GPU Overclocking 2026: 12 Steps, +25 MHz, 2 Hours

There is a specific species of optimism that types how to overclock GPU into a search bar at midnight, and this tutorial is written for that person with the weary sympathy of someone who has watched the same display driver crash and recover forty times in a single evening. The good news first: overclocking a graphics card in 2026 is no longer the pencil-on-the-shunt, flash-the-BIOS-and-pray dark art it was when Alexey Nicolaychuk's RivaTuner ruled the desktop and MSI Afterburner was still a fork of it. The software clamps the voltage. The firmware clamps the temperature. The card will, in almost every case, protect itself from you.

The bad news is that everything worth having still costs patience. A real overclock is not a slider you yank to the right; it is a documented, incremental, deliberately boring process of changing exactly one variable, testing it, writing down the result, and repeating until the card misbehaves. Do it properly and you spend an evening to gain a handful of frames. Do it impatiently and you spend the same evening reinstalling drivers with a hollow look in your eyes. This guide runs about twelve discrete steps and, if you actually respect the stability testing, roughly two hours of hands-on work spread across a longer validation window. Anyone who tells you it is faster than that has not tested their result, and their result will fail in a game they have not played yet.

Why Overclock in 2026 (and When Not To)

What an overclock actually buys you

Set expectations before you touch a slider. A modern GPU already boosts itself aggressively. NVIDIA's GPU Boost and AMD's equivalent spend the entire time you are gaming quietly probing for the highest clock the current voltage, temperature, and power budget will allow. What you are doing when you \"overclock\" is not unlocking a hidden gear that the manufacturer cruelly withheld. You are widening the box those algorithms are allowed to play in — raising the power ceiling, nudging the clock offset upward, improving the cooling — so the card's own boost logic can hold a higher number for longer before something forces it back down. The payoff is typically a single-digit percentage in real games, occasionally reaching into low double digits on a bandwidth-starved title, and it is entirely possible to spend two hours to move a 60 fps average to 63.

Whether that matters is your call, but measure it honestly. How much you actually feel a few percent depends heavily on your resolution and refresh target: a card that is comfortable at 1440p but gasping at 4K is exactly the case where a small uplift can drag a stuttering title back over your monitor's refresh floor, and a card that is already pinned to your frame cap will show you nothing at all for the effort. The ceiling also scales with the silicon; a flagship such as the one dissected in our RTX 5090 review, where the $1,999 sticker is only the beginning of the story, has a very different power and thermal envelope than a mid-tier board, and its stock boost behaviour already eats most of the low-hanging fruit.

The warranty and lifespan reality

Two things are true at once. Overclocking within the sliders a tuning utility exposes is genuinely low-risk on 2026 hardware, because the card refuses to exceed its own voltage and thermal limits no matter how enthusiastically you drag things. And yet the current crop of 2026 how-to guides still, correctly, note that overclocking can affect warranty status and long-term lifespan. In the United States the Magnuson-Moss Warranty Act means a manufacturer cannot void your entire warranty simply because you ran a tool; what they can do is decline a claim for damage they can demonstrate the overclock caused. That is a much narrower thing than the forums imagine, and it is precisely why the very first steps of this guide are about writing numbers down. Documentation is not bureaucracy here — it is the difference between \"the fan bearing failed\" and \"you ran the memory junction at 104°C for eight months.\"

The practical safety rule that the more sober 2026 writeups converge on is simple: keep sustained GPU temperatures below 80°C to reduce longevity risk, and avoid voltage tweaks unless they are absolutely necessary. Heat is the thing that ages silicon, solder, and thermal pads, and voltage is the thing that makes heat. If you internalise only one sentence from this section, make it that one.

Why undervolting is often the smarter move

Here is the heresy that every honest 2026 guide now leads with: for most people, the better project is undervolting, not overclocking. Instead of pushing more voltage to chase raw clocks, you find the lowest voltage that holds the card's rated clock, which drops power draw, temperature, and fan noise — frequently for identical or near-identical performance. It is the same voltage-frequency curve-editor skill set pointed in the opposite, cooler direction, and it is the reason we treat that curve as a first-class tool later in this guide rather than a footnote. If your goal is performance-per-watt rather than a leaderboard screenshot, read our companion walkthrough on shaving 50 mV off a CPU in twelve steps; the philosophy transfers directly, and the discipline it teaches — one variable, test, log, repeat — is the same discipline that makes a GPU overclock stick.

Prerequisites: Software, Drivers, Hardware

The software stack

Install everything before you change anything. Nothing derails a tuning session faster than discovering, three hours in, that you never had a stress test that hits memory. The 2026 consensus stack is short, free, and specific. Install a modern tuning utility first — the guides are near-unanimous on MSI Afterburner, with at least one 2026 walkthrough naming MSI Afterburner 4.6.7 explicitly and pointing you at either MSI's own page or Guru3D for the download. Grab it from the official MSI Afterburner page or the Guru3D mirror; both bundle RivaTuner Statistics Server, which draws your on-screen overlay and per-game profiles.

The rest of the shopping list:

• HWiNFO — the reference sensor tool. It logs peak and sustained core clock, edge temperature, hotspot, memory-junction temperature, and total board power. Get it from hwinfo.com.
• GPU-Z — for reading your exact vBIOS version and confirming which clocks are actually applied, from TechPowerUp.
• FurMark — the classic \"power virus\" stress test that pins the card at maximum load, from Geeks3D.
• Unigine Superposition — a graphics-heavy loop benchmark that doubles as a stability probe, from Unigine.
• 3DMark Time Spy — your repeatable scoring yardstick, from UL Benchmarks.

You want at least two synthetic tests because they stress different parts of the card. FurMark hammers thermals and power; Time Spy and Superposition exercise the render pipeline the way a game does, and they hand you a score, which is the number that will quietly expose a memory overclock that has gone backwards.

Drivers: the version that matters

A buggy driver looks exactly like an unstable overclock, and you will chase a phantom for hours if you skip this. For NVIDIA cards in 2026, one widely-cited guide recommends being on driver 595.76 or later before you tune, while explicitly warning away from the older 595.59 and 595.71 releases, which had their own instability unrelated to anything you do. Update from the official NVIDIA driver page, do a clean install, reboot, and only then begin. If you are on Radeon, the same principle holds: sit on a current, known-good driver, not the one that shipped a week ago and is still generating crash reports across the forums.

Hardware requirements and headroom

Overclocking is, in the end, a thermal and electrical problem wearing a software costume. Three hardware realities decide your ceiling. First, PSU headroom: raising the power limit in Step 3 literally lets the card pull more watts, and transient spikes can trip an already-marginal supply. If your PSU is sized with no margin, a bigger unit is the real upgrade, not a slider. Second, case airflow and ambient temperature: the same overclock that is rock-solid at 19°C in January can throttle in a warm room in August, so tune at the ambient you actually game in. Third, card sag: a heavy GPU drooping in its slot stresses the PCIe connector and can subtly worsen die and pad contact, which shows up as a stubborn hotspot. Fix that before you fight the thermals — our guide to a GPU support bracket that kills sag in fifteen minutes is a cheaper win than any overclock and removes a variable you would otherwise blame the silicon for.

Steps 1–2: Record a Stock Baseline

Step 1 — Run the stock benchmark

Before you change a single value, record your card's default 3DMark Time Spy or Unigine Superposition score, its peak temperature, and its peak power draw. This is the most-skipped step and the most important one, because an overclock is defined entirely by the delta from this number. Without a baseline you cannot prove a gain, you cannot detect a regression, and — critically — you cannot notice the day your memory overclock quietly started making the card slower while remaining perfectly \"stable.\" Run each benchmark twice, discard the first (it warms the card and the shader caches), and keep the second. Write it into a plain text file that lives next to your profile:

=== STOCK BASELINE (record BEFORE any change) ===
GPU / vBIOS      : <your card + vBIOS from GPU-Z>
Driver           : 595.76 (NVIDIA)     # never 595.59 / 595.71
Ambient room     : 22 C
--- 3DMark Time Spy (stock) ---
Graphics score   : ____   # this is YOUR number, not a target
--- Unigine Superposition 1080p Extreme (stock) ---
Score            : ____
--- Peak sensors during the run (HWiNFO) ---
GPU core clock   : ____ MHz
GPU temp (edge)  : ____ C
GPU hotspot      : ____ C
GPU mem junction : ____ C
Board power      : ____ W

Step 2 — Log temperature and power

Open HWiNFO in sensors-only mode and let it run in the background during the benchmark. You care about the peak column for temperatures and power, and you care about the sustained clock the card settles at once it is hot, because that settled clock — not the momentary boost spike you see for the first two seconds — is what you are actually trying to raise. A plausible stock snapshot on a hot Time Spy run looks like the shape below. The absolute numbers are illustrative only; your card, cooler, and room will produce entirely different figures, and the point of the exercise is that these become your reference, not a score to hit.

# HWiNFO64 sensor snapshot, PEAK column, mid stock Time Spy run
# (illustrative shape only -- your values WILL differ)
GPU Clock ................ 2775  MHz
GPU Memory Clock ......... 1313  MHz   # GDDR6X, double data rate
GPU Temperature (edge) ... 67    C
GPU Hot Spot ............. 79    C
GPU Memory Junction ...... 78    C
GPU Power (Total Board) .. 331   W
GPU Fan .................. 61    %

Why the baseline is non-negotiable

Beyond the warranty-documentation argument from the first section, the baseline is your instrument. Every subsequent step is a comparison against it: a core step \"passes\" only if the score went up and nothing crashed, and a memory step \"passes\" only if the score went up. If you did not capture the starting score, you have no way to tell a real gain from a rounding error, and no way to catch the single most counter-intuitive failure mode in the whole process — which is the memory overclock that keeps running but keeps losing frames. Save the file. You will read it again in an hour.

Step 3: Max the Power Limit First

Step 3 — Drag the slider to maximum

The safest first tuning step, and the one nearly every 2026 guide opens with, is to raise the Power Limit slider to its maximum allowed value before you go anywhere near core or memory clocks. In Afterburner it is the slider usually labelled \"Power Limit (%)\"; drag it to whatever ceiling your card's BIOS permits — often something like 110% or 120%, sometimes exactly 100% on a locked reference board — and click apply. If your utility links the power and temperature limits, raise the temperature limit alongside it, within reason, so the card stops throttling for power before it has a chance to show you its real clock behaviour.

Why this step cannot crash you

This is the one adjustment that carries essentially no instability risk, and understanding why teaches you how the whole system works. Raising the power limit does not force the card to run faster; it only removes a constraint. The boost algorithm was already trying to clock as high as voltage and temperature allowed, and on many cards it was hitting the power ceiling first — backing off not because the silicon was unstable but because it had run out of watts. Lift that ceiling and the card is simply permitted to hold its existing boost longer and reach into clocks it already knew were safe. Nothing about that can produce an artifact, because you have not asked for any clock the card had not already validated for itself.

What to expect

Re-run your benchmark. Expect higher sustained clocks, higher temperatures, higher board power, and frequently a small score bump for free — all before you have touched a single clock offset. This is also your early thermal warning. If maxing the power limit alone pushes you toward or past 80°C sustained, stop and fix cooling now — improve airflow, build the fan curve from Step 9 early, deal with sag — because every clock you add from here makes more heat, and a card that is already thermal-limited will simply throttle away whatever you dial in.

Steps 4–6: Core Clock in +25 MHz Steps

Step 4 — The first +25 MHz

Now the real work. Recommended core-clock adjustments in 2026 are deliberately small: typically +10 MHz to +25 MHz per step, with a stability check after every single change. Enter +25 on the Core Clock offset, apply, and run a short probe — five minutes of Superposition or a single Time Spy loop is plenty at this stage. The reason for tiny steps is not superstition; it is isolation. When the card eventually fails, you want the failure pinned to a 25 MHz window so you know exactly where the edge is. Take a +200 MHz leap and all you have learned is that the answer lies somewhere in a 200 MHz haystack.

Steps 5–6 — Climb and probe

Repeat: +25, apply, short test, log the result, +25 again. Some guides permit larger +50 MHz steps purely for fast probing — a legitimate shortcut to get into the rough neighbourhood quickly — but if you do that, drop back to +25 MHz granularity as soon as you are close to the edge, because the last 50 MHz is exactly where precision pays off. Keep a running log so you never have to guess what you already tried:

# One variable at a time. Test after EVERY line. Stop at the first failure.
Core offset   Result
+25           PASS (Superposition 5 min, no artifacts)
+50           PASS
+75           PASS
+100          PASS
+125          PASS
+150          PASS
+175          driver TDR during Time Spy   <-- first failure
# Back off per the rule of thumb (10-25 MHz):
+150          PASS again  -> candidate core offset

Reading the first artifact

The rule of thumb, unchanged for a decade and still front-and-centre in 2026 guides, is: stop at the first artifact or crash, then back off. When you hit it, dial back 10 to 25 MHz from the failing value and re-test to confirm the lower clock is genuinely solid. A core failure almost always looks like a hard event, not a visual one: a black screen followed by the desktop returning with a \"display driver stopped responding and has recovered\" notification (a TDR — timeout detection and recovery), an outright application crash, or a full system reset. That is distinct from the speckled, flickering artifacts that signal a memory problem, which is why you tune the two separately and in this order. Confirm your candidate core offset with a slightly longer run before you move on; a clock that survives five minutes but dies at fifteen is not your clock.

Steps 7–8: Memory Overclock, +50–100 MHz

Step 7 — Add memory only after the core is stable

With a confirmed core offset locked in, add the memory overclock — and only now. Common 2026 step sizes are +50 MHz to +100 MHz per increment, applied and tested the same way as the core: one change, one short check, one log entry. Memory tends to tolerate larger absolute offsets than the core, so you may find yourself hundreds of megahertz in before anything protests, but resist the urge to jump. The reason memory comes second is pure variable isolation: if you had raised core and memory together and the card crashed, you would have no idea which one to blame, and you would have to unwind both.

Step 8 — The error-correction trap

Modern GDDR6 and GDDR6X carry on-die error detection and correction. This is wonderful for reliability and treacherous for overclocking, because when you push memory past its real limit it very often does not crash. Instead it silently detects errors and re-transmits the data, and the cost of all that retrying is performance that goes backwards while the system stays perfectly stable. This is the failure mode that catches people who test only for stability and never look at their score — they proudly run +1500 on the memory, pass every stress test, and never notice they gave back four percent. This is why Step 1's baseline exists. Watch the benchmark score, not just the absence of crashes:

# Memory: +50 to +100 MHz steps AFTER the core is locked.
# Watch the SCORE TREND, not just stability. Error-correction hides the wall.
Mem offset   Score trend vs. previous   Verdict
+200         up                          good
+400         up                          good
+600         up                          good
+800         flat                        marginal
+1000        DOWN                        past the wall (EDC retransmitting)
# Roll back to the last offset that still GAINED:
+600         (best clean result)         final memory offset

Finding the memory wall

The technique, then, is to climb until the score stops rising, then step back to the highest offset where it was still clearly gaining, ignoring the tempting territory beyond it where the number is stable but slower. If you also start seeing visual artifacts — coloured dots, flickering textures, geometry spikes — you have found the harder wall where correction can no longer keep up, and you are already well past the useful clock. Either way the answer is the same: retreat to the last clean gain and confirm it with a longer run. Memory tuning is the one place where \"stable\" is a trap and \"faster\" is the only verdict that counts.

Step 9: Build a Fan Curve

Step 9 — Set the curve

Fan-curve tuning remains a major part of the process in 2026, because a stable clock that thermal-throttles is, in practice, a slower clock than a lower one that never gets hot. In Afterburner, enable the software \"user defined\" fan curve and shape it against edge temperature. The example profile the 2026 guides suggest lands around 70% fan speed at 70°C and 85% at 80°C — aggressive enough to keep the card out of throttle territory, gentle enough at idle that it is not screaming while you read email. Spell it out as a set of points:

# Software fan curve (GPU edge temp -> fan duty). Steeper = cooler + louder.
30 C  ->  30 %
40 C  ->  35 %
50 C  ->  45 %
60 C  ->  55 %
70 C  ->  70 %     # 2026 guide target
75 C  ->  78 %
80 C  ->  85 %     # 2026 guide target
85 C  -> 100 %     # you should almost never reach this

Acoustics versus thermals

The whole exercise is a negotiation between noise and heat, and the right answer depends on your tolerance and your case. Steepen the curve and you buy cooler, more consistent clocks at the cost of an audible fan under load; flatten it and you get silence and throttling. The one non-negotiable, echoing the safety guidance from the top of this guide, is to keep sustained temperatures below 80°C. A brief peak to 81°C during a benchmark spike is not a crisis; a card that lives at 84°C through a two-hour raid is aging faster than it should. Tune the curve so that under your worst realistic load, the temperature plateaus in the mid-to-high 70s rather than climbing until it hits the thermal limit and clocks fall off a cliff.

Memory junction and the hotspot

Two temperatures matter that the fan curve does not directly target: the die hotspot and the memory-junction temperature. A memory overclock in particular drives the junction temperature up, and on some cards it is the first thing to reach an uncomfortable number even when the edge sensor looks calm. If HWiNFO shows the junction creeping into the 90s while the edge sits comfortably in the 60s, that is your airflow and pad contact talking, not your fan duty — and it is another argument for fixing sag and case ventilation before blaming the silicon. Watch all three sensors; the edge temperature is the one the fan curve controls, but it is not the one that fails first.

Step 10: Voltage vs. Undervolting

Overvolting — small, or not at all

You can, on many cards, add a little core voltage to push a slightly higher stable clock. The 2026 guidance on this is uniformly cautious and you should be too. A safety-focused approach recommends very small voltage changes only — on the order of +0.02 to +0.04 V — and not exceeding roughly 1.15 V on some cards. The broader advice is blunter: avoid voltage tweaks unless they are absolutely necessary, because voltage is the fastest route to heat, and heat is the thing that shortens hardware life. The honest arithmetic is that overvolting typically buys you the last 15–30 MHz for a disproportionate thermal and longevity cost. If a leaderboard is your goal, fine; if a quiet, durable daily driver is your goal, this is the step to skip.

The curve editor

Both overvolting and its far more useful cousin live in the voltage-frequency curve editor, opened in Afterburner with Ctrl+F. The graph plots the clock the card will run at each voltage point. To overclock through the curve — the more controlled method than the flat offset slider — you grab a point at a given voltage, drag it up to the clock you want, and then flatten every point to the right of it so the card never asks for more voltage than that. To undervolt, you do the same thing in reverse: pick a lower voltage, set your target clock at that point, and flatten the curve to the right so the card holds your clock without ever climbing to the higher, hotter voltages it would normally use.

Undervolting for performance-per-watt

Several 2026 writeups frame undervolting as the preferable alternative to heavy overvolting, and for good reason: it delivers better performance-per-watt instead of chasing raw clocks, which for most users in a warm room is the outcome they actually wanted. The practical recipe is to choose a voltage in the efficient part of the curve — many cards are happy somewhere around 900–950 mV — pin your desired clock at that point, flatten everything to the right, and then stress-test exactly as you would an overclock. Done well, you land at stock-or-near-stock performance with a card that runs ten or more degrees cooler and noticeably quieter. It is the same discipline as the overclock, and if you have already worked through our CPU undervolting walkthrough the muscle memory carries straight over. No overclock, incidentally, closes a generational gap — if you are hoping tuning will make a card behave like the next tier up, settle that question with an honest comparison first, the way we did in our RTX 5080 versus 4080 breakdown, rather than voltage.

Steps 11–12: Stability Testing and Final Validation

Step 11 — Short checks after every step

Stability testing in 2026 is not one event at the end; it is woven through the whole process. After each individual change in Steps 4 through 10 you run a short check — a few minutes of Superposition or a single Time Spy loop — because the point of the short check is only to catch an obvious failure before you build more on top of it. The toolset the guides converge on is FurMark for raw thermal and power stress, 3DMark and Unigine Superposition for render-pipeline stress with a score attached, and, eventually, real games. Different tests find different failures, which is exactly why you use more than one.

Step 12 — The 30–60 minute synthetic pass

Once you have a candidate profile — confirmed core offset, confirmed memory offset, fan curve in place, voltage decided — subject it to a proper synthetic soak. The 2026 recommendation is a final synthetic test of at least 30 to 60 minutes, long enough for the card to reach full heat-soak and for the memory-junction temperature to plateau at its real worst case. Many overclocks that pass a five-minute probe fail somewhere in minute twenty, when the card is thoroughly hot and the boost algorithm is operating in a thermal regime the short test never reached. Watch for a single artifact, a single driver reset, or a score that has quietly drifted downward — any of which sends you back to trim the offending clock.

The 1–2 hour gaming validation

Synthetics are necessary and insufficient. One 2026 guide is explicit that the real final pass is one to two hours of actual gaming after the synthetic stress tests, because game workloads reveal instability that benchmarks simply never trigger. A benchmark is a fixed, repeatable loop; a game throws load transitions, wildly varied shaders, ray-tracing spikes, streaming stalls, and sudden idle-to-full swings that no synthetic reproduces. The classic signature is a profile that aces every benchmark and then hard-crashes forty minutes into one specific title. If that happens, it is not the game's fault — it is your last 15–25 MHz of core clock, and the fix is to trim it and re-validate. Only after a couple of hours of real play across a game or two would a careful person call an overclock \"done.\"

Five Common Pitfalls (and Fixes)

Every failed overclock is a failure of process, not of nerve. These are the ones that recur.

1. Pitfall: skipping the baseline. You max everything, it feels faster, and you have no idea whether you gained three percent or lost two to memory error-correction. Fix: record the stock Time Spy or Superposition score, peak temperature, and peak power before touching anything, and compare every step against it.
2. Pitfall: moving core and memory together. The card crashes and you cannot tell which clock did it, so you unwind both and start over. Fix: lock a stable core offset first, then and only then begin the memory climb, changing exactly one variable at a time.
3. Pitfall: trusting a five-minute stress test. The profile passes every synthetic and then dies in a real game after half an hour. Fix: follow the short checks with a 30–60 minute synthetic soak and then a 1–2 hour gaming validation, because game workloads expose instability benchmarks miss.
4. Pitfall: chasing the last 30 MHz with voltage. You add core voltage for a marginal clock and buy a large heat and longevity penalty. Fix: keep any voltage change tiny (+0.02 to +0.04 V, and not past ~1.15 V on cards where that applies), or skip it entirely and consider undervolting instead.
5. Pitfall: ignoring the fan curve. A perfectly stable clock throttles under sustained load and ends up slower than a cooler, lower one. Fix: build a curve around 70% at 70°C and 85% at 80°C, and keep sustained temperatures below 80°C.
6. Pitfall: never saving the profile. A reboot or a driver update wipes your evening's work. Fix: save the Afterburner profile, enable \"apply at startup,\" and keep the baseline text file with your final offsets written into it.
7. Pitfall: tuning on a bad driver. You spend hours fighting instability that was never yours. Fix: update to a known-good release (NVIDIA 595.76 or later; avoid 595.59 and 595.71) before you begin, and re-validate after any future driver change.

The Troubleshooting Table

Symptoms map to causes more reliably than beginners expect. Find yours, apply the fix, re-test.

The Complete Working Configuration

The Afterburner profile

Here is the whole result assembled in one place, as an exported Afterburner profile. Read the values as the UI offsets in kilohertz, the way the tool stores them (a +150 MHz core offset is written 150000). Treat the specific numbers as an example tune for a hypothetical high-end card carried through this guide's stepping logs — your card will land somewhere else entirely, and the only correct values are the ones your own testing produced.

; MSIAfterburner Profile (exported). Values are UI OFFSETS in kHz.
; EXAMPLE tune only -- derive your own from your stepping logs.
[Profile1]
Format=2
Flags=0
CoreClkBoost=+150000     ; +150 MHz core (last confirmed-stable step)
MemClkBoost=+600000      ; +600 MHz memory (last CLEAN score gain)
PowerLimit=112           ; slider maxed in Step 3
ThermalLimit=84          ; keep sustained temps in the 70s
ThermalPrioritize=0
FanMode=2                ; 2 = custom software fan curve
; Undervolters: keep CoreClkBoost modest and pin the V/F curve instead.

The fan curve, spelled out

The fan curve is stored with the profile but is worth keeping in your notes as plain points, because it is the part you will re-tune when the seasons change and your ambient temperature moves ten degrees.

# Daily-driver fan curve (edge temp -> duty). Retune when ambient shifts.
30 C -> 30 % | 40 C -> 35 % | 50 C -> 45 % | 60 C -> 55 %
70 C -> 70 % | 75 C -> 78 % | 80 C -> 85 % | 85 C -> 100 %

The Linux equivalent

None of this requires Windows. On Linux you enable the overclocking controls through Coolbits, set the power limit with nvidia-smi, and apply clock offsets with nvidia-settings. The one gotcha that trips everyone: the memory control on Linux is a transfer-rate offset, which is double the clock offset — so a +600 MHz memory clock is entered as +1200 here.

# --- Linux (X11) overclocking via Coolbits ---
# 1) Enable the OC controls, then reboot:
sudo nvidia-xconfig --cool-bits=28

# 2) Raise the power limit (within the min/max shown by 'nvidia-smi -q'):
sudo nvidia-smi -pl 350

# 3) Apply tested offsets. Memory is the DATA RATE = 2x the clock:
nvidia-settings -a '[gpu:0]/GPUGraphicsClockOffsetAllPerformanceLevels=150'
nvidia-settings -a '[gpu:0]/GPUMemoryTransferRateOffsetAllPerformanceLevels=1200'

# 4) Watch it live, once per second:
nvidia-smi --query-gpu=clocks.gr,clocks.mem,temperature.gpu,power.draw \\
           --format=csv -l 1

Advanced tips for the last two percent

Once the fundamentals are solid, a few refinements separate a decent tune from a considered one. Use RivaTuner Statistics Server to build per-game profiles — a bandwidth-bound title benefits from every megahertz of memory you can hold, while a shader-bound one wants the core, and there is no rule that one profile must serve both. When probing a fresh card, take the +50 MHz fast-probe shortcut to find the rough core ceiling quickly, then drop to +25 MHz granularity for the last 100 MHz where precision decides the outcome. Consider the efficiency tune — a modest undervolt combined with a small core offset — which frequently beats a raw overclock on sustained clocks precisely because it runs cooler and never throttles. Always validate at your real summer ambient, not the artificially cool conditions of a late-night session, and watch the hotspot and memory-junction sensors rather than only the edge temperature, since those are what fail first. And resist the deeper rabbit holes — vBIOS flashing, shunt mods, aftermarket power tools — unless you fully accept that they move you decisively out of the safe, self-clamping zone this guide has kept you inside. If you find yourself weighing them just to catch the next tier of card, that is your answer that you wanted a different GPU, not a different profile.

The whole thing, in one breath

Record a baseline; update the driver; max the power limit; climb the core in +25 MHz steps until it crashes and back off; add memory in +50–100 MHz steps while watching the score, not just stability; build a fan curve for 70°C/70% and 80°C/85%; skip or barely touch voltage; then soak it for 30–60 minutes of synthetics and 1–2 hours of real games before you trust it. Save the profile, keep sustained temperatures under 80°C, and write down what you did. That is the entire craft. It is unglamorous, it is roughly two hours of hands-on patience, and it is the only version of \"how to overclock GPU\" that produces a result still standing next week.