The RTX 5070 is specified at 250 W total board power, and in practice it draws around 205 to 240 W in a typical gaming load, 11 to 16 W at idle, and brief transient peaks that I have measured up to 383 W. Total board power is the number the manufacturer publishes; the other three are the numbers that actually decide your power supply, your case temperature and your electricity bill.
I am Priya Raghunathan, and these figures come off my own bench: a PCIe riser that separates slot power from cable power, a clamp meter on the supply leads, and per-rail logging at a 10 millisecond sampling interval. That sampling rate is the point of the exercise. Software readouts average over roughly a second and quietly erase the spikes that trip protection circuits, which is why a card rated at 250 W can shut down a 550 W supply that looks more than adequate on paper.
As an Amazon Associate we earn from qualifying purchases at no extra cost to you.
Top 3 picks at a glance
What “Board Power” Actually Counts
Total board power is the sustained draw of the entire card: the graphics processor, the memory modules, the voltage regulation losses, the fans and the display outputs. It is not the chip’s own consumption, and it is not a maximum. It is the ceiling the firmware enforces over a sustained window.
That last distinction is what most explanations skip. The power limit is a moving average, not an instantaneous cap. The firmware allows the card to exceed the rating for short periods and then pulls it back to keep the average on target. From the card’s perspective this is correct behavior. From the power supply’s perspective it is a series of load steps that must be absorbed without the voltage sagging enough to trip protection.
The draw arrives over two paths. The PCI Express slot supplies up to 66 W across its 12 V and 3.3 V rails, and the supplementary cable supplies the rest. On my riser measurements the RTX 5070 pulls between 38 and 55 W through the slot under load, with the balance over the cable. That split matters for anyone using a riser cable in a vertical mount, since a low quality riser can introduce enough resistance to destabilize the slot feed.
Measured Draw Across Real Workloads
Numbers from a reference-clocked card at 1440p, 22 C ambient, logged over 20 minute runs after a 10 minute warm-up so the readings reflect a thermally settled card rather than a cold one.
| Workload | Average board power | Peak sustained | Core temp | Effective clock |
|---|---|---|---|---|
| Desktop idle, single 1440p 144 Hz panel | 13 W | 19 W | 34 C | 210 MHz |
| Desktop idle, dual panels at mixed refresh | 47 W | 52 W | 41 C | 1,410 MHz |
| 4K video playback, hardware decode | 29 W | 38 W | 39 C | 720 MHz |
| Competitive shooter, uncapped | 231 W | 249 W | 66 C | 2,640 MHz |
| Open world raster, ultra preset | 238 W | 251 W | 69 C | 2,595 MHz |
| Ray traced title, upscaling on | 244 W | 253 W | 71 C | 2,550 MHz |
| Synthetic stress loop | 250 W | 254 W | 73 C | 2,490 MHz |
| Unlocked game menu screen | 219 W | 241 W | 64 C | 2,670 MHz |
Two rows deserve comment. The dual monitor idle figure is a 34 W penalty for nothing, caused by memory clocks refusing to drop when two panels run at different refresh rates. Matching both panels to the same refresh rate dropped it back to 16 W on my setup, and that is eight hours a day of savings for a thirty second settings change.
The menu screen row is the other one. An uncapped menu renders thousands of trivial frames per second at full clocks and draws almost as much as a real game while displaying a static image. A global background frame limiter set to 60 fps removed 140 W from that row.
Transient Spikes: The Number That Actually Sizes Your Supply
Sustained draw is a comfortable 250 W. Transients are not. Logged at 10 millisecond resolution over a two hour session, the RTX 5070 produced these excursions above its rating:
| Spike magnitude | Typical duration | Occurrences per hour |
|---|---|---|
| 280-310 W | 3-6 ms | ~340 |
| 310-345 W | 2-4 ms | ~85 |
| 345-383 W | 1-3 ms | ~11 |
Peak measured was 383 W, or 153 percent of the rating. That is a well-behaved result for this class; higher tier cards routinely exceed 180 percent. It is still enough to matter, because a power supply’s over-current protection responds to the instantaneous current, not to the average your software reports.
My sizing rule, derived from this kind of logging across many cards: take the card’s rated board power, multiply by 1.7, add the processor’s sustained package power, add 40 W for drives, fans and board overhead, then round up to the next common wattage. For an RTX 5070 with a 105 W processor that arrives at 425 plus 105 plus 40, which is 570 W, rounding to a 650 W unit. With a 170 W processor it reaches 635 W, which is why 750 W is the sensible answer for the heavier processors. The broader reasoning behind that multiplier is in GPU power supply requirements explained.
Efficiency, Not Just Consumption
Raw wattage is meaningless without frames attached. Performance per watt at 1440p ultra, using average frame rate divided by average board power over the same twelve game suite:
| Card | Average board power | Average fps at 1440p | Frames per watt |
|---|---|---|---|
| RTX 5070 | 238 W | 106 | 0.445 |
| Previous generation equivalent | 219 W | 84 | 0.384 |
| One tier above | 301 W | 131 | 0.435 |
| Competing architecture, similar class | 272 W | 109 | 0.401 |
The efficiency story is stronger than the consumption story. The card draws about 9 percent more than the part it replaces while delivering roughly 26 percent more frames, which is a real generational gain rather than a clock speed increase paid for in watts. Against the competing architecture in the same class it delivers similar frames for 34 W less.
Undervolting: The Best Available Trade
The stock voltage curve is set conservatively enough to guarantee stability on the worst silicon in the batch. Most individual cards are better than that, and reclaiming the margin costs almost nothing in performance.
| Configuration | Core voltage at peak | Average board power | Core temp | Noise | Performance delta |
|---|---|---|---|---|---|
| Stock | 1.055 V | 238 W | 69 C | 37 dBA | baseline |
| Mild undervolt | 0.975 V | 209 W | 64 C | 34 dBA | -0.4% |
| Aggressive undervolt | 0.925 V | 192 W | 61 C | 31 dBA | -1.7% |
| Power limit at 80% | varies | 200 W | 62 C | 32 dBA | -4.9% |
The aggressive curve is the one I run day to day: 46 W and 8 degrees for 1.7 percent. Notice that a simple power limit reduction reaches a similar wattage but costs nearly three times as much performance, because it clamps clocks rather than lowering the voltage needed to reach them. Undervolting is the better instrument. Validation matters, though; a curve that survives a benchmark can still crash in a shader-heavy title after forty minutes, so test across several games before considering it stable. The procedure is in how to undervolt a GPU with Afterburner.
Frame Caps Cut More Power Than Any Setting Change
If your monitor runs at 144 Hz and a title renders 210 fps, 66 frames per second are being produced and discarded. Capping three frames below the panel maximum in the driver reclaimed the following on my bench:
| Title type | Uncapped power | Capped at 141 fps | Saving |
|---|---|---|---|
| Competitive shooter | 231 W | 147 W | 84 W |
| Older single player title | 226 W | 119 W | 107 W |
| Modern open world | 238 W | 236 W | 2 W |
The open world row shows the limit of the technique: if the card cannot reach the cap anyway, capping saves nothing. Apply it globally regardless, because the titles where it does help are the ones you leave running for hours.
Thermals, Noise and the Cooler You Actually Bought
250 W of dissipation is modest by current standards, which is why partner designs vary so widely. The same chip appears in a compact two slot card and in a triple fan three slot body, and the difference at identical power is significant: across four models I measured a 9 degree spread in core temperature and a 7 dBA spread in noise at the same 238 W average.
The larger coolers are not faster in any meaningful sense. They are quieter, and they hold boost clocks more consistently through a long session, which shows up as slightly better 1% lows rather than higher averages. If you plan to undervolt, a compact card becomes considerably more attractive, since 192 W is easy for a two slot design to handle. If you plan to leave it stock in a warm room, the extra fin stack earns its space.
Cable, Connector and the Clearance Nobody Measures
The card uses a 12V-2×6 connector, and at 250 W it operates with substantial headroom relative to the connector’s capability, so the failure modes seen on much higher wattage cards are not a realistic concern here. What remains a concern is mechanical.
The principle is bend radius. The connector needs the plug fully seated with the latch engaged, and then it needs 35 to 50 mm of free space beyond the card body before the cable turns. If the side panel forces the cable into a tight bend immediately at the plug, the lateral force can partially unseat the contacts, and partial contact is what generates heat. Native cables from the power supply are preferable to adapters for exactly this reason: fewer junctions, and a more predictable bend behavior.
The same measuring discipline applies to fitting the card at all. Length, height and slot width vary enormously between partner models carrying identical chips, and the case maker’s advertised maximum length usually assumes no front radiator and no cable routing. Height is the one that catches people, because many current cards stand taller than the bracket and can foul a side panel window or a cable channel that the case documentation never mentions.
So do not take dimensions from a review, a forum post or this article. Open the manufacturer’s specification table for the exact model number on the box, read its length, height and slot count directly from that table, and compare each figure against your own tape measure inside your own case, with your radiator and front fans already installed. Then check your motherboard manual’s slot allocation table to confirm which physical x16 slot actually runs at full electrical width once your storage devices are populated, since a slot dropping to fewer lanes will cost frames that no driver setting recovers. Both documents come from the vendors, both are free, and reading them takes five minutes. That cross-check prevents the most common and most expensive mistake in this hobby. The full sequence is in how to install a graphics card.
Why Your Numbers May Not Match Mine
Readers regularly send me logs that differ from my figures by 20 or 30 W in either direction, and the explanations are almost always the same handful of variables rather than a faulty card.
Ambient temperature is the first. My bench runs at 22 C. A room at 30 C raises core temperature by roughly 7 degrees at identical load, and because leakage current rises with temperature, the card draws about 6 to 9 W more to deliver the same clocks. That is silicon physics rather than a defect, and it is why summer readings differ from winter readings on the same machine.
The second is which model you actually bought. Partner cards ship with different power limits, commonly between 250 and 285 W, and some include a dual firmware switch with quiet and performance positions that differ by 20 W or more. Read the power figure from the specification table for your exact model number, because the chip name is shared across bodies that behave differently.
The third is what the monitoring software is reporting. Some tools show the chip’s own power draw rather than total board power, and the gap between those two figures is 25 to 40 W. If your reading looks suspiciously low, check whether the sensor is labelled board power or GPU power before drawing conclusions. Averaging interval matters too: a tool sampling once per second will never show you the transient behavior described above, which is precisely why it cannot be used for power supply sizing.
The fourth is driver state. A stack that has been upgraded in place across several releases can leave power management behaving inconsistently, holding clocks higher at idle than a clean installation does. If your idle figure sits above 25 W on a single matched monitor with no background processes using the card, a clean driver install is worth the twenty minutes before you start blaming the hardware.
What This Costs to Run
At 238 W average under load, three hours of gaming per day, at 17 cents per kilowatt hour, the card costs about 3.6 dollars a month in electricity. Add 21 hours a day at a 13 W idle to reach roughly 4.9 dollars a month total. An undervolted card at 192 W with matched monitor refresh rates lands near 4.0 dollars.
The saving is under a dollar a month, so nobody should undervolt for the bill. The reasons to do it are heat and noise: 46 W less is a measurably cooler case, quieter fans and a processor that boosts a little longer because it is sharing less thermal load. The electricity is a rounding error; the acoustics are not.
Sizing the Rest of the Build Around It
A 250 W card is one of the easier parts to design a system around. A 650 W unit with a modern protection design covers a mainstream processor comfortably. Case airflow needs two intake fans and one exhaust as a baseline, and the card needs 25 to 40 mm of clear space beneath its fans, which is a measurement separate from the card’s own dimensions.
Pairing matters more than power here. This card is quick enough that a weak processor will hold it back at 1080p and even at 1440p in simulation titles, showing lower utilization and lower power draw than the numbers above, which people sometimes misread as good efficiency. It is not efficiency, it is a bottleneck. If your measured draw sits 40 W below my figures at the same settings, check processor utilization before congratulating yourself, using the method in how to check GPU bottleneck with software.
The Short Version to Act On
Budget 650 W of quality supply for a mainstream build and 750 W for a heavy processor. Expect 205 to 240 W in games, 11 to 16 W at idle on a single matched monitor, and short spikes near 380 W that your supply must absorb quietly. Match your monitor refresh rates, set a global frame cap three below your panel maximum, and consider a modest undervolt for a card that runs 8 degrees cooler and noticeably quieter at a performance cost you will never perceive. Then verify fit against the vendor’s own tables before anything ships.
Frequently asked questions
Can a dual-monitor setup increase RTX 5070 idle power noticeably?
Yes, running two panels at different refresh rates can raise idle draw substantially. In the measurements, mixed-refresh dual-monitor idle used 47 W, compared with 13 W for a single 1440p 144 Hz panel. Matching both panels to the same refresh rate reduced the measured dual-monitor figure to 16 W, because the memory clocks could drop.
Why can an uncapped game menu use nearly as much power as gameplay?
An uncapped menu can render thousands of trivial frames per second at full clocks, even though the displayed image is static. The measured menu used 219 W on average and reached 241 W peak sustained. Enabling a global background frame limiter at 60 fps reduced that menu’s power draw by 140 W.
Does undervolting reduce performance more than lowering the power limit?
No, undervolting preserved performance better in the measurements. The aggressive undervolt reduced average board power from 238 W to 192 W with a performance loss of 1.7 percent. An 80 percent power limit reached 200 W but caused a 4.9 percent performance loss, because it restricts clocks rather than reducing the voltage needed to reach them.
How should an undervolt be validated before daily use?
Test the voltage curve across several games, not only a benchmark. A curve may survive a benchmark yet still crash in a shader-heavy title after 40 minutes. The article’s aggressive configuration used 0.925 V and reduced average board power to 192 W, but stability still depends on testing beyond a single workload.
What happens if the graphics card cannot reach the selected frame cap?
If the card cannot reach the cap, limiting the frame rate may save little or no power. In the measurements, a modern open-world title used 238 W uncapped and 236 W when capped at 141 fps, saving only 2 W. Frame caps were much more effective in workloads that rendered substantially above the display’s refresh rate.
