Key takeaways
- The best balance for most graphics cards in 2026 is a mild undervolt combined with a power limit set to 85–95% of stock and a fan curve that holds the core around 72–78 °C under sustained load. That combination typically…
The best balance for most graphics cards in 2026 is a mild undervolt combined with a power limit set to 85–95% of stock and a fan curve that holds the core around 72–78 °C under sustained load. That combination typically costs only 1–4% of average frame rate while cutting 10–20% of board power and 5–10 °C of core temperature. Entry-level cards are the exception — they are already power-starved at stock, so leave the limit alone and tune voltage only.
Our top picks
As an Amazon Associate we earn from qualifying purchases at no extra cost to you. Ratings are shown as listed on Amazon.
Quick comparison: what each tier should actually change
Board power figures below reflect manufacturers’ published total graphics power for reference designs; partner cards with factory overclocks often draw 5–10% more.
| Tier | Example cards | Power limit | Undervolt target | Fan target under load | Typical FPS change | Typical temp change |
|---|---|---|---|---|---|---|
| Entry / 1080p | GeForce RTX 5060, Radeon RX 9060 XT, Arc B580 | 100% (leave alone) | −50 to −80 mV | 68–72 °C | 0 to +2% | −3 to −5 °C |
| Mainstream / 1440p | RTX 5070, RX 9070, RTX 5060 Ti 16 GB | 90–95% | −60 to −100 mV | 72–75 °C | −1 to −2% | −5 to −8 °C |
| Enthusiast / 1440p–4K | RTX 5080, RTX 5070 Ti, RX 7900 XTX | 85–90% | −80 to −120 mV | 74–77 °C | −2 to −4% | −6 to −10 °C |
| Flagship / 4K | RTX 5090, RTX 4090 | 75–80% | −100 to −150 mV | 76–79 °C | −3 to −6% | −8 to −12 °C |
Why the power limit is the first lever, not the last
Modern boost algorithms sit near the flat top of the voltage–frequency curve, where every extra 100 MHz costs a disproportionate amount of voltage and therefore heat. Trimming the power limit pushes the card back down the steep part of that curve, where clock speeds are cheaper in watts. In practice, a 575 W flagship capped at 450 W usually loses low single digits in real gameplay, because frame rate scales far more slowly than power once the card is near its clock ceiling. The same logic holds at 250 W: capping to 220 W on a mainstream card costs roughly 1–2% of average FPS but removes a meaningful chunk of heat from the case.
The mistake is treating the power slider as a blunt throttle. It is better understood as a budget the boost algorithm spends intelligently — it will still hit high clocks in light scenes and only back off in heavy ones.
Step-by-step: a stable undervolt in about 20 minutes
An undervolt is a fixed voltage ceiling paired with a raised clock at that voltage; the goal is more clock per volt, not less performance.
- Log stock behaviour first. Run 30 minutes of a demanding title or benchmark loop with hardware monitoring open, recording core clock, core voltage, core temperature, hotspot temperature, fan RPM and board power. Reason: every later comparison depends on a baseline, and you need to know whether the card is thermally limited or power limited before changing anything.
- Set the power limit to your tier target from the table above. Reason: it caps worst-case heat immediately, and because voltage demand falls as boost clock headroom shrinks, it makes the undervolt easier to stabilise.
- Open the voltage–frequency curve editor (in MSI Afterburner press Ctrl+F; AMD Adrenalin and Intel’s graphics tuning software expose equivalent sliders). Pick a target voltage around 0.900–0.950 V for current Nvidia parts, or the equivalent point that sits two or three steps below stock on Radeon cards.
- Raise the clock at that voltage by roughly 150 MHz over stock, then flatten every point above it. Reason: flattening prevents the driver from silently jumping back to a higher, hotter voltage bin under load.
- Validate for 30 minutes, then validate again in a ray-traced or path-traced title. Reason: transient-heavy workloads expose instability that a synthetic loop misses, and an unstable undervolt usually shows up as a driver timeout, not a crash.
- Tune the fan curve last. Reason: once the card runs cooler at the same noise, you can lower fan speed targets rather than fight the original curve.
Worked example. An RTX 5070 reference board at its 250 W limit typically boosts near 2.5 GHz at around 1.05 V under a sustained 1440p load. Setting the limit to 92%, capping voltage at 0.925 V with a 2,650 MHz point, and targeting 73 °C produces roughly 215 W board power, a core temperature near 68 °C, and an average frame rate within about 2% of stock — with noticeably less fan noise and less heat dumped into the room.
Fan curves: the temperature-to-noise trade-off
Fan speed is the only variable that changes noise without changing performance at all, which makes it the most personal setting in the whole profile. The table below shows typical behaviour for a triple-fan card measured at roughly one metre in a quiet room.
| Fan speed | Typical noise | Best used for | Trade-off |
|---|---|---|---|
| 0% (fan stop) | 0 dBA | Idle, desktop, video playback below ~50 °C | Heat soaks the heatsink; first minute of a game is hotter |
| 30–40% | 22–26 dBA | Light gaming, older titles | Not enough for a 300 W+ card at full load |
| 50–60% | 29–33 dBA | The everyday sweet spot for mainstream cards | Audible in a silent room, fine under headphones |
| 70–80% | 36–40 dBA | Flagships at 4K, small cases, hot rooms | Clearly audible; the point where most people cap the curve |
| 90–100% | 44–50 dBA | Emergency fallback only | Rarely buys more than 2–3 °C over 80% |
Two habits matter more than the exact numbers. First, set a ramp delay or hysteresis of 3–5 seconds so the fans do not oscillate audibly as temperatures cross a threshold. Second, never let the curve reach 100% before 85 °C — if a card needs maximum fan speed to stay under 80 °C, the problem is airflow or dried thermal paste, not the curve.
Which profile fits your situation
| Your situation | Recommended profile | Why |
|---|---|---|
| Small-form-factor or two-fan card in a tight case | Power limit 80–85%, aggressive undervolt, fan target 78–80 °C | Accepts more noise in exchange for keeping a restricted heatsink out of throttle range |
| Living-room or quiet-focused build | Power limit 85%, undervolt, fan curve capped at 55–60% | Accepts 4–6 °C higher temperatures to stay below the audible threshold |
| 550–650 W power supply driving a flagship | Power limit 70–75% | Reduces transient spikes that trip over-current protection on older units |
| Room regularly above 30 °C, no air conditioning | Shift the whole fan curve 5 °C lower and drop the power limit another 5% | Ambient temperature sets the floor; software cannot beat physics |
| 1080p esports titles only, GPU rarely above 70% load | Undervolt only, keep fan stop enabled | The card never reaches its power ceiling, so limiting it does nothing useful |
| Already 80 °C+ with a hotspot delta above 25 °C | Repaste and check thermal pads before tuning anything | A large core-to-hotspot gap points to mounting pressure or paste pump-out, not settings |
What wears out, and what settings can’t fix
Software tuning buys temperature headroom; it does not stop the parts that degrade on their own. Thermal paste dries and pumps out over roughly two to four years of daily gaming, which shows up first as a widening gap between core and hotspot temperature rather than a sudden spike. Fan bearings develop a tick or a rumble long before they fail, usually after 20,000–30,000 hours of spinning — and a fan curve that sits at 70% instead of 45% gets there noticeably sooner. Dust accumulates fastest on cards mounted horizontally in cases with negative pressure, and a clogged fin stack can cost 5–8 °C on its own.
Three-slot and four-slot coolers on flagship cards also weigh well over 1.5 kg, so sag is a real long-term concern; a support bracket costs very little compared with a cracked PCB or a warped slot. And on cards with high-power memory, memory junction temperature matters as much as core temperature — keep an eye on it separately if your monitoring tool exposes it.
How to verify a profile is genuinely working
Compare the same 30-minute run before and after, logging average core clock, average board power, average core temperature, hotspot delta and average frame rate. A good result shows a lower power draw, a lower temperature, a stable clock that does not oscillate, and an average frame rate within a few percent of baseline. If the clock graph looks like a sawtooth, the undervolt is marginally unstable — drop the clock at your chosen voltage by 25–50 MHz rather than raising the voltage. Re-check the profile after major driver updates, since boost tables occasionally shift, and keep a stock profile saved so you can switch back in one click for a new game that behaves oddly.
Frequently Asked Questions
Does undervolting or lowering the power limit void my warranty?
In most jurisdictions, no — software tuning through the manufacturer’s own utility or a widely used third-party tool is not a physical modification, and regulators generally treat it as normal use. Warranty problems typically arise from physical damage, such as a failed BIOS flash or a card damaged during disassembly. Keep the original BIOS intact and you are on solid ground.
How much performance do I actually lose at an 80% power limit?
Less than most people expect. On mainstream and enthusiast cards, dropping to 80% of stock power usually costs around 3–5% of average frame rate while cutting board power by a fifth and lowering core temperature by roughly 6–10 °C. Flagships lose a slightly larger share in synthetic benchmarks than in real gameplay, because game engines rarely sustain the absolute worst-case load.
Can I use the same tuning software on any brand of card?
Yes. MSI Afterburner works across Nvidia, AMD and Intel cards regardless of board partner, and each vendor’s own utility — AMD Adrenalin’s tuning tab, Nvidia’s automatic tuning option, and Intel’s graphics software — offers a subset of the same controls. Use whichever you find clearer, but avoid running two tuning tools at once, as they can overwrite each other’s settings.
What is a normal GPU hotspot temperature?
A hotspot-to-core delta of 10–20 °C under load is normal for a healthy card, and a delta that creeps above 25–30 °C usually signals dried paste, uneven mounting pressure or worn thermal pads. Absolute hotspot limits vary by architecture, with AMD publishing a 110 °C junction limit for its RDNA 3 parts and Nvidia’s core throttle points sitting in the high 80s to low 90s °C range. Whatever the limit, a wide delta is the signal worth acting on.
Should I use automatic tuning or set the curve manually?
Automatic tuning is a reasonable starting point: it applies a conservative undervolt with no input from you and is difficult to get wrong. Manual tuning wins on results, because you choose the voltage point and fan target that match your case, room temperature and noise tolerance. A practical compromise is to run the automatic tune first, note the temperature and noise, then hand-tune the fan curve on top of it.
How often should I redo my profile?
Re-validate after major driver releases and after any hardware change such as a new case, extra fans or a repositioned radiator, since ambient airflow changes what the card can sustain. Otherwise a stable profile can last for years. If temperatures slowly drift upward across several months at identical settings, the profile is fine — the paste or the dust filter is the thing that changed.
