NVIDIA’s DLSS 5 neural rendering technology is driving hardware power consumption to high levels.
The 600-Watt Threshold and the Fatal Hardware Failure
The incident occurred as hardware enthusiast erek tested the first officially supported DLSS 5 title, NBA 2K27, on an MSI GeForce RTX 5090 Gaming Trio OC. According to metrics captured via GPU-Z, running the game under standard conditions previously drew roughly 450W. Once DLSS 5 was engaged, that figure climbed rapidly and sustained levels above 610W, logging a peak Board Power Draw of 613.5W.
Mid-session, the card lost stability and shut down entirely. The user reported a distinct burning odor and discovered that the plastic housing of the 16-pin power plug had fused directly to the graphics card’s onboard socket. Disassembly required significant physical force, revealing at least five terminal pins with severe burn marks and one completely melted surrounding plastic shroud. The cable used was an MSI-supplied variant featuring a yellow visual indicator designed to verify complete insertion.
It is vital to note that GPU-Z’s 613.5W metric represents total board power, not current delivered exclusively through the single 16-pin interface, as the PCIe slot also supplies auxiliary power. However, this total significantly exceeds the MSI RTX 5090 Gaming Trio OC’s rated 575W specification, which recommends a 1000W power supply unit.
Independent Benchmarking Reveals Broad Power Surges
To isolate the power demands of NVIDIA’s neural rendering pipeline, Tom’s Hardware conducted extensive lab tests using an MSI RTX 5090 Lightning Z—an extreme overclocking variant equipped with dual 16-pin connectors and an Extreme vBIOS capable of pushing power limits up to 1000W. Testing at 4K resolution with DLSS Performance mode and maximum path tracing revealed massive generation-wide power spikes:
- Lightning Z / Cyberpunk 2077: Power draw increased from 580W to 723W (approx. 25% jump).
- Lightning Z / Hogwarts Legacy: Power draw surged from 480W to 720W (a 50% increase).
- RTX 5090 Founders Edition / Hogwarts Legacy: Power draw rose from 417W to 547W (approx. 31% increase).
- Lightning Z / Control: Power draw climbed from 691W to 802W.
For games lacking native integration, testers utilized community mods like OptiScaler to inject the models, which incurred a performance penalty. When testing native integration in NBA 2K27 using professional NVIDIA PCAT hardware monitoring equipment, the RTX 5090 Lightning Z averaged nearly 850W under maximum 4K workloads. This represents an increase in power draw over the Founders Edition while yielding a performance gain.
Distinguishing Correlation from Causation in Power Delivery
Despite the severe thermal events, industry analysts emphasize that DLSS 5 cannot be definitively blamed as the direct root cause of the melted connector. Multiple cases of 16-pin thermal degradation on RTX 5090 cards have surfaced independently of DLSS workloads since the architecture’s market debut.
As VideoCardz points out, neural rendering models increase raw GPU compute loads significantly, which in turn stresses the entire delivery path. Industry discussions tracked by outlets like PCM highlight underlying concerns regarding current imbalances across the 12V-2×6 connector standard. Delivering 600W at 12V requires roughly 50 amperes, which divides to about 8.3A across six individual positive terminals. If contact resistance spikes on even a single pin due to microscopic tolerance variances or incomplete seating, localized Joule heating escalates rapidly. Because resistance increases with temperature, this creates a dangerous thermal runaway loop.
Hardware commentary from engineering observers notes that current graphics cards lack per-terminal telemetry monitoring. Historical testing has occasionally caught individual wires carrying up to 23 amperes while adjacent conductors sit nearly idle. While DLSS 5 does not inherently break hardware, its intense computational demands push voltage regulation modules, contact resistances, and thermal safety margins to their absolute limits.
