NVIDIA's spec sheet lists the RTX 5090 at 575 W TGP. Real workloads tell a more interesting story. We instrumented a Founders Edition card with a Hall-effect clamp on the 12V-2x6 cable, sampled at 1 kHz, across 12 workloads.
Setup
- Ryzen 9 9950X3D, 64 GB DDR5-6400, 1,200 W Seasonic Prime TX
- Ambient 21 °C, case fans at default curve
- Driver 581.42, Resizable BAR enabled
Measured power draw (W)
| Workload | Sustained | Peak 10ms transient |
|---|---|---|
| Desktop idle (single monitor) | 22 | 41 |
| Desktop idle (3 monitors) | 38 | 62 |
| YouTube 4K HDR playback | 47 | 89 |
| Cyberpunk 2077 (4K, Path Tracing, DLSS 4) | 568 | 642 |
| Black Myth: Wukong (4K Ultra) | 542 | 618 |
| Counter-Strike 2 (1440p, 360 Hz cap) | 312 | 408 |
| Blender Classroom render | 581 | 612 |
| Furmark stress | 591 | 631 |
| 3DMark Speed Way | 572 | 645 |
| Stable Diffusion XL batch | 489 | 538 |
What stands out
The card respects its 575 W TGP envelope in sustained gaming but spikes ~70 W above it on millisecond transients — well within the ATX 3.1 PCIe 5.0 spec (which allows up to 600 W sustained on the 12V-2x6 connector and short transients above that), but a real consideration for older ATX 2.x PSUs that don't budget for transient spikes. Owners of pre-2023 1,000 W PSUs should consider whether their unit handles the spike or whether to upgrade to an ATX 3.1 unit.
The idle figure is notably good: 22 W on a single monitor is lower than the RTX 4090 (28 W) at the same workload — NVIDIA's idle power gating on Blackwell appears more aggressive. Three monitors push idle to 38 W (the display engine has to drive three pixel clocks); HDMI 2.1 driver state could potentially trim that further in future drivers.
Frame caps: the single biggest power-saving lever
The frame-rate cap experiment matters most: capping CS2 at 360 fps dropped power from a runaway 480 W (uncapped, ~890 fps with no monitor benefit) to 312 W with no perceptual difference at the display refresh ceiling. Frame caps are the single biggest power-saving lever on a 5090.
Practical settings that meaningfully reduce power without harming experience:
- Cap frame rate to monitor refresh (or refresh × 1.5 for VRR headroom) — saves 100–200 W on titles that can run far above.
- Use DLSS 4 Quality instead of Native at 4K — saves 60–110 W with minimal visual difference.
- Cap power target via MSI Afterburner to 85% — saves ~85 W with ~3% performance loss in our test.
- Disable RGB on the card if your case PCB-mounts lights — saves 4 W.
Compared to RX 9070 XT
The 9070 XT pulled 304 W sustained in Cyberpunk 2077 at the same settings — 47% less power for 76% of the frame rate. The 5090 wins on absolute performance and feature set; the 9070 XT wins on perf-per-watt by a wide margin.
| Workload | RTX 5090 | RX 9070 XT |
|---|---|---|
| Cyberpunk 2077 4K Ultra | 568 W | 304 W |
| Black Myth: Wukong 4K | 542 W | 296 W |
| Idle (single monitor) | 22 W | 18 W |
| 4K YouTube playback | 47 W | 32 W |
For 1440p targets, the 9070 XT delivers playable framerates at significantly lower power draw, lower case temperature, lower PSU spec requirement, and a smaller bill at the wall socket.
PSU sizing
NVIDIA's recommendation is 1,000 W PSU minimum for the RTX 5090. We tested with 850 W (Seasonic Focus GX-850 ATX 3.1) and the system was stable across all workloads except Furmark + Prime95 simultaneous stress, where the PSU's OCP tripped during a 720 W combined draw spike. For real gaming workloads (not synthetic stress combo), 850 W ATX 3.1 is sufficient. 1,000 W gives headroom.
Thermal observations
The Founders Edition cooler held 78 °C peak under sustained Cyberpunk loading in a well-ventilated case (3× 140 mm intake, 2× 140 mm exhaust, ambient 21 °C). Case fan curves matter: switching to a quieter fan profile pushed the GPU to 82 °C — still safe, but reduces boost clock by 5–8% under sustained load.
See our best GPUs 2026 guide for the broader GPU landscape, or the RTX 5090 vs RX 9070 XT 4K ray-tracing lab for the head-to-head benchmark numbers.
See the RTX 5090 vs RX 9070 XT comparison.