A report has emerged of a GeForce RTX 5090 whose power connector melted while the owner was using DLSS 5. VideoCardz covered the case of an owner of the MSI GeForce RTX 5090 32G GAMING TRIO OC, who said they saw high power draw while playing NBA 2K27, after which the card stopped working. The peak value published from GPU-Z was 613.5W, above MSI's rated 575W. However, that figure is for the entire card and does not equal the power that passed through the damaged terminals. How AI rendering changes the load, and how the connector handled that load, need to be considered separately.
The 613.5W peak and the claim of sustained high load
In a post on HardForum, the owner, erek, explained that power fluctuated around 450W without DLSS 5, whereas with it enabled the draw stayed above 610W. When another participant pointed out that the image showed a peak value, erek acknowledged that the display was set to show maximum values, and added that the high readings continued until they exited the game.
In other words, the peak value captured in the image and the owner's statement about duration are different kinds of evidence. 613.5W cannot simply be called the average power draw. To verify the claim of sustained high load, a time-series log is needed.
According to VideoCardz's report, the owner noticed an unusual smell, and the plug and socket had begun to fuse as the resin melted, making the cable extremely difficult to remove. The outlet also published photos showing heat damage on multiple terminals. These details convey how concrete the failure was, but they do not reveal the state of the contacts before the damage occurred.
MSI's specification page lists power consumption of 575W, one 16-pin power connector, and a recommended power supply of 1000W. These three items refer, respectively, to the card's rated power consumption, its connection type, and the capacity of the PC's power supply unit. The gap between those figures and 613.5W alone cannot identify an undersized power supply or a power-control fault. Checking the settings and BIOS of the failed unit is also essential.
Furthermore, the card draws power from both the auxiliary power connector and the motherboard's PCIe slot. If the total includes both, learning the power through the connector requires separating out the slot's share at the same moment. Nor can the standard's maximum value be treated as a measured slot value and subtracted. The 613.5W record alone cannot establish that more than 600W flowed through the connector.
Why DLSS can increase power consumption
In its technical explanation of DLSS 5, NVIDIA distinguishes the purpose of earlier DLSS technologies from this new generative processing. While earlier technologies efficiently reconstruct images that would otherwise be rendered at higher computational cost, DLSS 5 uses learned real-world appearance to add generative processing on top of conventional rendering output. Assuming that the DLSS name always means lighter rendering overlooks this difference.
The processing uses information from the current rendered frame and in-game motion, and it is also given state carried over from previous frames. In NVIDIA's description, a one-step diffusion model operating in pixel space is built into the game's rendering pipeline on GeForce RTX 50 series cards. The computation for generation is handled by the user's GPU.
Because a process that improves image quality is added, new computational cost arises compared with prior settings. This is why the argument about efficiency, getting the same image quality with less computation, should be separated from the argument about changing image quality with additional computation. How large the power difference is must be measured with the scene and settings held fixed.
However, this explanation of the mechanism does not allow the power difference observed in erek's environment to be attributed entirely to DLSS 5's internal processing. Nor does it prove that DLSS 5 disabled the power limit. The change in load the owner reported is a lead to investigate; establishing the cause of the failure requires further measurement.
Per-terminal heating inside the 600W standard
Intel's power supply design guide describes 12V-2x6 as a connector delivering up to 600W of 12V power. There are 12 power contacts in total, counting the positive and ground sides, plus small contacts for signaling. The 600W figure builds in a design in which multiple contacts share the current.
Supplying 600W at 12V means a total of 50A, which, if split evenly across the six positive-side terminals, comes to roughly 8.3A per terminal. The calculation is 600W ÷ 12V ÷ 6, an illustrative value based on the guide's terminal configuration. It does not reproduce the actual terminal currents of the failed card.
The guide's connector performance requirements specify 9.2A per contact with all 12 contacts energized, and a temperature rise of no more than 30°C above ambient. The 30°C here is not an absolute terminal temperature. Nor does it mean that exceeding 9.2A causes instant melting. It is a figure describing the performance of a connection that meets the specified conditions.
| Figure / condition | What it refers to | What the figure alone doesn't tell us |
|---|---|---|
| 613.5W | Peak total card power recorded in GPU-Z | Power through the auxiliary connector; duration |
| 600W | Power delivery rating of 12V-2x6 | How current is actually shared across terminals in a real unit |
| ~8.3A | Calculated value for 600W, 12V, even split across 6 positive terminals | Actual terminal currents of the failed unit |
| 9.2A, ≤30°C rise | Contact performance conditions in Intel's guide | Failure temperature of a unit with degraded contact |
The figures in the table cannot be chained together to conclude that the limit was exceeded and therefore the connector melted. There is still missing information between the measurement of total power and what happens at each contact.
Joule heating at a contact is determined by the square of the current multiplied by the resistance. With the same resistance, doubling the current quadruples the heat generated. If the current across multiple terminals is uneven, the burden is concentrated on some contacts. And since resistance changes when the contact state changes, local temperature will not necessarily be the same even when the card's total wattage is the same.
That said, finding an imbalance does not by itself predict a failure. In Igor'sLAB's 2025 measurements, the RTX 5090 Founders Edition and an MSI Suprim showed slight current imbalances, and values also changed after reseating, but thermal imaging showed no abnormalities. This is an example of load sharing and contact state being subjects of measurement, not a test that reproduced the failure of this Gaming Trio OC.
In 12V-2x6, the power terminals were made longer and the sense pins shorter, so that the sense side connects only after the power-carrying terminals are fully mated. But a mechanism that orders the contact sequence is distinct from one that measures the temperature of each terminal during use. The fact that it is an improved standard cannot be equated with temperature monitoring on actual hardware.
From monitoring total power to protecting the connection
CORSAIR has commercialized "ThermalProtect," which monitors cable temperature. In the company's explanation, temperature is sensed inside a cable comb 30mm from the GPU connection, and if it exceeds 65°C, a stop signal is sent to the GPU through the sense pin. It is a design that adds monitoring on the cable side, separate from the power supply unit's own overheat protection.
This 65°C is the temperature picked up by the sensor at that location, not the temperature of the metal contacts themselves. It also concerns a different subject from the "30°C rise above ambient" in Intel's performance conditions. The existence of such products shows that monitoring beyond total power is a design option, but it is not evidence that this failure could have been reliably prevented.
A test to investigate the cause would need to match the game scene and resolution, fix the frame rate cap, and compare DLSS 5 on and off. Recording only the peak total power would not capture how long the high load lasted or where things got hot. Aligning the card's overall power trend with per-terminal current and connector temperature along a time axis is what would link the change in load to the damage.
In this case, work also remains to confirm the contact state and power settings through examination of the failed unit. A temporal coincidence between use and failure cannot yield a DLSS 5-specific failure rate. Conversely, dismissing the owner's claim of sustained high load without measuring it, on the grounds that the cause is unconfirmed, would be premature.
When evaluating DLSS 5's image quality improvements, it is worth checking not only frame rate but also the power needed to sustain that quality and the temperature of the connection. Once long-duration tests under identical conditions and analysis of the failed unit are available, users will be able to concretely weigh the benefits of AI rendering against the costs borne by the hardware.