DLSS 5 Power Usage on RTX 5090: Watts, TGP Limits, and 12V-2x6 Headroom (2026)
What lab tests actually show about DLSS 5 board power: RTX 5090 FE near 575 W TGP, dual-connector cards climbing toward ~720–802 W, and how to read connector / usage claims without inventing numbers.
TL;DR — Early lab tests show DLSS 5 neural rendering raises board power draw, not just frame-time cost. On an RTX 5090 Founders Edition (575 W TGP, single 12V-2x6), Tom’s Hardware repeatedly saw the card near or at its power limit with neural rendering on. On an MSI RTX 5090 Lightning Z (dual 12V-2x6, Extreme ~1000 W BIOS headroom), the same workloads jumped as high as ~720–802 W. A separate enthusiast report claimed >600 W sustained and a melted 16-pin connector in NBA 2K27 — useful as a caution, not proof that DLSS 5 alone melts cables. This piece is about watts, TGP ceilings, and connector headroom — not an “FPS tax” explainer.
Primary links
| What | Link |
|---|---|
| Tom’s Hardware — FE vs Lightning Z power under DLSS 5 (OptiScaler) | tomshardware.com/…/12v-2x6-power-connector-may-hold-it-back-on-the-rtx-5090 |
| Tom’s Hardware — enthusiast melted-connector report + lab context | tomshardware.com/…/pushed-rtx-5090-past-600w-and-melted-the-16-pin-connector |
| TechPowerUp — same melted-connector write-up | techpowerup.com/352412/… |
| Wccftech — single 12V-2x6 may hold FE back | wccftech.com/dlss-5-test-reveals-a-single-12v-2x6… |
| NVIDIA — GeForce RTX 5090 specs (575 W TGP / 1000 W system) | nvidia.com/…/rtx-5090 |
| NVIDIA — DLSS 5 / NBA 2K27 (official neural rendering launch) | nvidia.com/…/dlss-5-3d-guided-neural-rendering |
| FMKorea community 5090 play report (load / temp — not a lab) | fmkorea.com/10283437374 |
Sister Terapep posts (different angles): DLSS 5 Swapper how-to · DLSS 5 Manager · RTX 40 official support status.
Why power draw climbs
NVIDIA’s own messaging frames DLSS 5 as a computationally heavy neural / diffusion-style pass on Tensor Cores (the same theme as the “most demanding model to date” language used when discussing RTX 40 timing). Tom’s notes that community OptiScaler injection showed ~8 ms model cost at 4K — in line with NVIDIA’s published guidance — so early power behavior is unlikely to be “mod-only fluff.”
Practically: neural rendering keeps shaders + Tensor busy on top of RT/PT, so board power rises. If the card is already near its TGP / connector budget, clocks get squeezed and you feel it as lower FPS — that secondary effect is not the focus here.
Caveats that matter
- Tom’s early numbers used OptiScaler community injection, not every title’s final Streamline path.
- Tests were 4K, DLSS Performance upscaling, no Multi Frame Generation, max RT/PT where available.
- Board-power sensors (GPU-Z etc.) report total board draw, which includes the ~75 W PCIe slot — a 600 W+ reading is not automatically “600 W through the 16-pin alone.”
Confirmed lab power table (Tom’s Hardware)
Conditions summarized from Tom’s Sep 2026 piece: 4K output, DLSS Performance, max RT/PT settings as applicable, neural rendering on vs off.
| Game | Card | DLSS 5 / NR off | DLSS 5 / NR on | Δ |
|---|---|---|---|---|
| Cyberpunk 2077 (path tracing) | MSI Lightning Z | 580 W | 723 W | +143 W (~+25%) |
| Cyberpunk 2077 | RTX 5090 FE | 482 W | 551 W | +69 W (~+14%) |
| Hogwarts Legacy | MSI Lightning Z | 480 W | 720 W | +240 W (~+50%) |
| Hogwarts Legacy | RTX 5090 FE | 417 W | 547 W | +130 W (~+31%) |
| Control (max RT) | MSI Lightning Z | 691 W | 802 W | +111 W (~+16%) |
| Control | RTX 5090 FE | — | hits ~575 W TGP ceiling | power-limited |
What that means for usage
- FE / single-connector 5090s — reference 575 W TGP. With neural rendering on, Tom’s saw FE practically power-limited (mid-500 W readings that sit on the ceiling in Control).
- High-headroom dual-connector boards — Lightning Z’s twin 12V-2x6 + Extreme BIOS let draw climb into the 700–800 W range in the same titles. That is available headroom being used, not a free lunch.
- Connector design is part of the story — Tom’s argument is that a single 12V-2x6 design can cap how much sustained juice DLSS 5 wants on flagship silicon, independent of how pretty the frames look.
Official FE envelope (NVIDIA product page): 575 W Total Graphics Power, 1000 W recommended system power, supplementary power via 1× 600 W PCIe Gen 5 cable (or adapter from 4× 8-pin).
Community / incident reports (labeled)
Melted connector (one enthusiast case)
Tom’s / TechPowerUp covered a reader (“erek”) whose MSI RTX 5090 Gaming Trio OC (rated around 575 W) reportedly drew ~450 W before and >600 W sustained with DLSS 5 in NBA 2K27 (GPU-Z peak cited ~613.5 W), then failed with a melted 16-pin. Outlets stress:
- Melted 12VHPWR / 12V-2x6 failures already happen under other workloads.
- Loose seating is a common root cause historically.
- One incident does not prove DLSS 5 uniquely melts connectors — only that sustained high draw makes seating / cable quality less forgiving.
Treat as: higher sustained usage → higher risk if the connector stack is already marginal.
FMKorea 5090 play report (load / temperature)
A widely circulated Korean board post (FMKorea 10283437374) claimed roughly 94–99% GPU load, about +6–7 °C, and a base-FPS drop when enabling DLSS 5 on a 5090 (with MFG discussion in comments). That is a community play report, not a calibrated power lab — useful for “cards sit hotter / busier,” not for watt numbers.
Reader checklist (power usage, not frame-tax slogans)
| Check | Why |
|---|---|
| Note your card’s TGP / power limit and whether it is FE-like (~575 W) or a high-limit dual-plug board | Same DLSS 5 toggle can mean “ceiling” vs “+100–200 W” |
| Watch board power and power limit % in HWInfo / GPU-Z with NR off vs on | Confirms whether you are usage-bound |
| Reseat / fully click the 12V-2x6; prefer a quality Gen5 cable over a stressed adapter daisy-chain | Sustained high watts punish bad contacts |
| PSU headroom | NVIDIA’s 5090 guidance already assumes a ~1000 W system class for FE; dual-plug extremes need more |
| Don’t confuse board watts with connector-only watts | Slot contributes ~75 W |
| Official Streamline / game toggles may differ slightly from OptiScaler injections | Early lab = early lab |
FAQ
Is DLSS 5 “using more electricity” or just costing FPS?
Both can happen, but they are different measurements. Tom’s data shows clear wattage increases when neural rendering is enabled. FPS can fall partly because the card is power-limited and clocks cannot climb — that is a power-budget story first.
Will every RTX 5090 hit 800 W?
No. 800 W-class readings in Tom’s piece came from the Lightning Z extreme configuration. FE stayed near its 575 W design point.
Does this mean you should avoid DLSS 5?
Not automatically. It means you should treat neural rendering like other heavy RT stacks: monitor thermals and power, keep the connector seated, and know whether your SKU has headroom. Image quality / artistic taste is a separate debate covered elsewhere.
Is the melted-cable story confirmed as caused by DLSS 5?
No. Outlets report the user’s hypothesis and note that 5090 connector melts are already common enough that causality is not settled.
Sources
- Tom’s Hardware — early DLSS 5 power / 12V-2x6 limits on RTX 5090
- Tom’s Hardware — enthusiast >600 W / melted connector report
- TechPowerUp — melted connector / power past 600 W
- Wccftech — single 12V-2x6 may hold RTX 5090 back
- NVIDIA GeForce RTX 5090 specifications
- NVIDIA GeForce News — DLSS 5 in NBA 2K27
- FMKorea community report 10283437374
Hero still: NVIDIA / Visual Concepts / 2K — NBA 2K27 DLSS 5 comparison image from NVIDIA CDN (hotlinked for illustration).
← Back to all posts