps." Fix: download only from MSI.com or Guru3D. There is no legitimate third mirror. Pitfall 2: mistaking the 4.6.7 beta for a stable release. The brief that lands in a lot of inboxes calls 4.6.7 the current stable build; it is not — 4.6.7 is Beta 3 (build 17352), and betas carry expiration timers. Fix: run 4.6.6 stable for daily use; treat 4.6.7 as a tools preview only. Pitfall 3: not updating the driver before you baseline. Baseline on one driver, overclock on another, and you have measured the driver, not your tune. Fix: update first, reboot, then baseline.

Tuning mistakes

Pitfall 4: chasing the memory clock past the score peak. Because GDDR silently retransmits errored data, the benchmark score drops before you see artifacts — so people who tune memory "until it glitches" have already sailed well past the point of negative returns. Fix: watch the score; the instant it stops climbing, back off 50 to 100 MHz to the peak. Pitfall 5: taking +100 MHz core jumps to "save time." Big steps blow past the ceiling and hide which offset actually failed, turning a one-hour job into an afternoon of guessing. Fix: +15 MHz steps, retest each. Pitfall 6: reaching for voltage first. Voltage is quadratic heat for linear frequency and is very often locked or clamped anyway; leading with it is how beginners cook their thermals for nothing. Fix: leave voltage at zero until you have exhausted the core and memory offsets and read the advanced section.

Validation mistakes

Pitfall 7: calling a synthetic-only pass "stable." Synthetic stability does not guarantee real-world stability; the card that survives an hour of OCCT can still crash in an actual game. Fix: always finish with 1 to 2 hours of the games you play. Pitfall 8: ignoring VRAM and hotspot temperatures. GDDR7 and GDDR6X run hot, and a memory junction quietly cooking at 100°C-plus will throttle or degrade regardless of how happy your edge temperature looks. Fix: log the VRAM and hotspot sensors specifically, not just the headline edge temperature, and if the memory junction is the limiter, improve airflow before you push the memory offset further.

Troubleshooting Table

When something goes wrong — and on a first overclock, something will — the symptom tells you exactly which slider overreached. Read the symptom, do not panic, apply the fix.

Symptom, cause, and fix

Reading the symptom correctly

The single most useful diagnostic distinction is when it crashes. A crash at idle or on the desktop points at the core offset failing in the low-voltage boost states, where the card runs at reduced voltage and a too-high core offset has no margin. A crash under heavy load with visible artifacts points at memory. A crash that only happens in games while synthetics pass points at a marginal tune that real-world bursty load exposes. Match the symptom to the slider and you will fix it in one or two steps instead of flailing at all four at once.

The nuclear option: DDU and a clean slate

If the driver itself becomes unstable — repeated timeouts that persist even after you reset every offset to zero — stop tuning and clean-slate the driver. Boot into safe mode, run Display Driver Uninstaller to strip every trace of the current driver, reboot, and install the current branch driver fresh. This clears corrupted driver state that no amount of slider-wiggling will fix, and it is the correct first move whenever the card misbehaves at stock after a session of overclocking.

Advanced: Voltage and Linux

Everything above gets you a stable, sensible overclock without touching the two things that separate careful tuners from reckless ones: voltage and the operating system underneath. Here is where the real control lives — and the real risk.

Voltage and the V/F curve

Voltage is the last resort, not the first, because of that quadratic power term. But if you want to try it, a 2026 beginner-friendly path is to open Afterburner's General tab and enable Unlock voltage control and Unlock voltage monitoring — but only if your card actually supports it, and many partner cards restrict or lock the voltage table entirely. The more sophisticated technique, and the reason the 4.6.7 beta's Voltage/Frequency curve editor exists, is not to raise voltage but to flatten the curve: you pin a target frequency to a lower voltage point, which is an undervolt and an overclock at once — more sustained clock for less heat and power. That is the same efficiency logic we apply to processors in our CPU undervolting guide, and it is genuinely the smartest way to tune a thermally-limited card. If all of this sounds like more risk than you want, the NVIDIA App's built-in Automatic Tuning does a safe, warranty-preserving version automatically — a 10-to-20-minute scan for roughly +3 to +5% — and it is a perfectly respectable place to stop.

AMD specifics

AMD's Adrenalin software handles this differently from Afterburner. The power limit is hard-capped at +10% on RX 9000 (RDNA 4) cards, so there is no +25% to chase; the built-in tuning exposes core and memory sliders plus a voltage offset expressed in millivolts, and the RDNA-native move is to lower that voltage offset for an efficiency-focused undervolt-overclock rather than raise it. RX 9000 runs GDDR6, which is more forgiving on temperature than GDDR7 but still wants its junction temperature watched. Adrenalin's own tuning presets are a reasonable starting point, and unlike the NVIDIA path everything lives in one first-party application.

Linux: Coolbits and amdgpu sysfs

Overclocking on Linux is entirely possible and, once configured, arguably cleaner than Windows. On NVIDIA under X11 you enable the controls with the Coolbits mask, then apply offsets through nvidia-settings:

# NVIDIA on Linux (X11) -- enable overclocking controls
sudo nvidia-xconfig --cool-bits=28        # 28 = clock + fan + voltage bits

# then, per session (offsets in MHz):
nvidia-settings -a '[gpu:0]/GPUGraphicsClockOffsetAllPerformanceLevels=150'
nvidia-settings -a '[gpu:0]/GPUMemoryTransferRateOffsetAllPerformanceLevels=1000'
nvidia-settings -a '[gpu:0]/GPUFanControlState=1'
nvidia-settings -a '[fan:0]/GPUTargetFanSpeed=65'

# GUI alternative (Wayland-friendly): LACT
#   https://github.com/ilya-zlobintsev/LACT

On AMD you first unlock the overdrive interface with a kernel parameter, then read and write the clock/voltage table through sysfs:

# AMD on Linux -- kernel param unlocks overdrive, then sysfs:
#   GRUB_CMDLINE_LINUX_DEFAULT="... amdgpu.ppfeaturemask=0xffffffff"

cat /sys/class/drm/card0/device/pp_od_clk_voltage        # read current table
echo 's 1 2600' | sudo tee /sys/class/drm/card0/device/pp_od_clk_voltage  # SCLK state 1 -> 2600 MHz
echo 'm 1 1075' | sudo tee /sys/class/drm/card0/device/pp_od_clk_voltage  # MCLK state 1
echo 'vo -30'   | sudo tee /sys/class/drm/card0/device/pp_od_clk_voltage  # voltage offset -30 mV
echo 'c'        | sudo tee /sys/class/drm/card0/device/pp_od_clk_voltage  # COMMIT the changes

# GUI alternative: CoreCtrl
#   https://gitlab.com/corectrl/corectrl

LACT and CoreCtrl wrap all of this in a GUI if you would rather not hand-edit sysfs, and both handle fan curves and persistence across reboots, which the raw commands do not.

The Complete Configuration

Here is the whole thing distilled into a starting profile and a verification checklist. Do not paste these offsets and walk away — they are a starting point to tune up or down per your silicon using the twelve steps. But they encode the sensible-defaults consensus of every credible 2026 guide, and they will not hurt a healthy card.

A conservative daily profile

Two consensus profiles worth knowing. The safe profile, for people who want the efficiency win without the risk: core +50 to +100 MHz, memory +300 to +500 MHz, power limit +5 to +10%, no voltage change. The typical profile, one notch more aggressive but still voltage-neutral: core +100 MHz, memory +500 MHz, power limit +10%, no voltage change. Both deliberately leave voltage alone, because the efficiency and stability of a voltage-neutral tune is worth more than the last one percent of frequency a voltage bump might buy. Start at the safe profile, validate, then push toward typical only if you want to.

The Afterburner profile, annotated

; ===== staresback.gg -- conservative daily GPU profile (2026) =====
; MSI Afterburner 4.6.6 stable. Values are OFFSETS, not absolutes.
; Start here, then tune UP per the 12 steps. No voltage change.

[Profile1]
PowerLimit     = +10        ; % over stock (AMD RX 9000 hard-caps at +10)
TempLimit      = MAX        ; raise the thermal ceiling; link to power
CoreClock      = +100       ; MHz offset; back off to +75 on first instability
MemoryClock    = +500       ; MHz offset; back off ~50 to the last SCORE peak
VoltageOffset  = 0          ; leave voltage alone unless you know why
FanCurve       = 20/35/65/85/100  @  30/50/70/80/85 C
ApplyAtStartup = 1          ; ONLY after 1-2 h of real-game validation

; "Safe" starter (even more conservative):
;   PowerLimit +5..10 / CoreClock +50..100 / MemoryClock +300..500 / Voltage 0

The verification checklist before you trust it

Before you enable apply-at-startup and forget about it, walk the checklist one last time. Driver current and rebooted. Baseline recorded on that driver. Power and temperature limits maxed. Core offset validated in +15 MHz steps and backed off 15 to 25 MHz from its ceiling. Memory offset tuned to the score peak and backed off, with OCCT's VRAM error count reading exactly zero. Custom fan curve applied and holding temperatures in the 70 to 85°C band under load. Thirty to sixty minutes of synthetic soak passed. One to two hours of real games passed. Only when every one of those is true do you save the profile and enable apply-at-startup. An overclock is not a number you set once; it is a state you validated and can re-validate after the next driver update. Tune it like you mean to keep it, and it will keep you in frames — modestly, honestly, and without ever melting anything you cannot afford to replace.

For the canonical method references this tutorial builds on, see Tom's Hardware's overclocking how-to and PCMag's step-by-step guide; grab the tools from MSI Afterburner or Guru3D, benchmark with Unigine Superposition, and validate cooling with Geeks3D FurMark 2 and memory with OCCT.