RTX 5080 Undervolt Guide: Less Heat, Same Performance (2026)

By LK Wood IV · 2026-06-12 · ~13 min read · St. Louis County, MO

Start at 900 mV. In MSI Afterburner’s curve editor, lift the 900 mV point to your card’s stock sustained clock, flatten every point to its right and leave the power limit at default; once that survives a 30-minute game loop, try 875 mV.

On my ASUS ROG Astral RTX 5080 OC the clock to hold is its default profile’s 2,784 MHz in Time Spy, and at stock the card games at 266 to 340 W (my bench, June 2026).

The tuned results come from owners in ComputerBase’s Blackwell thread: 0.900 V at 2,910 MHz cut one card’s peak draw from 318 W to 265 W while running 30 MHz above its stock clock, and 0.875 V at 2,760 MHz ran 2 to 3 percent faster than stock at 250 to 260 W peak.

The 900 mV post gives no FPS comparison; the 875 mV result includes a +1000 memory offset. At 800 mV, HCD traded clock for lower draw.

This guide covers the voltage-frequency curve methodology for undervolting the RTX 5080, what to expect, and the settings I’ve landed on with the ROG Astral.

On AMD’s RDNA 4 cards the process runs through the Adrenalin driver instead, so see the RX 9070 XT undervolt guide; this page is the NVIDIA worked example in the broader guide to undervolting your GPU. Settings from other cards, with what each owner measured, are collected in the GPU undervolt settings database.

My ROG Astral at stock: the numbers to beat

Every figure in this table is my own card on its stock settings, from the Astral bench dataset (3DMark and HWiNFO64, test window June 8 to 11, 2026; the dataset CSV dates the 3DMark scores 2026-06-11). Run the same tests on your card before and after, and compare.

The same dataset’s manual-overclock profile (GPU Tweak III, +328 MHz core, +2,996 MHz memory, +12% power target) games at about 271 W and reached 400 W under stress, at about 58 °C with the fans at up to about 51 dBA. Those are overclock numbers. Do not compare an undervolt against them.

August 19 capture, stock 448 W power limit: decoding four LLMs the card ran a median 3,060 to 3,266 MHz at 265 to 344 W and 49 to 52 °C, and idled at 52 to 71 W at the Windows desktop across a week of captures (LLM power-efficiency dataset).

In those desktop captures multi-monitor output held the card in P0.

What I have measured and what I have not

The ladder: stock, then 900 mV, 875 mV and 850 mV, each flattened at the clock a same-day stock run holds, plus one row with only the power limit lowered to match the 875 mV row’s gaming draw.

Each row gets 3DMark Steel Nomad and Speed Way, a 30-minute game loop with average FPS and 1% lows, board power, effective clock, GPU and memory junction temperature, fan speed and whether it held, plus the four-model LLM decode test for inference readers.

Why undervolt instead of just setting a power limit

The RTX 5080 is a 360W card at NVIDIA’s spec. The ROG Astral OC variant I run draws up to ~400W in practice (450W VBIOS ceiling) on its high-performance BIOS.

Power limit reduction tells the GPU to throttle when it hits a ceiling: 90% is 324 W on NVIDIA’s 360 W spec and 360 W on the Astral’s 400 W factory limit. The card lowers clocks dynamically until it stays under the limit. Simple, effective, blunt: the GPU decides how to trade clock speed against power.

Voltage-frequency curve undervolting maps a specific voltage to a specific clock and flattens the curve to its right. You find the lowest voltage at which the card is stable at your target clock. The goal is the same clock as stock at lower power, with lower temperatures.

The one owner report I found that posts both numbers cut peak draw from 318 W to 265 W at 900 mV, with the clock 30 MHz higher than stock.

For an always-on inference box the curve is the better tool, because the saving runs every hour the card is loaded.

Every 50 W an undervolt takes off, at six hours of load a day, is 109.5 kWh a year: about $20 at 18.34 cents per kWh, EIA’s US residential average for June 2026, or $60 over three years.

California residential ran 34.74 cents in the same EIA Table 5.6.A, which makes it about $38 a year. The Power & Cost Calculator runs the same math for your rate and hours.

Understanding the voltage-frequency curve

The RTX 5080 carries a voltage-frequency table: for each voltage level there is a maximum boost clock the GPU will run at. The driver’s boost algorithm picks a point on that curve from the card’s temperature and power headroom.

Stock behavior: the GPU slides up and down the curve, following its thermal and power limits, and a factory-overclocked card sits well above the spec sheet. NVIDIA rates the RTX 5080’s boost at 2.62 GHz; my Astral’s default profile sustained 2,784 MHz in Time Spy.

The thread opener in ComputerBase’s forum runs the same Astral 5080 OC, and at stock in Steel Nomad that card held 2,766 MHz at 0.980 V.

Undervolted behavior: you take that stock point, lower the voltage while keeping the frequency, and flatten the upper portion of the curve to cap the target clock. The GPU can still move to lower points as load and power or temperature limits change; flattening is not a voltage-frequency lock.

Where a given RTX 5080 lands is silicon lottery. In ComputerBase’s Blackwell undervolting thread, three RTX 5080 owners posted points running from 2910 MHz at 900 mV down to 2572 MHz at 800 mV for a 220 W average, with 2760 MHz at 875 mV between them.

You find out where yours is in the testing below.

Tools

MSI Afterburner works with every RTX 5080 regardless of board partner. Version 4.6.6, the stable release TechPowerUp reported on October 1, 2025, added official RTX 50 support. Its voltage-frequency curve editor is the undervolting interface. Download from msi.com/Landing/afterburner.

ASUS GPU Tweak III is the tool I used for my Astral’s manual profile, which is offsets and a power target, not a curve (the dataset lists them). The curve steps below are Afterburner’s. Run one tuning tool at a time.

HWiNFO64 logs GPU temperature, GPU memory junction temperature, board power, core voltage and effective clock under load. Run it alongside whatever workload you’re testing. Test against GPU temperature and memory junction temperature.

RTX 50 cards launched with the hotspot sensor hidden from GPU-Z and HWMonitor (TechPowerUp, July 14, 2026), and CPUID lists the reading as fixed in HWMonitor 1.65.1 (CPUID, July 16, 2026). Download HWiNFO64 from hwinfo.com.

Step 1: Baseline measurements

Run HWiNFO64 and put the GPU under load:

• For gaming: run a GPU benchmark or play a demanding title for 15 minutes
• For AI inference: run your inference workload for 15 minutes
• For video encoding: run a long encode pass

Record from HWiNFO64:

• GPU clock (the curve target) and GPU effective clock (compare sustained readings under load)
• GPU core voltage (sustained under load)
• GPU power (total board power)
• GPU temperature and GPU memory junction temperature

These are your baselines, and the stock table above is mine to compare them against. The goal is to match or exceed the stock clock at lower power and temperature.

Step 2: Open the voltage-frequency curve editor

In MSI Afterburner, click the curve editor button (or press Ctrl+F). You’ll see a scatter plot: X axis is voltage (mV), Y axis is clock frequency (MHz). Each point represents a voltage-frequency pair.

The curve rises from left to right and flattens at the top where boost speeds cap out.

Note your stock operating point.

Under load, HWiNFO64 or GPU-Z shows your actual core voltage and clock; find that point on the curve.

Two stock points posted in ComputerBase’s thread: the same Astral 5080 OC at 0.980 V and 2,766 MHz in Steel Nomad (February 5, 2025), and another RTX 5080 at 1.030 V and 2,880 MHz by default (April 9, 2025).

Step 3: Apply the undervolt

Method: raise the lower-voltage point to the stock clock, then flatten the curve to its right.

1. Press Ctrl+F to open the curve.
2. Click the point at 900 mV and drag it up to your stock sustained clock. On my Astral’s default profile that is 2,784 MHz in Time Spy. Never set it below your card’s stock clock; compare average FPS and 1% lows against stock to check the performance cost.
3. Hold Shift and drag from empty space just left of the 900 mV point, above or below the curve, all the way to the right. Press Shift+Enter twice to flatten the selected range, as shown in MSI’s curve-editor guide.
4. Press Apply (the checkmark) in Afterburner’s main window. The curve editor alone does not change the card. Check that the applied curve is flat from 900 mV to the right at your target clock.

The GPU can still downclock at idle or when it hits a power or temperature limit.

Step 4: Stress test and validate

Run your GPU workload for at least 30 minutes. Watch for:

• Crashes or driver resets: the voltage is too low for that frequency. Raise it a step and retry; MSI’s guide adjusts in 30 or 50 mV steps.
• Clock sagging below target: check for a power or temperature limit first; raising voltage can make either worse. If neither limit is active, step the voltage up and retest before you lower the clock. Compare FPS and 1% lows with stock.
• Stable at target: record the power draw and temperature delta against your baseline.

Stability testing tools:

• 3DMark Time Spy or Port Royal for gaming workloads
• FurMark for thermals and power behavior. MSI’s guide lists it for checking an overclock. I do not count it as a curve-undervolt test. My reasoning: it drives the card to its power limit, where the card runs below the flattened point, so a pass says little about that point. MSI’s own sequence ends in a game: “run a 10 minute benchmark, and if it’s stable, try gaming.” Long game sessions that hold the flattened clock are the test I count.
• Ollama or llama.cpp with the model you actually run, for an inference rig

For the ROG Astral specifically, what the card does at stock and at a manual overclock, measured on my bench, is in the ASUS ROG Astral RTX 5080 OC review. A 25-minute gaming log there, on the manual-overclock profile, averaged 261 W at about 3,200 MHz with a 60 °C peak and zero throttle events.

On this cooler the voltage sets your stable point and the heat never enters into it.

Step 5: Make settings persistent

In MSI Afterburner:

1. Click the save icon (disk icon in the profile row)
2. Save to Profile 1 (or 2/3 if you want multiple configs)
3. Turn on the Apply at Windows Startup button in the main window’s upper-right corner so the profile applies at boot (the profile lock must be open, or the button stays grayed out)

Leave it off until the curve has passed the 30-minute test. An unstable curve that applies at boot is harder to back out of.

Power limits and fan curve

Afterburner’s fan curve follows GPU temperature. On my Astral at stock the core ran 63 to 72 °C while gaming with the fans near-silent, so there is nothing for a custom curve to fix; the manual-overclock profile ran cooler (about 58 °C) because its fans ran harder, at up to about 51 dBA.

On a louder card, an undervolt lowers the temperature the curve reacts to, which is what brings the fans down. Set a custom curve after the undervolt holds.

What to expect in practice

The owner reports in ComputerBase’s Blackwell overclocking and undervolting thread and on its page 14 post specific voltage, clock and power figures:

• Tech-Dino, April 2025 (the post names the RTX 5080): stock 1.030 V at 2880 MHz, 318 W peak. Tuned 0.900 V at 2910 MHz, 265 W peak and about 250 W typical. That is 53 W off the top at a higher clock.
• Wan900, February 2025 (his attached Time Spy result lists an RTX 5080): 0.875 V at 2760 MHz with +1000 on memory. Peak 250 to 260 W, and 2 to 3 percent faster than stock.
• HCD, April 2025 (his attached Speed Way result lists an RTX 5080): 0.800 V, the deep end. 2572 MHz at about 220 W average in Speed Way, 2580 MHz at about 210 W in Steel Nomad. A lower clock, traded for lower draw.

Those are owner posts in a shared-results thread, one card each. Read them as the shape of the range. Tech-Dino’s 900 mV point holds a higher clock than stock. Wan900 reports a usable 875 mV point. HCD’s 800 mV point costs clock.

My own Astral’s stock profile games at 266 to 340 W and peaked at 360 W under stress; its ladder is described under What I have measured and what I have not.

For AI inference: LLM decode is memory-bandwidth bound, not shader-bound. On my card at stock, board power sat inside a narrow 265 to 344 W band across a 7x spread in model size, at a median 3,060 to 3,266 MHz, and never touched the limit. The ladder adds the four-model decode test to each row.

Always-on AI inference setup

If you’re running the RTX 5080 as a local AI workstation running LLM inference around the clock, prioritize:

1. Undervolt for sustained efficiency (as above)
2. Lower the power limit only if you want a hard ceiling on top of the curve; MSI’s guide calls the power-limit slider an alternative and says “the Curve method seems to deliver slightly better results”
3. Monitor with HWiNFO64 + logging enabled to catch thermal drift over time

At 24/7 operation, 50 W of sustained savings is 438 kWh and about $80 a year at 18.34 cents per kWh.

The idle hours count too: what this card draws around the clock at stock, 52 to 71 W at my Windows desktop and 265 to 344 W decoding, I measured in RTX 5080 performance per watt for 24/7 AI inference and the efficiency dataset.

The ROG Astral RTX 5080 OC that I use for this testing is documented with photos and full spec sheet at the dataset page. To size a PSU and case around the same card, load it in the PC Builder.

The Power & Cost Calculator models what a GPU workstation node costs to run 24/7.

Sources

• NVIDIA — GeForce RTX 5080 specifications: 360 W Total Graphics Power, 2.62 GHz boost. Checked 2026-09-24.
• ComputerBase — Nvidia GeForce RTX 5080 im Test: Lautstärke, Temperatur, Leistungsaufnahme und OC, 2025-01-29. Founders Edition: 282 W average telemetry across the benchmark suite, 353 W worst case at Ultra HD in Metro Exodus Enhanced, spikes to 405 W, 15 W at a 4K60 desktop. Checked 2026-09-24.
• ComputerBase forum — Blackwell (5070/5080/5090) Overclocking/Undervolting Sammelthread and page 14. Owner posts: Goozilla (ASUS Astral 5080 OC at stock in Steel Nomad, 2766 MHz at 0.980 V, February 5, 2025), Wan900 (0.875 V, 2760 MHz, 250 to 260 W peak, 2 to 3 percent faster than stock, February 2025; the post names no card and its attached Time Spy result lists an RTX 5080), Tech-Dino (0.900 V at 2910 MHz, 265 W peak, against 1.030 V at 2880 MHz, 318 W stock, April 2025; the post names the RTX 5080), HCD (0.800 V, 2572 to 2580 MHz, about 210 to 220 W average in Speed Way and Steel Nomad, April 2025; the post names no card and its attached Speed Way result lists an RTX 5080). A shared-results thread, one card per post. Checked 2026-09-24.
• Club386 — Asus boosts ROG Astral GeForce RTX 5080 power limits to 450W via BIOS update, 2025-04-04: 400 W out of the box, and the March 14 BIOS adds 50 W. Checked 2026-09-24.
• TechPowerUp — MSI Pushes Afterburner 4.6.6 Stable Update After Years of Beta, 2025-10-01, and HWMonitor v1.65 Brings Back Hotspot Reading for NVIDIA RTX 50 Series, 2026-07-14. Checked 2026-09-24.
• CPUID — HWMonitor version history: HWMonitor 1.65.1, July 16, 2026, “Hotspot temperature on NVIDIA RTX 50 GPUs (fixed in 1.65.1)”. Checked 2026-09-24.
• MSI — MSI Afterburner Overclocking Guide & Undervolting Guide: Ctrl+F opens the curve editor; Shift-drag and Shift+Enter twice flatten the curve; Apply at Windows Startup needs the profile lock open; 30 or 50 mV steps; the curve method rated slightly better than the power-limit slider. Checked 2026-09-24.
• EIA — Electric Power Monthly, Table 5.6.A, June 2026 data, released 2026-08-26: residential 18.34 cents per kWh US total, 34.74 cents California. Checked 2026-09-24.
• TechFuelHQ, first-party — ROG Astral RTX 5080 OC dataset, test window June 8 to 11, 2026. Stock profile: gaming 266 to 340 W, 360 W maximum, idle 13 to 20 W, gaming core 63 to 72 °C, memory junction 72 to 82 °C, Time Spy 32,350 (CSV date 2026-06-11), sustained 2,784 MHz in Time Spy. Manual-overclock profile: gaming about 271 W, stress 400 W, 450 W VBIOS ceiling. RTX 5080 LLM power efficiency, August 19, 2026: 265 to 344 W loaded, 3,060 to 3,266 MHz median clock, 52 to 71 W desktop idle, stock 448 W limit.