DLSS 5 pushes RTX 5090 power draw sharply higher

Early community-mod testing found that DLSS 5 Neural Rendering can add a substantial power burden to RTX 5090-class hardware. In the clearest result, an MSI RTX 5090 Lightning Z rose from 480 W to 720 W in Hogwarts Legacy when DLSS 5 was enabled under the same reported conditions. That is a 50% increase—but it is a result from an early OptiScaler injection path, not a universal measurement for every RTX 5090 or official DLSS 5 implementation.

The distinction matters because a separate user later reported instability and a melted 16-pin connector after testing DLSS 5. The power concern is measurable in the tests below; the claim that DLSS 5 caused that particular failure remains unconfirmed.

The short answer: DLSS 5 can add serious RTX 5090 power demand

The GeForce RTX 5090’s reference specification lists 575 W of Total Graphics Power (TGP). TGP is the card’s rated graphics power, not the same thing as every reading taken during a specific game test. The MSI RTX 5090 Lightning Z used in the high-power tests had a reported 1,000 W Extreme vBIOS limit, so it was not operating under the same ceiling as the Founders Edition.

The results also came from community implementations using OptiScaler. They used 4K output with DLSS Performance upscaling—equivalent to a 1,920 × 1,080 input—maximum graphics settings, ray tracing or path tracing, and no Frame Generation. NVIDIA’s official Streamline integration may not behave identically.

A small sample of these tests showed a 39%–49% performance reduction with DLSS 5, alongside the higher power readings. Neural rendering is therefore not a free layer of image polish in this setup: it adds work at the output resolution, and the cost depends heavily on the game, card, power limit and implementation.

The measured numbers, game by game

Here are the comparable readings from the same testing approach. “DLSS 5 off” and “DLSS 5 on” refer to the reported GPU power measurements for each exact card and game.

The strongest increase in this set is the Lightning Z’s 480 W-to-720 W result in Hogwarts Legacy. The Founders Edition repeatedly encountered its 575 W power ceiling in Control, so its exact off/on comparison is not included here.

These figures should not be read as a new standard operating target. The Lightning Z’s Extreme vBIOS, the demanding graphics settings and the community injection path all shape the result. A 575 W reference TGP, a measured board-power reading and a power-supply recommendation describe different things.

What happened to the reported connector

A user reported that an MSI RTX 5090 Gaming Trio OC drew more than 610 W consistently while testing DLSS 5 in NBA 2K27, compared with approximately 450 W during earlier gaming. The card then became unstable, and its 16-pin connector was reported to have at least five scorched pins, with part of the housing melted.

That sequence is a reported hardware incident, not proof of a DLSS 5 failure mechanism. The 16-pin interface was described as capable of carrying up to 600 W, while the PCIe slot supplies up to 75 W separately. A connector fit and contact quality can also affect how power is delivered. The available account does not establish that DLSS 5 was the sole cause—or the cause—of the damage.

The practical conclusion is narrower and more useful: early DLSS 5 testing can push some RTX 5090 configurations into unusually high power readings, but one damaged connector cannot turn that observation into a product-wide safety claim.

Why the three-8-pin RTX 5090 experiment is not a fix

TecLab modified a Galax RTX 5090D HOF OC LAB XOC by removing its original connectors, soldering three 8-pin connectors directly to the power regulators and rewiring the sense pins. The liquid-cooled modification reportedly reached 900 W, more than 120 A, and 3,400 MHz.

That changes the board itself. It does not show that a normal RTX 5090 can be converted through a conventional cable swap, nor does it establish a supported or safe consumer setup.

The photo makes the physical arrangement clear: three connectors sit along the edge of a modified circuit board. It does not, by itself, establish the card’s exact model, its wattage, its clock speed or the safety of the modification.

A second GPU can move the neural workload—but this is another experiment

A separate unofficial MGPU Bridge demonstration moved DLSS 5 neural post-processing to a second GPU rather than running it on the card handling the game’s rendering. The setup used two RTX 5060 Ti cards, a second display and added display latency.

The underlying idea is straightforward: the render GPU produces a finished frame, then another device processes that frame. It is a technical demonstration, not a requirement for DLSS 5 and not an ordinary two-card RTX 5090 configuration. The normal power tests used one RTX 5090.

This experiment does, however, clarify why the power question is more complicated than a single card specification. The neural workload can be separated from the rendering workload in an unofficial setup, but doing so changes the hardware, display path and latency characteristics.

What RTX 5090 owners should take from this

The different setups answer different questions, so they should not be blended into one sweeping claim:

For an RTX 5090 owner, the sensible reading is not “DLSS 5 will melt every connector.” It is that early community implementations can materially change the workload and power behavior of particular cards under particular settings. Keep the card model, BIOS or power limit, game, graphics mode and software path attached to every number you see.

The 720 W result is real within those stated conditions. The connector-causation claim is not established. And the three-8-pin experiment demonstrates how far a modified board can be pushed—not what a normal RTX 5090 should be asked to do.