Quick answer

Start at 900 mV in MSI Afterburner’s curve editor, flattened at the clock your card holds at stock, with the power limit at 100%. One owner’s RTX 5090 ran 0.895 V at 2,895 MHz, kept its stock frame rates and peaked 75 to 150 W lower. A power limit costs frames: ComputerBase’s Founders Edition lost 5% at 450 W.

Start at 900 mV. In MSI Afterburner’s curve editor, raise the 900 mV point to the clock your RTX 5090 sustains at stock, flatten every point to its right, leave the power limit at 100% and apply. Once that survives ten minutes of a benchmark and an hour of your games, try 875 mV. The reason to bother is in two owner results: one card at 0.895 V and 2,895 MHz matched its stock frame rates while peaking 75 to 150 W lower, and a Founders Edition at 870 mV and 2,572 MHz went from 579 W to 444 W for 4 to 5% less performance.

The power-limit slider is the blunter tool on this card. In ComputerBase’s test the Founders Edition lost 5% of its average frame rate at 450 W and 10% at 400 W, the lowest value it accepts.

I run an RTX 5080, not a 5090, so the curve walkthrough with my own card’s stock numbers is the RTX 5080 undervolt guide. Every 5090 figure below is someone else’s measurement, named where it appears.

What the power limit costs: ComputerBase’s Founders Edition

NVIDIA rates the RTX 5090 at 575 W Total Graphics Power and a 2.41 GHz boost clock. ComputerBase’s Founders Edition averaged 2,645 MHz while gaming at 4K, measured 578 W at 4K in its worst-case game, and spiked to 655 W for short moments in games. Then it capped the card and re-ran 19 games at 3840 × 2160.

Founders Edition settingAverage FPS, 4K rasterAgainst stock
600 W limit with +228 MHz core and +2,000 MHz memory102.5+4.9%
575 W, stock97.7—
450 W (78% limit)92.6-5.2%
400 W (69% limit, the lowest it accepts)88.2-9.7%

The average hides the spread. At 450 W single games lost 7 to 14%, and Warhammer 40,000: Space Marine 2 lost the 14%. At 400 W, Space Marine 2 ran 28% slower than at 575 W while Final Fantasy XVI lost only 4%. Why? Final Fantasy XVI never pulls the full 575 W even unrestricted, ComputerBase explains, and Space Marine 2 does. A power limit costs nothing in a game that sits under it and a lot in a game that lives at it.

The same test cut an RTX 4090 from 450 W to 350 W, the same 22% reduction, and the older card lost about 2%. The 5090 gives power back at a higher price.

What the curve did on two owners’ cards

These are owner posts, one card each. Yours will land wherever its silicon lets it.

Card and settingPeak board powerFrames against stockPeak GPU temp
MSI Vanguard SOC LE (Aorus Master VBIOS), stock: 2,912 MHz, about 1.045 V600 W—68 °C
Same card, 0.975 V at 3,052 MHz, +1,500 MHz memory525-575 W+3.2%63 °C
Same card, 0.895 V at 2,895 MHz, +2,000 MHz memory450-525 W+1.8%58 °C
Founders Edition, stock579 W—77 °C
Founders Edition, 870 mV at 2,572 MHz444 W4-5% lower68 °C

The first three rows are Panchovix’s on Overclock.net (March 2025), averaged over Monster Hunter Wilds, Forza Horizon 5 and Cyberpunk 2077, and his four 3DMark tests agree, averaging 2.0% above stock at 0.895 V and 4.0% at 0.975 V. Two things to know before you copy them. His card runs another board’s VBIOS with a 600 W limit, so his stock row is not a Founders Edition. And both tuned rows overclock the memory, which carries part of the gain. He uses the 0.895 V profile day to day.

The Founders Edition rows are an owner’s 4K results, with the power read in Port Royal and Steel Nomad, as GameGPU reported them on February 3, 2026. That is 135 W off the top. His 2,572 MHz is below the 2,645 MHz ComputerBase’s Founders Edition averaged at stock, so his curve gave up clock as well as voltage. I would flatten at the stock clock and keep it.

Put the two tables side by side. The 450 W limit took 125 W off and cost 5% of the frame rate, while the 0.895 V curve took a similar amount off its own card’s peak and still came out ahead of stock in all three games. That gap is the case for the curve.

The 12V-2x6 connector: what lower draw buys

Intel’s power supply design guide rates the 12V-2x6 connector to deliver up to 600 W, carried on six 12 V pins rated at 9.2 A each. A card pulling 575 W through the connector puts 47.9 A on those six pins, 8.0 A each if the split were perfectly even. It is never perfectly even.

One project measured it. astral-hwmon is a Linux driver that reads the per-pin sensors on an ASUS ROG Astral RTX 5090 OC, and its August 2026 captures give two points on the same card and cable:

  • CUDA burn: 581 W, 48.26 A total, 8.56 A on the busiest pin, 93% of Intel’s 9.2 A rating.
  • The Witcher 3 at 4K with ray tracing: peak 477 W, 38.2 A total, 6.82 A on the busiest pin, 74% of the rating. That is a game peak. GPU faults killed the game seven times, and the README notes most of the session was idle or menu load.

Neither capture is an undervolt. They show that on this card the busiest pin’s current falls with total draw. A curve like the 0.895 V profile above lowers total draw. So does a 450 W limit. I would count that as the main reason to tune a 5090 at all, ahead of the electricity bill and the few degrees the core loses, because at full load the measured card ran its busiest pin within 7% of the rating. It is still current, not contact. Different problem. A lower draw does nothing for a plug that is not pushed fully home or a cable bent hard at the connector. Seat it until it latches. The RTX 5090 power supply section covers the cable and the supply, and the 12V-2x6 pin current census collects per-pin readings from owners whose hardware can measure them.

Before you start

You need MSI Afterburner from msi.com (MSI warns about modified installers on look-alike download sites) and a monitor for board power, core voltage, effective clock and GPU temperature; HWiNFO64 logs all four, and the RTX 5080 guide’s tools section covers its setup. The version I would install and the route for each brand are in the complete GPU undervolting guide.

A curve is software. MSI’s guide notes that Afterburner’s reset button (Ctrl+D) restores the card’s defaults and keeps your saved profiles. The real risk is an unstable curve that crashes a game or the driver. Keep Apply at Windows Startup off until the curve has passed every test below, so a bad curve cannot load itself at boot. Flashing another board’s VBIOS, as Panchovix did, is not part of this method.

Step 1: Record your stock numbers

Play your heaviest game for 15 minutes, or loop a benchmark, with HWiNFO64 logging. Write down average FPS and 1% lows, then the sustained clock, core voltage, board power and GPU temperature under that load. That clock is the target. ComputerBase’s Founders Edition held 2,645 MHz at 4K; a factory-overclocked card holds more, like Panchovix’s 2,912 MHz.

Step 2: Set the curve

  1. In Afterburner, press Ctrl+F to open the curve editor.
  2. Click the 900 mV point and drag it up to your stock sustained clock. Do not set it lower.
  3. Hold Shift, then drag from empty space just left of that point, above or below the line, all the way to the right. Press Shift+Enter twice and the selected range flattens, the method in MSI’s guide.
  4. Click Apply in the main window, then save the curve to a profile slot.

The card can still drop below the flat line when it hits a power or temperature limit. Flattening caps voltage; it does not lock the clock.

Step 3: Test it, then go lower

Run a benchmark for 10 minutes. Then play. MSI’s guide uses that order, and it lists the failure signs: frame stutters, artifacts, worse 1% lows, a crash. On a 5090 I would test in the game that pulls the most power, the Space Marine 2 kind, because that is where the card runs at its limit.

  • Crash, black screen or driver reset: raise the voltage. MSI’s guide steps in 30 or 50 mV. A game that now quits with DXGI_ERROR_DEVICE_REMOVED points at the curve. If the error stays after you reset to stock, the DXGI_ERROR_DEVICE_REMOVED guide reads the code and the GPU driver crash fix takes it from there.
  • Clock sitting under the line: check whether the card is at its power or temperature limit before you add voltage.
  • Stable for an hour of gaming: keep it, then try 875 mV the same way. Stop at the lowest point that holds, and step back up one for margin.

Step 4: Make it stick

Save the final curve to a profile. Then open the profile lock and switch on Apply at Windows Startup, so the profile loads every time the PC boots and you never run the card at stock by accident after a restart. MSI notes the button stays grayed out while the profile lock is closed. Check the curve after every driver install.

If you want a hard ceiling too

The Power Limit slider in Afterburner’s main window works on its own or on top of the curve, and MSI suggests trying 70 to 90% in 5% steps and rates the curve method slightly better. On a Founders Edition the slider stops at 69%, which is 400 W; 78% is 450 W. Panchovix’s 0.895 V profile still peaked at 450 to 525 W, so a limit set on top of a curve only trims the heaviest games.

A frame cap saves more than either in games that run past your refresh rate. ComputerBase capped Doom Eternal at 144 FPS at 4K and the Founders Edition drew 273 W.

What changes and what does not

Temperature falls with power: 10 °C on Panchovix’s card at 0.895 V and 9 °C on the Founders Edition at 870 mV. NVIDIA lists 90 °C as the 5090’s maximum GPU temperature, and both cards ran well under it even at stock.

The supply does not shrink. ComputerBase recommends at least a 1,000 W supply for the 5090 and 1,200 W for more room in an Intel system, and NVIDIA’s own figure is 1,000 W. A reset or a profile that fails to load puts the card back at stock, so size the supply for stock, the case the RTX 5090 vs RTX 5080 comparison covers. Results for other cards, each with its source, are in the GPU undervolt settings database, and the RTX 5070 Ti undervolt guide runs the same method on a smaller Blackwell card.

Sources

Frequently asked questions

Is it worth undervolting an RTX 5090?
Yes, for power and heat. One owner’s card at 0.895 V and 2,895 MHz ran 1.8% faster than stock across three games while peaking 75 to 150 W lower and 10 °C cooler; that run also overclocked the memory. A Founders Edition owner reported 579 W falling to 444 W at 870 mV and 2,572 MHz, for 4 to 5% less performance.
What are good RTX 5090 undervolt settings?
Start at 900 mV in MSI Afterburner’s curve editor, raise that point to the clock your card sustains at stock, flatten everything to its right and apply. Test for 10 minutes in a benchmark, then in your games. If it holds, try 875 mV. Owner results sit between 0.870 V at 2,572 MHz and 0.975 V at 3,052 MHz.
How much performance does the RTX 5090 lose at a 450 W power limit?
About 5% on average in ComputerBase’s 4K test of the Founders Edition: 92.6 FPS at 450 W against 97.7 at the stock 575 W, across 19 games. Single games lost 7 to 14%, with Warhammer 40,000: Space Marine 2 the worst. At 400 W the average loss was about 10%, and Space Marine 2 lost 28%.
How low can the RTX 5090 power limit go?
On the Founders Edition, 400 W. ComputerBase reports that 69% of 575 W is the lowest value the card accepts. A 300 W target is out of reach on that card through the power-limit slider; an undervolted curve or a frame cap is the way to draw less than 400 W in a game.
Should I undervolt or lower the power limit on an RTX 5090?
Undervolt first. The curve lowers voltage at the clock you choose, so frame rates stay near stock; in one owner’s three games they rose. A power limit cuts clocks in the heaviest games, which is why ComputerBase’s Space Marine 2 result fell 14% at 450 W. Add a limit on top of the curve only if you want a hard ceiling.
Does undervolting protect the RTX 5090’s 12V-2x6 connector?
It lowers the current on every pin, and it does not fix a plug that is not fully seated. Intel rates each of the six 12 V pins at 9.2 A. One ROG Astral RTX 5090 measured 8.56 A on its busiest pin at 581 W and 6.82 A at 477 W. Seat the plug fully whatever you set.

Sources and corrections

Last updated
Methodology
See our methodology for research and review standards. It draws on 8 cited sources, listed below, each checked against the original page on the date above. I do not own an RTX 5090; my card is an RTX 5080, measured on its own guide. Every RTX 5090 figure here is someone else’s measurement, named where it appears: ComputerBase’s Founders Edition review for the power-limit and stock-draw numbers, two owners for the curve results, and the astral-hwmon project for the per-pin connector currents. GameGPU’s report was read in place of the Reddit post it summarizes. Every source was opened on 2026-10-02. Percentages are my arithmetic on the sources’ figures; the per-pin even-split figures are my arithmetic on Intel’s rating.
Sources
Update log
  • 2026-10-02 — Page updated.
Corrections
Spotted an error or a stale number? Email contact@techfuelhq.com. Confirmed corrections are added to the update log above.

About the author

Written by Lowell K. Wood IV, who builds and runs TechFuelHQ from St. Louis, Missouri.