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

Somewhere on the internet right now, a freshly minted guide is telling you to install NVIDIA driver 595.76, grab the 4.6.7 stable build of MSI Afterburner, and leap your core clock in +25 MHz bounds until glory arrives. Two of those three claims are wrong and the third is a coin flip. There is no 595.76 branch on the current tree; NVIDIA has been shipping the 610 series since spring, and 610.88 WHQL landed on 28 July 2026. And 4.6.7 is a beta, not a stable. This is the state of overclocking content in 2026: confidently specific, frequently fabricated, and written by things that have never once watched a benchmark score climb and then quietly fall.

This guide does the boring, correct version instead. Twelve numbered steps, real version numbers, and an honest ceiling. On a modern card you are fighting for five to ten percent, sometimes less, because the silicon already overclocks itself several thousand times a second while you are not looking. We will max the two sliders that actually matter, climb the core clock in small rungs, tune the memory until its error-correction logic betrays it, soak the result for an hour, validate it in real games, and then, because it is 2026 and heat is the only real enemy, probably talk you into undervolting instead. Bring a card, a benchmark, and some patience. Leave the fake driver numbers at the door.

GPU Overclocking in 2026: The Reality Check

Before you touch a single slider, calibrate your expectations, because the gap between what overclocking promises and what it delivers is where all the disappointment lives.

The Five-to-Ten-Percent Truth

A modern graphics card is not a 2010 GPU with a fixed clock and a fat margin waiting to be claimed. It is a self-tuning device that already runs as fast as its power budget, temperature, and voltage curve allow, recalculating that decision thousands of times per second. NVIDIA's own one-click Automatic Tuning inside the NVIDIA App nets you roughly three to five percent on average, takes ten to twenty minutes, and does not void your warranty. A careful manual pass, the twelve steps below, typically buys five to ten percent, and on a genuinely lucky sample maybe twelve. If someone promises you thirty, ask what they are selling. The honest framing: you are not unlocking hidden power, you are renegotiating terms the card already set for itself, and it drives a hard bargain.

Silicon Lottery and the Guardband

Every die that comes off the wafer is tested and binned, and the vendor sets stock clocks with a guardband, a deliberate safety margin so the worst chip in the batch still works at forty degrees ambient, three years from now, under a dusty cooler. Overclocking is the act of reclaiming part of that guardband. How much you get back is the silicon lottery: two identical SKUs on the same shelf can differ by a hundred megahertz at the same voltage because their transistors leak differently. This is the same physics that governs the inverse move, undervolting, where you spend the margin on lower voltage rather than higher clocks. It is the P = C x V squared x f relationship we walk through in the CPU undervolting guide, where voltage is squared and therefore dominates the heat equation. Your card, your die, your lottery ticket. Nobody else's numbers are yours, which is exactly why you record a baseline instead of chasing a stranger's screenshot.

What Changed: The Boost Algorithm Fights You Now

The reason a +150 MHz offset does not add 150 MHz to the number on your screen is that you are not setting a clock. You are nudging a curve. Boost, which NVIDIA brands as GPU Boost and AMD implements without the marketing, is a closed control loop that reads temperature, power draw, and voltage, then picks the highest frequency on its voltage/frequency curve that fits inside all three ceilings at once. Your overclock is an offset applied to the entire curve. If the card is already pinned against its power limit, shifting the curve upward does nothing until you also raise the power ceiling, which is precisely why Step 4, not Step 5, is where most of your free performance actually comes from.

Prerequisites: Versions, Tools, and Headroom

Half the instability people blame on a bad overclock is actually a stale driver, a beta tool, or a case with the airflow of a sealed lunchbox. Fix the foundation first.

The Software Stack, With Exact Versions

As of August 2026 the current NVIDIA Game Ready driver is 610.88 WHQL, released 28 July 2026; the 610 branch is what your RTX 40 or 50 card wants, and any guide citing 595.76 is quoting a version that does not exist on the current tree. AMD users want the current Adrenalin WHQL release. WHQL means the driver is signed and passed Microsoft's lab, which matters because a beta driver bug reads exactly like an unstable overclock and will waste an hour of your life. The overclocking tool is MSI Afterburner 4.6.6 Final, the first stable build in roughly two years, shipped October 2025, which added RTX 50 and Radeon RX 9000 support and bundles RivaTuner Statistics Server 7.3.7 for the on-screen display. There is a 4.6.7 Beta 3 (build 17352, June 2026) with a nicer voltage/frequency curve editor; install it only if you specifically want the curve tools, and know that it is a beta. Download Afterburner from MSI or Guru3D and nowhere else. The search results are wall-to-wall SEO farms repackaging it with invented step counts and, occasionally, a payload.

Toolchain — verified August 2026
--------------------------------------------------
GPU-Z             latest (TechPowerUp)   identify silicon
NVIDIA driver     610.88 WHQL (28 Jul 26) or newer 610-series
AMD Adrenalin     current WHQL release
MSI Afterburner   4.6.6 Final (stable)   bundles RTSS 7.3.7
   4.6.7 Beta 3   build 17352 (Jun 26)   beta only, V/F editor
Unigine Superposition  1.1 (free Basic edition)
FurMark 2         2.10.2 (Geeks3D)
OCCT              latest (VRAM error test)
HWiNFO64          latest (sensor logging)

You also want a way to see your silicon and beat on it. GPU-Z confirms the exact model, VRAM type, and stock clocks; Unigine Superposition gives you a repeatable score for the climb; FurMark 2 and OCCT handle the thermal soak and the VRAM error test respectively. The method itself, if you want a second authority to check my work against, follows Tom's Hardware's overclocking walkthrough almost exactly.

Hardware and Thermal Requirements

Raising the power limit means the card draws more watts, so your PSU needs headroom for the higher sustained draw and the nasty transient spikes modern GPUs produce, which can momentarily double the rated figure. A +15 percent power limit on a 350 W card asks for roughly 400 W steady and considerably more in bursts; a marginal supply will trip its own protection and reboot the machine under load, and you will blame the overclock. You also need airflow. Overclocking inside a hotbox with two intake fans and a nest of cables is a way to discover thermal throttling, not performance. And if your card droops at the PCIe slot like a wilting houseplant, sort that out first, because a sagging card stresses the slot and can shift the die's contact with the cooler; the GPU support bracket walkthrough covers the twenty-minute fix.

A Word on Warranties and the Law

The Machine reads the fine print so you do not have to. Software overclocking through Afterburner does not, on its own, void your warranty, and in the United States the Magnuson-Moss Warranty Act means a manufacturer cannot void your entire coverage simply because you moved a slider; they must show the modification caused the failure. NVIDIA's App-based Automatic Tuning is explicitly warranty-safe. The line you do not cross is flashing a modified vBIOS to lift hard voltage or power caps, which alters the card's firmware and is exactly the kind of thing a denied RMA is built on. Keep it in software, keep your receipts, and keep your expectations legal.

How Modern GPU Boost Actually Works

You cannot tune a system you do not understand, and the single most useful thing to internalize is that your card is not waiting for permission. It is already trying to run as fast as physics allows.

The Four Ceilings

At any instant the boost algorithm is choosing the lowest of four ceilings and clocking to it. There is the power ceiling measured in watts, the temperature ceiling measured in degrees, the voltage/frequency ceiling set by the card's internal curve, and a reliability ceiling the vendor bakes in and never shows you. Whichever it hits first is the one dragging your clocks down. This is why Step 4 raises the power and temperature ceilings before you ever touch a clock offset: on the vast majority of cards, especially Founders Edition and blower designs, power is the first wall, and simply moving that wall is the biggest safe gain available. You are not making the card faster so much as removing the reason it was holding back.

Why Setting a Clock Is a Lie

The core clock slider in Afterburner is labeled in megahertz, which fools people into thinking they are dialing an absolute frequency. They are not. The value is an offset added to every point on the voltage/frequency curve at once. Set +150 and the whole curve shifts up by 150, but the card still picks the point that fits inside the four ceilings, so your observed clock might rise by 150, or by 40, or by nothing if you are power-limited. The memory slider is also an offset, and on NVIDIA cards it is expressed on the data-rate scale, so a +1000 there is not as wild as it sounds; it corresponds to a far smaller change in the underlying clock. Read the offset as a request, not a command.

Reading the Telemetry

You tune by watching sensors, not by faith. Set up the RTSS overlay or HWiNFO64 to show core clock, memory clock, GPU edge temperature, hotspot temperature, memory-junction temperature, power in percent of TDP, voltage, fan speed, and, most usefully, the PerfCap reason, which tells you which of the four ceilings is currently limiting you. Pwr means power, Thrm means temperature, Vrel means the voltage/reliability curve, and Util means the workload simply is not asking for more. If PerfCap says Pwr, no core offset will help until you raise the power limit.

RTSS / HWiNFO on-screen readout (example, load)
-----------------------------------------------
GPU Clock       2865 MHz
Mem Clock      10502 MHz  (effective ~21 Gbps)
GPU Temp          71 C
Hotspot           84 C
Mem Junction      86 C
Power            102 % TDP  (398 W)
Voltage        1.050 V
Fan               78 %
PerfCap          Pwr        <-- limited by power, not heat

Steps 1-4: Model, Driver, Baseline, and Power

The first four steps are the ones amateurs skip and then wonder why their results are unrepeatable. Do them in order.

Steps 1-2: Identify the Card, Update the Driver

Step 1 — Confirm your exact GPU. Open GPU-Z and read the model, the VRAM type and size, the stock boost clock, and the BIOS version. The rationale is that overclocking is card-specific: NVIDIA and AMD have different power headroom, different voltage behavior, and different stability thresholds, and even within a brand a GDDR7 RTX 50 behaves nothing like a GDDR6 RX 9000. You cannot follow generic numbers; you can only follow your own silicon. Write down the stock boost clock, because that is the number every later gain is measured against.

Step 2 — Install the current WHQL driver. Clean-install NVIDIA 610.88 WHQL or newer, or the current AMD Adrenalin, using the installer's clean-install option or DDU if your driver stack is a mess. The rationale is blunt: driver bugs masquerade as overclock instability, and a brand-new card on an old driver will hand you phantom crashes you will spend hours chasing at the wrong slider. New silicon needs new drivers. Start from a known-good, signed baseline.

Step 3: Install the Toolchain

Step 3 — Install Afterburner 4.6.6, RTSS, and a benchmark. Put down Afterburner 4.6.6 Final (RTSS 7.3.7 rides along in the same installer), plus Superposition and HWiNFO64. The rationale is that this is the reference stack every credible guide's numbers assume, including this one and Tom's Hardware's. Use anything else and you are comparing your results against measurements taken on different tools, which is no comparison at all. Reboot after install so the kernel-mode monitoring driver loads cleanly.

Step 4: Baseline, Then Max Power and Temp

Step 4 — Record a stock baseline, then raise the two ceilings. First, run your benchmark at stock and write down the score, the average and maximum temperatures, the average sustained clock, and the peak power draw. The rationale is arithmetic: without a before, there is no measurable after, and you have no way to know whether an offset helped or merely felt like it did. Then, and only then, drag the Power Limit slider to its maximum and, on cards where it is separate, drag the Temp Limit to maximum and link the two so they move together. Typical power headroom runs from about +10 percent to +25 percent depending on the card; the AMD RX 9000 series is hard-capped at +10 percent in Adrenalin, while many NVIDIA cards allow +15 percent or more. Re-run the same benchmark. You will frequently see a two-to-four percent gain from this alone, before a single clock offset, because you just moved the wall the card was leaning on.

Superposition 1080p Extreme — baseline vs power-maxed
-----------------------------------------------------
                     Score   Avg FPS  Max Temp  Avg Clock
Stock                 9847     73.6      74 C     2760 MHz
Power +15%, Temp max 10190     76.2      76 C     2850 MHz
                    (+3.5%)              (no core offset yet)

Steps 5-7: Core Clock, +15 MHz at a Time

Now the actual overclock. Patience here is not a virtue, it is the entire method, because the only way to find your edge is to walk up to it one small rung at a time.

Step 5: The +15 MHz Ladder

Step 5 — Add +15 MHz to the core clock offset. Type +15 into the core clock field and apply. The rationale is resolution: a small step isolates the exact rung where instability begins, and a large jump only tells you that somewhere between +100 and +200 things broke, which is useless. The 2026 guides scatter across a range here, with the conservative camp recommending +10 MHz per step and the aggressive camp using +25 MHz. Fifteen is the Tom's Hardware convention and the sensible balance: fine enough to find the edge, coarse enough to get there before your coffee is cold. If you want maximum caution, use +10 and expect to click more.

Step 6: Test Every Rung

Step 6 — Test after every single step. After each +15 MHz, run a short stress loop, a single Superposition pass or five to ten minutes of a looped benchmark, and watch the screen. The rationale is that instability compounds and hides; catch it at the rung where it starts and you know your ceiling exactly. You are watching for three failure signs: visual artifacts (stray colored dots, flickering or corrupted textures, geometry spikes), an outright application crash, or a black screen that recovers itself accompanied by the Windows message Display driver stopped responding and has recovered, which is a TDR, a timeout-detection-and-recovery reset. A recovered TDR is Windows saving you from a hard lock; treat it as a failed rung.

Step 7: Back Off After the First Crash

Step 7 — On the first instability, back off. The moment you hit an artifact, crash, or TDR, drop the core offset back down by 15 to 30 MHz, with some guides preferring a fuller 25 to 50 MHz retreat. The rationale is margin: the exact edge is temperature-dependent, and the value that survives a cool evening test will fail on a thirty-degree afternoon three months from now with a dustier heatsink. You want last-stable-minus-a-cushion, not last-stable. On many current cards a realistic stable core result lands somewhere between +100 and +200 MHz, but yours is a lottery ticket, so trust your own ladder over any number in this paragraph.

Steps 8-9: Memory and the Error-Correction Trap

Memory is where overclocking stops being intuitive, because the failure mode is not a crash. It is a lie your benchmark tells you, and if you do not know the tell, you will happily run a memory clock that is quietly destroying your performance.

Step 8: Tune Memory After Core, in +50 MHz Steps

Step 8 — Only now raise the memory clock, in +50 to +100 MHz steps. Lock and validate the core clock first, then move to memory. The rationale, echoed across the 2026 sources, is that memory instability appears later and differently than core instability, so tuning it while the core is still in flux confounds both; isolate the variable. Step memory in +50 MHz increments (up to +100 if you are impatient), applying and testing each time. Memory has enormous headroom compared to core: on GDDR6X and GDDR7 the total offset can reach +1000 MHz or more, so feel free to take larger +100 to +200 jumps early and switch to fine +50 steps as the score stops improving. A conservative, lower-risk all-in target is +300 to +500 MHz, and if you never want to think about it again, stop there.

Step 9: The Score-Drop Tell

Step 9 — Watch the benchmark score for the drop, not for artifacts. Here is the trap. GDDR6X and GDDR7 include error detection with retransmission, conceptually the same CRC-and-resend scheme that Ethernet and Wi-Fi use. Push the memory past its sweet spot and it does not immediately throw artifacts; instead it silently detects corrupted transfers and re-sends them. Everything still works, no dots, no crash, but the re-sends cost bandwidth, so your benchmark score climbs, peaks, and then falls while stability looks perfect. That score drop is the tell. It means you have passed the useful ceiling and the memory is now spending its extra clock on cleaning up its own mistakes. Back the memory offset down until you are sitting just below the peak, roughly 15 MHz of effective clock below where the score turned over. You are hunting the highest useful memory clock, not the highest number the slider will accept. This single mechanism is why no-artifacts is not proof of a good memory overclock.

VRAM Type Matters

The reason a friend's memory numbers are useless to you starts with which memory you have, and this is where the fabricated guides give themselves away by claiming the RTX 50 series uses GDDR6X. It does not. Match your card to the table before you touch the slider, and keep an eye on the memory-junction temperature the whole time, because GDDR6X in particular runs hot enough to throttle on its own near 105 C, which will crater your score in a way that looks exactly like the error-correction drop but is not.

Steps 10-12: Stress, Validate, and Persist

You now have numbers that survive a short test. That is not the same as numbers you can trust, and the difference is measured in hours.

Step 10: The 60-Minute Soak

Step 10 — Stress test for at least sixty minutes. Run FurMark 2, OCCT's combined 3D-and-VRAM test, or a looped Superposition for a solid hour minimum; the cautious among the 2026 sources push two to three hours before trusting a profile for daily use. The rationale is thermal: a thirty-second pass runs cool, but the edge of stability shifts as the card and its VRAM heat-soak, and a clock that is rock-solid at 60 C can fold at 78 C forty minutes in. Watch for artifacts, TDRs, and thermal throttling, and if you use OCCT, watch its VRAM error counter specifically, because it will surface memory errors the benchmark score alone might hide.

OCCT VRAM + 3D — 60-minute soak (example, PASS)
-----------------------------------------------
Elapsed              01:00:00
GPU Temp max            79 C
Hotspot max             91 C
Mem Junction max        94 C
VRAM errors              0
TDR / driver resets      0
Throttling            none
Verdict            STABLE - promote to game validation

(FAIL looks like: VRAM errors 137, or one TDR at 00:42
 -> drop memory 50 MHz or core 30 MHz and re-soak)

Step 11: Real Games for 1-2 Hours

Step 11 — Validate in actual games for one to two hours. Synthetic loops are consistent, which is their strength and their blind spot. Real game engines hit clock, voltage, and memory-access patterns that a benchmark's tight inner loop never touches, so a profile that passes an hour of FurMark can still crash in one specific title's shader-heavy loading zone. Play the games you actually play, for one to two hours, and treat a single reproducible game crash as proof the overclock is too aggressive even if every synthetic test passed. This is the step that separates a stable overclock from a benchmark trophy. Once it holds, the frames are yours to spend, ideally on a display that can show them; whether that is worth chasing past a certain refresh rate is its own argument, which we settle in 144Hz vs 240Hz.

Step 12: Fan Curve, Save, and Persist

Step 12 — Build a fan curve, save the profile, and enable apply-at-startup. Open Afterburner's fan settings, enable the user-defined curve, and shape it to keep the card in its comfort zone; a widely recommended shape is 70 percent fan at 70 C and 85 percent at 80 C, with a hard ramp to 100 percent by 85 C. The rationale is that stock fan curves are tuned for quiet, and an overclock needs cooling more than it needs silence. Then save everything to a profile slot and tick Apply overclocking at system startup so the settings survive a reboot. One hard rule: enable auto-apply only after Steps 10 and 11 pass, because an unstable overclock that loads at boot can crash before the desktop appears, and digging yourself out of that is a Safe Mode chore nobody enjoys.

Afterburner custom fan curve
----------------------------
Temp (C)   Fan (%)
   40         30
   55         45
   70         70
   80         85
   85        100

The Smarter Move: Undervolting

Here is the part the overclocking guides bury, because it undercuts their own premise: on most 2026 cards, the better move is to go down, not up. The 2026 safety guides treat undervolting as the recommended third phase after baseline and overclock attempts, and they are right.

Why Down Beats Up

Recall the four ceilings. Boost is almost always limited by power or temperature, not by the frequency the chip can theoretically reach. Undervolting sets a target where the card runs the same clock at less voltage, and because power scales with voltage squared, a modest voltage cut yields a disproportionate drop in heat and watts. Less heat and fewer watts mean the card stops throttling against its ceilings, which frequently produces a higher sustained clock than stock, along with lower temperatures, lower fan noise, and lower power draw. You get the overclock's performance and cooling you did not have to pay for. It is the same trade we make on the processor side in the CPU undervolting guide, and the physics is identical.

The V/F Curve Method

In Afterburner, press Ctrl+F to open the voltage/frequency curve editor (the tooling is noticeably better in the 4.6.7 beta). Pick a voltage target, commonly somewhere around 875 to 950 mV, set the clock you want at that point, then flatten every point to the right of it so the card can never exceed that voltage. Apply and test with the same soak-and-game discipline from Steps 10 and 11. On AMD, Adrenalin exposes a simpler voltage offset in millivolts, so you can nudge it negative without the curve dance. The result you are aiming for is the same clock as your overclock, or close to it, at a voltage that keeps the card ten or fifteen degrees cooler.

When to Undervolt Instead of Overclock

If your card throttles on temperature or power, which describes most Founders Edition cards, blower designs, small-form-factor builds, and every laptop, undervolt. If you have a triple-fan cooler with genuine thermal headroom to spare, an overclock will net a little more. The best of both worlds is a mild core overclock combined with an undervolt, riding the efficiency curve for higher clocks at lower heat. Laptops and handhelds are a special case where you should undervolt or power-limit only and forget overclocking entirely; the thermal budget in a chassis that thin, discussed in the 2026 gaming laptop guide, has no room for the extra watts.

Five Pitfalls That Kill Overclocks

Every unstable overclock I have ever seen traces back to one of these. Read them now and save yourself the reboots.

The Six Ways People Wreck It

1. Trusting a thirty-second benchmark. A short pass runs cool and proves nothing; the edge moves as the card heat-soaks. The fix is the full sixty-minute soak plus one to two hours of real games from Steps 10 and 11, no shortcuts.
2. Judging memory by artifacts instead of score. GDDR error-correction hides instability by retransmitting, so a bad memory clock looks perfectly stable while your score quietly drops. The fix is Step 9: watch the score, find the peak, sit just below it.
3. Maxing everything at once. Slam power, core, and memory up together and when it crashes you have no idea which one did it. The fix is one variable at a time, in order: power ceiling, then core, then memory.
4. Enabling apply-at-startup before validation. An unstable overclock that loads at boot can crash before the desktop, turning a slider mistake into a Safe Mode rescue. The fix is to enable persistence only after the card has passed the soak and the games.
5. Ignoring the hotspot and memory-junction temperatures. The edge temperature looks fine at 72 C while the hotspot is at 95 C and the GDDR6X junction is throttling near 105 C. The fix is to monitor hotspot and memory-junction, not just the headline number, and to treat a large hotspot delta as a mounting or paste problem.
6. Installing whatever the top search result served. The fabricated guides hand you a nonexistent driver, a beta labeled as stable, and the wrong VRAM type. The fix is to pull drivers from NVIDIA or AMD directly, Afterburner from MSI or Guru3D, and to verify every version number against the source before you trust a single instruction.

Troubleshooting: Symptoms and Fixes

When it goes wrong, and on the first pass it will, match the symptom to the cause rather than flailing at random sliders.

The Table

Decoding the Crash Types

The three failure signatures each point somewhere specific. A recovered TDR (the black-screen-that-comes-back) is almost always the core clock, because the core is what stalls the render pipeline long enough to trip Windows' timeout. Visual artifacts without a crash are usually memory, or a core clock right at its edge. A hard lock or full system reboot with no recovery message points at power delivery, either the card's power limit or the PSU itself, not the clock at all. Match the signature before you touch a slider and you will fix it in one move instead of five.

Advanced Tips (and Linux)

Once the basics hold, a few refinements separate a working overclock from a considered one.

Per-Game Profiles and RTSS

Afterburner supports multiple profile slots and RTSS can auto-apply a profile when a specific game launches, which lets you run an aggressive profile on a well-behaved title and a conservative one on the game that hates your memory clock. Keep the RTSS overlay showing PerfCap while you play; if it reads Pwr in your most demanding game, you are power-limited and an undervolt will do more than any further core offset. Log a session to CSV in HWiNFO64 and you can see exactly where and why the card throttled instead of guessing.

Overclocking on Linux

There is no Afterburner on Linux, so you work closer to the metal. On NVIDIA under X11, you unlock the controls with Coolbits and then set offsets through nvidia-settings; the graphics and memory-transfer-rate offsets map to the same curve Afterburner nudges. On Wayland or for a headless setup, nvidia-smi handles the power limit directly. On AMD, you enable the full overclocking interface with a kernel parameter and write to the amdgpu sysfs node, or you save yourself the syntax and use a GUI: LACT or CoreCtrl both expose clocks, voltage, power limits, and fan curves in a window that behaves much like Afterburner.

# NVIDIA (X11) — enable the overclock controls
nvidia-xconfig --cool-bits=28
# then, per session:
nvidia-settings -a "[gpu:0]/GPUGraphicsClockOffsetAllPerformanceLevels=150"
nvidia-settings -a "[gpu:0]/GPUMemoryTransferRateOffsetAllPerformanceLevels=800"
nvidia-settings -a "[gpu:0]/GPUFanControlState=1"
# power limit (works under Wayland too):
sudo nvidia-smi -pl 400

# AMD (amdgpu) — add kernel param first:
#   amdgpu.ppfeaturemask=0xffffffff
echo "s 1 2900" | sudo tee /sys/class/drm/card0/device/pp_od_clk_voltage
echo "c"        | sudo tee /sys/class/drm/card0/device/pp_od_clk_voltage
# or skip the syntax entirely: use LACT or CoreCtrl

Laptops, Handhelds, and the Connector Aside

Laptop and handheld GPUs are usually locked to power-limit and undervolt adjustments only, which is fine, because that is what their thermals want anyway. And a word of lore for the high-end desktop crowd raising power limits on a 5090: more watts mean more current through the 12V-2x6 connector, and a connector that is not fully seated is how you get the melted-plug photographs. When der8auer put a thermal camera on a poorly seated 5090 connector, one pin was pulling enough current to hit roughly 150 C. Push the plug in until it clicks, verify it, and do not run 500-plus watts through a half-mated connector because a guide told you to max the power slider.

The Complete Working Profile

Here is a conservative, known-good starting point you can dial in and then push from, assembled from the lower-risk all-in figures the 2026 guides converge on. It is a starting line, not a finish line: validate it on your own silicon with the soak and the games before you trust it.

The Afterburner Profile

=== staresback known-good starting profile ===
=== validate before you trust it ===
Power Limit ....... +15 %     (AMD RX 9000: +10 %, hard cap)
Temp Limit ........ max (linked to power)
Core Clock ........ +75 MHz   (conservative; push toward +150 if stable)
Memory Clock ...... +400 MHz  (watch for the score-drop tell, Step 9)
Voltage ........... stock     (better: undervolt at ~900 mV instead)
Fan Curve ......... 30/45/70/85/100 % at 40/55/70/80/85 C
Apply at startup .. ON  (only AFTER a 60-min soak + 2h of real games)
Profile slot ...... 1

The Startup Checklist

The whole method in one pass, so you can run it without scrolling: (1) confirm the card in GPU-Z; (2) clean-install 610.88 WHQL or current Adrenalin; (3) install Afterburner 4.6.6, RTSS, and a benchmark; (4) baseline, then max power and temp; (5) core in +15 MHz rungs; (6) test every rung; (7) back off 15-30 MHz on the first failure; (8) memory in +50 MHz steps; (9) find the score-drop peak and sit below it; (10) soak sixty minutes; (11) two hours of real games; (12) fan curve, save, apply at startup. Then, and this is not optional if you value your thermals, seriously consider replacing the whole thing with an undervolt.

Numbers to Expect

When it is done, expect five to ten percent more performance, maybe twelve on a lucky die, or, if you undervolted, roughly the same performance at ten to fifteen degrees cooler and noticeably quieter. What you should not expect is a transformation, because the card was already doing most of this for you before you opened Afterburner. The gain is real, it is free once you own the hardware, and it is worth an evening. It is also, in 2026, a rounding error next to buying the right card in the first place. Tune what you have, temper your expectations, and never trust a guide that quotes a driver version that does not exist.