Folders of profile shortcuts per card range are offered. They are based on instrumented reviews and large owner discussion threads, cited inline below.
- Extract
simple-nvidia-undervolt-profiles.zipfrom releases. - Run
~install-simple-nvidia-undervolt.exeto install the app to Program Files, where the shortcuts are targeting. - Double-click shortcuts from the folder that lists your card to apply them.
Every profile is one of three goals at one of four aggressiveness levels:
| Family | Goal |
|---|---|
| Perf boost | A few percent more FPS at roughly stock power: a memory overclock plus, from tier 2 up, a core clock push. |
| Power cut, same perf | The classic undervolt: much less power/heat/noise at roughly the performance you have today. |
| Deep power cut | Maximum savings, giving up a few percent of performance (quiet/SFF/summer profile). |
| Tier | Meaning |
|---|---|
| 1 low risk | Should just work on ~95% of cards. |
| 2 moderate risk | Safe for a clear majority; a small minority steps down. |
| 3 high risk | Tuned for the median card — roughly a 1-in-3 chance you need to step down. |
| 4 very high risk | Silicon lottery: roughly half of all cards will NOT hold this. Good samples only. |
Three facts shape all the values:
- Voltage behavior is uniform within an architecture. Peak load voltage, the driver's voltage
floor, and core overclock headroom don't vary between cards of the same generation, so the
--mv/--peak-mvanchors and the core push are per-generation. Perf-boost profiles cap at 1000 / 1030 / 1050 / 1050 / 1040 mV for Pascal / Turing / Ampere / Ada / Blackwell — each generation's real peak voltage under load, except on Pascal, which peaks at ~1050 mV; power and thermal limits rarely let sustained load sit that high anyway, so its lower cap costs about nothing while shaving some power. - Memory headroom is what splits a generation. Safe memory-overclock percentages differ by memory type (GDDR5 / 5X / 6 / 6X / 7) and bus width, which is why most generations get more than one folder. Memory overclocks fail soft: GDDR5-and-newer error correction retries instead of crashing, so past the limit FPS silently drops. Validate perf-boost tiers with a benchmark, and if a tier is slower than the one below it, use the one below.
- Deep power cut is pure caps — each tier only requests a point the card's own factory curve already validated, so the risk is performance given up, not instability. The driver stops honoring caps below a per-generation floor (roughly 800 mV on GTX 10/16/RTX 20, 750 mV on RTX 30, 850–875 mV on RTX 40/50); a cap below the floor is silently clamped and simply saves no more than the tier above — harmless.
Folder names list their cards. The ambiguous cases:
- 3060 Ti: two memory variants exist — the 2022 refresh is GDDR6X (check the box or GPU-Z's memory type); it belongs with the GDDR6X folder, the original with the GDDR6 one.
- 1650: the GDDR6 revision belongs in the GTX 16 GDDR6 folder, the original GDDR5 card in the GDDR5 one.
- 5050: the 8 GB desktop card is GDDR6; a 9 GB GDDR7 variant belongs with the GDDR7 folder.
- Titan-class: Titan X/Xp use the Pascal GDDR5X folder, Titan RTX the RTX 20 folder, and the RTX 4090 D the RTX 40 4070-and-up folder. Workstation and laptop GPUs are out of scope.
Pascal's voltage/clock behavior is generation-wide: sustained load tops out near 1.05 V despite the 1.093 V firmware ceiling (Tom's Hardware's 1080 Ti measurements), max core OC converges to ~2050 MHz across samples, ~5–8% above a card's own peak boost (TechPowerUp 1080 Ti), 875–900 mV holds ~1900 MHz on most cards for a 20–30% power cut (Undervolting Pascal, Overclock.net), and a full 1070 cap scan measured 159 W → 124 W at 950 mV for <1% performance (SFF.life). The folders differ only on the memory overclock, by memory type.
| Family | 1 low | 2 moderate | 3 high | 4 very high |
|---|---|---|---|---|
| Perf boost | mem +5% | core +3% · mem +8% | core +5% · mem +10% | core +8% · mem +12% |
| Power cut, same perf | 925 mV | 900 mV | 875 mV, −2% | 850 mV, −2% |
| Deep power cut | 900 mV | 875 mV | 850 mV | 825 mV |
GDDR5 takes a large percentage overclock: the everyday-stable range on the weakest 1070-class members is ~+12% (TechSpot MSI 1070 Ti: +12%, TechPowerUp Gigabyte 1070), with record samples near +20%. The ladder is sized to the weakest member and stays below the useful-throughput ceiling — GDDR5's error correction makes FPS peak around +12–15% and then decline before anything crashes. These cards are strongly power-limit-bound (75 W hard cap on the 1050/1050 Ti), so the perf-boost design of a near-peak-voltage core hold plus memory OC is what pays real FPS.
| Family | 1 low | 2 moderate | 3 high | 4 very high |
|---|---|---|---|---|
| Perf boost | mem +3% | core +3% · mem +5% | core +5% · mem +6% | core +8% · mem +8% |
| Power cut, same perf | 925 mV | 900 mV | 875 mV, −2% | 850 mV, −2% |
| Deep power cut | 900 mV | 875 mV | 850 mV | 825 mV |
GDDR5X tops out at a ~+8–10% useful overclock, measured on the 1080 Ti itself — its error correction fails silently past that — so the memory ladder stays conservative. Both cards are power- and temperature-bound at stock (the 1080 Ti FE throttles to ~1650 MHz under load, TechPowerUp), so the near-peak core hold converts freed wattage to FPS.
Turing shares one voltage story across GTX 16 and RTX 20: load voltage sits at ~1.03–1.04 V, the power limit — not voltage — bounds sustained clocks, so a core OC pays only ~5% real FPS (TechPowerUp 2080 Ti FE, TechSpot: +110 core/+700 mem = +5.6% average). The multi-user ComputerBase Turing thread pins the undervolt: a 75% power target held the same FPS at −45 W with load voltage falling to ~0.83 V. Caps below ~800 mV are ignored under load (Overclockers.com). All three folders share the voltage ladders; memory splits them.
| Family | 1 low | 2 moderate | 3 high | 4 very high |
|---|---|---|---|---|
| Perf boost | mem +4% | core +3% · mem +6% | core +5% · mem +7% | core +8% · mem +8% |
| Power cut, same perf | 925 mV | 900 mV | 875 mV, −5% | 850 mV, −5% |
| Deep power cut | 900 mV | 875 mV | 850 mV | 825 mV |
8 Gbps GDDR5 has modest headroom — a Zotac 1660 held +6.25% stable over a week of testing (Wccftech) and a Gigabyte 1650 settled at ~+7.9% (Modders-Inc) — and its error-correction plateau caps the useful window well below any crash point. The 1650 is a hard 75 W card (SkatterBencher #42). The GTX 1630's tiny 64-bit bus puts it closer to these cards than to the 12 Gbps GDDR6 parts, so it lives here.
| Family | 1 low | 2 moderate | 3 high | 4 very high |
|---|---|---|---|---|
| Perf boost | mem +8% | core +3% · mem +12% | core +5% · mem +16% | core +8% · mem +20% |
| Power cut, same perf | 925 mV | 900 mV | 875 mV, −5% | 850 mV, −5% |
| Deep power cut | 900 mV | 875 mV | 850 mV | 825 mV |
Most of this folder runs 12 Gbps GDDR6 that is physically 14 Gbps-rated silicon clocked down, so +16.7% is essentially free and review samples reach +26–28% with net-positive FPS (TechPowerUp MSI 1660 Ti: +27%, Gigabyte 1650 GDDR6: +26%, ASUS 1650 Super: +28%, +12.4% real FPS). The exception is the 1660 Super, whose memory ships at 14 Gbps already and tops out at ~+11–12% (BabelTech 1660 Super vs 1660 Ti) — tier 1 (+8%) is sized so it still holds there, while tiers 2–4 pay off on the 12 Gbps parts and merely plateau (soft, via error-correction retry) on the 1660 Super.
| Family | 1 low | 2 moderate | 3 high | 4 very high |
|---|---|---|---|---|
| Perf boost | mem +5% | core +3% · mem +6% | core +5% · mem +8% | core +8% · mem +10% |
| Power cut, same perf | 925 mV | 900 mV | 875 mV, −5% | 850 mV, −5% |
| Deep power cut | 900 mV | 875 mV | 850 mV | 825 mV |
A homogeneous block: every card (including the Super refreshes) runs 14 Gbps GDDR6, stable at +700 to +1000 MHz (~+11–14%), which brackets the +5–10% ladder with margin.
Ampere's curve tops at ~1.05–1.08 V (a 3080 FE observed at 1.081 V under load) and the hot Samsung 8 nm node undervolts famously well: 900 mV ≈ the stock 325 W performance on a 3080, and each further 50 mV off the cap saves ~30 W (Igor's Lab 3080, 3090 follow-up). The whole generation is power-limit-bound at stock, so the core push stays small everywhere; the folders split on memory type and on where each card sits on its voltage/frequency curve.
| Family | 1 low | 2 moderate | 3 high | 4 very high |
|---|---|---|---|---|
| Perf boost | mem +10% | core +3% · mem +12% | core +4% · mem +14% | core +5% · mem +16% |
| Power cut, same perf | 925 mV | 900 mV | 875 mV, −3% | 850 mV, −3% |
| Deep power cut | 900 mV | 875 mV | 850 mV | 825 mV |
Plain GDDR6 on these cards runs cool and overclocks far beyond GDDR6X: 3070 FE stable at +13%, 3060 Ti FE at +11%, 3050 at +18.6% (thefpsreview 3070 FE, 3060 Ti FE, SkatterBencher #62). Tier 1 sits just under the weakest confirmed sample; the upper tiers extrapolate above it, which is what their risk labels say. The "power cut, same perf" ladder sits 25 mV above the GDDR6X folders' because these lower-TDP cards boost to ~1.0–1.05 V at stock for their ~1900 MHz — unlike a 3080/3090, which self-settles near 900 mV to hold its power ceiling — so owner data puts "900 mV ≈ stock, 850 mV ≈ −5–6%" on a 3070; holding stock performance on ~95% of cards needs the higher caps. The narrower-bus 3060 8GB and 3050 6GB variants share the folder; percentages and caps apply unchanged.
| Family | 1 low | 2 moderate | 3 high | 4 very high |
|---|---|---|---|---|
| Perf boost | mem +5% | core +3% · mem +6% | core +4% · mem +7% | core +5% · mem +8% |
| Power cut, same perf | 900 mV | 875 mV | 850 mV, −3% | 825 mV, −3% |
| Deep power cut | 900 mV | 875 mV | 850 mV | 825 mV |
These cards are hard power-limited and already boost-throttled to ~900 mV at stock, so a 900 mV cap is free, a core OC pays only ~1%, and hot GDDR6X limits the memory ladder to +5–8% (TechPowerUp 3080 FE, thefpsreview 3070 Ti: mem OC discouraged beyond ~+5%). The 1900 MHz @ 887 mV and 1890 MHz @ 850 mV sweet spots recur through the large Overclockers UK 3080 thread.
| Family | 1 low | 2 moderate | 3 high | 4 very high |
|---|---|---|---|---|
| Perf boost | mem +3% | core +3% · mem +4% | core +4% · mem +5% | core +5% · mem +6% |
| Power cut, same perf | 900 mV | 875 mV | 850 mV, −3% | 825 mV, −3% |
| Deep power cut | 875 mV | 850 mV | 825 mV | 800 mV |
The binding constraint is the double-sided 24 GB 3090, whose back-side GDDR6X modules run 100–110 °C junction (Tom's Hardware) — over-overclocking hot GDDR6X is silently counter-productive, so the memory ladder stays junction-safe on the worst card (a 3080 Ti or single-sided 3090 Ti has headroom to spare at these tiers). The deep-cut ladder goes one step lower than the rest of Ampere because on 400–450 W dies absolute watts dominate: ~30–40 W per 50 mV, a 3090 at 793 mV still ran 1785 MHz at 285 W (Igor's Lab 3090), and a 3090 Ti held to 300 W keeps ~90% of its 4K performance (Tom's Hardware).
Ada is voltage-limited: the curve tops at ~1.05 V (4090; 4080-class ~1.07 V) and clock tracks voltage, so caps trade clock for power — 850 mV measured −5.3% performance for −33% power, and near-half TDP loses only 8% at 4K (QuasarZone via VideoCardz). Hours-long ray-traced sessions expose undervolts that pass short tests (Overclock.net 4080S thread). The split is by die class: small 128-bit power-limited parts vs the wide-bus voltage-limited rest.
| Family | 1 low | 2 moderate | 3 high | 4 very high |
|---|---|---|---|---|
| Perf boost | mem +11% | core +3% · mem +13% | core +6% · mem +15% | core +8% · mem +17% |
| Power cut, same perf | 975 mV | 950 mV | 925 mV, −5% | 900 mV, −5% |
| Deep power cut | 925 mV | 900 mV | 885 mV | 875 mV |
These parts run plain 17/18 Gbps GDDR6 on a bandwidth-starved 128-bit bus, so memory is the dominant lever: +1000 MHz (to 20 Gbps, ~+11%) is the everyday pick and +1250 was stable where +1500 crashed (Tom's Hardware 4060 Ti, thefpsreview 4060 Ti FE: +1220). Tier 1 maps to the 20 Gbps plateau, tier 4 to the ~21 Gbps crash edge. The power limit slider only allows ~8–9% on the FE, so an undervolt frees TDP that becomes real boost (HotHardware 4060 Ti). The deep-cut ladder bottoms at 875 mV because these small dies report a vBIOS voltage minimum near ~860 mV — a lower cap would clamp to nothing (single-thread sourcing; 875 is the conservative read).
| Family | 1 low | 2 moderate | 3 high | 4 very high |
|---|---|---|---|---|
| Perf boost | mem +5% | core +3% · mem +6% | core +6% · mem +8% | core +8% · mem +10% |
| Power cut, same perf | 975 mV | 950 mV | 925 mV, −5% | 900 mV, −5% |
| Deep power cut | 925 mV | 900 mV | 875 mV | 850 mV |
The GDDR6X-class, voltage-limited regime (Supers included): caps trade clock for power at a
near-fixed ratio, GDDR6X memory headroom is +5–8%
(HotHardware 4090 undervolt),
and a non-golden FE holds +230 core (+8%) and +750 mem (+6.5%)
(thefpsreview 4080 Super).
The curve floor sits at ~875 mV on many cards; lower caps are silently clamped
(Overclockers UK: how far down can you downvolt a 4090).
The 2024 GDDR6 revision of the 4070 performs within 0–2% of the GDDR6X original
(VideoCardz)
and takes the same values.
Blackwell's load voltage tops at ~1.04–1.06 V (forcing the 1.075 V slider max just throttles) and core headroom is unusually large: +270 MHz ≈ +9% on the 5090 FE, +350 MHz ≈ +12.6% on the 5080 FE (thefpsreview 5090, 5080). The ComputerBase Blackwell thread collects the undervolt spread — e.g. 870 mV/2572 MHz cut a 5090 from 579 W to 444 W for ~4–5% — with ~875 mV about the lowest most cards genuinely hold. The lone GDDR6 card splits off; everything else is uniform.
| Family | 1 low | 2 moderate | 3 high | 4 very high |
|---|---|---|---|---|
| Perf boost | mem +5% | core +3% · mem +6% | core +7% · mem +8% | core +10% · mem +10% |
| Power cut, same perf | 975 mV | 950 mV | 925 mV, −1% | 900 mV, −1% |
| Deep power cut | 925 mV | 900 mV | 875 mV | 850 mV |
GDDR7's on-die ECC turns memory over-OC into silent slowdown (Guru3D), keeping the memory ladder conservative across the stack (the 5060 Ti 8 GB and 16 GB run identical memory). One known soft spot: the 5090 is pinned at its 575 W power limit, which eats part of a held core overclock — perf-boost tier 4 realizes only ~+4–6% net on it, less than on the smaller cards. That fails soft (the cap just throttles), so the 5090 shares the folder anyway.
| Family | 1 low | 2 moderate | 3 high | 4 very high |
|---|---|---|---|---|
| Perf boost | mem +9% | core +3% · mem +11% | core +7% · mem +13% | core +10% · mem +15% |
| Power cut, same perf | 975 mV | 950 mV | 925 mV, −1% | 900 mV, −1% |
| Deep power cut | 925 mV | 900 mV | 875 mV | 850 mV |
The desktop 5050 8 GB is the only Blackwell card on GDDR6 (20 Gbps, 128-bit), free of the GDDR7 ECC clamp: a review sample held ~+14.8% fully stable with crashes only near +16%, for ~+12% real performance on this bandwidth-starved card (TechPowerUp Gigabyte 5050). The upper-tier magnitudes rest on that single sample — the risk labels carry that uncertainty. The card is power-limited at stock (130 W, +15% slider), so memory is the perf-boost lever (HotHardware MSI 5050).
- Perf boost: roughly +1–2% FPS at tier 1 (memory-bound games gain most), up to +5–10% at the top tiers on cards that hold them. Power stays roughly flat. Wide-memory folders (GTX 16 GDDR6, Ampere GDDR6, Ada 128-bit, the 5050) gain the most, since their cards are bandwidth-starved.
- Power cut, same perf: about −10–25% board power at tiers 1–2 for 0 to −5% performance; −20–35% at tiers 3–4 for ~0–3% if the card holds the held clock.
- Deep power cut: about −15–35% board power and noticeably lower temperatures/noise, for a real −1% to −10% performance cost that grows with the tier and is steepest on RTX 40/50 (their clock follows voltage more tightly than older generations).
Apply a profile with the card cooled down (idle desktop, not right after a gaming session). The
tuning resolves against the live curve, which shifts slightly with temperature, so applying hot
bakes in a slightly different tuning than applying cold. Or run
simple-nvidia-undervolt set-reference-curve once (idle and cool) and the temperature stops mattering:
tuning then resolves against the saved stock curve instead of the live one.
Games crashing, artifacts, or a driver reset after applying a profile is the normal failure mode of an unlucky sample, not damage — apply a lower tier (or Reset to stock) and you're done. Validate a high/very-high tier with 20–30 minutes of a demanding (ray-traced) game, not just a benchmark: short synthetic passes over-report stability. A memory overclock past a card's limit can also show up as silently LOWER performance instead of crashes (the error correction on GDDR5 and newer memory retries instead of crashing) — if a perf-boost tier benchmarks worse than the tier below, use the tier below.
Because a profile re-applies at logon, an unstable one comes back after a crash reboot — that's harmless at the desktop (instability shows under 3D load), but step down or Reset to stock before returning to the game.