Nvidia DLSS 5 Power Draw Pushed an RTX 5090 Past 600W Before Its Connector Melted, Owner Says
Nvidia DLSS 5 power draw reportedly pushed an MSI RTX 5090 above 600W before its 16-pin connector melted during NBA 2K27 testing. The owner posted photographs of the damaged cable and socket, but the claim that DLSS 5 caused the failure remains unverified.
The distinction matters. There is direct evidence that the connector suffered heat damage, and independent testing shows DLSS 5 can sharply increase GPU consumption. There is not yet enough evidence to prove that the neural rendering workload caused this particular connector to fail.
That uncertainty does not make the episode irrelevant. Nvidia's RTX 5090 already operates near the practical limits of a single 12V-2x6 power connection. A demanding new rendering model now exposes how little margin can remain when software asks the GPU to sustain an unusually intensive workload.
The Reported RTX 5090 Connector Failure
The incident combines visible connector damage with a sustained workload that the owner says drove the card beyond its normal gaming consumption.
A HardOCP forum member named Erek reported the failure on September 4, 2026. The owner said the MSI RTX 5090 Gaming Trio OC was running DLSS 5 in NBA 2K27 when its reported board power remained above 600W.
The original forum post said the card was drawing “over 600+” watts before it stopped working. A later post described the load as consistent rather than a momentary spike.
According to the owner, GPU-Z showed consumption hovering above 610W. The same card reportedly consumed about 450W during ordinary gaming before DLSS 5 entered the test.
If those readings are accurate, the workload added more than 150W. That is roughly one-third above the owner's earlier gaming figure, although software telemetry cannot identify how current was divided across individual connector contacts.
The system reportedly became unstable and then stopped operating. Erek noticed a burning smell and inspected the power connection, where the plug had become difficult to remove.
The owner said melted plastic had begun bonding the cable to the graphics card socket. Published photographs appear to show several scorched contacts and substantial damage around at least one cable terminal.
The cable reportedly had a yellow insertion marker. That feature is designed to help users confirm that the plug has entered the socket completely because exposed yellow material indicates an incomplete connection.
A marker can reduce ambiguity during installation, but it cannot record what happens later. It does not prove that every terminal maintained low electrical resistance throughout repeated heating, cooling, movement, or sustained use.
The affected card was an MSI Gaming Trio model, not the dual-connector Lightning Z used in several high-power laboratory tests. It relied on one 16-pin connection for most of its external power.
Tom's Hardware published the account on September 7 and explicitly said it could not establish causation. Its incident coverage nevertheless found the owner's explanation plausible enough to investigate because its own power measurements showed a similar pattern.
This is therefore a documented component failure paired with an unconfirmed causal theory. Treating it as proof against DLSS 5 would go beyond the available evidence. Dismissing it as an unrelated anecdote would overlook the matching power behavior found in separate tests.
Nvidia DLSS 5 Power Draw Is the Larger Warning
The most important finding is not one damaged card, but the repeatable increase in consumption when DLSS 5 runs on heavily loaded RTX 5090 hardware.
DLSS 5 is Nvidia's generative neural rendering system. It uses scene data and a diffusion-based model to infer details such as lighting, materials, and shadows while a game renders.
That process differs from ordinary upscaling. The model performs substantial additional work on the GPU's Tensor Cores, which are processors optimized for the matrix operations used by neural networks.
Nvidia's DLSS 5 research describes a system guided by game-engine data rather than a simple post-processing filter. This approach gives developers more control over which reconstructed visual properties belong in a scene.
The quality goal carries a measurable compute cost. Nvidia's published research indicates that the model takes about eight milliseconds to process a 4K frame under its documented conditions.
Eight milliseconds is a large portion of a real-time rendering budget. A complete frame must finish in about 16.7 milliseconds to sustain 60 frames per second before frame-generation techniques alter the displayed rate.
Tom's Hardware tested early DLSS 5 implementations in Cyberpunk 2077, Hogwarts Legacy, and Control. Those tests used community integration software, so they do not perfectly represent developer-tuned releases.
The publication still measured similar execution behavior to Nvidia's research. That makes the results useful as an early stress test, though not a definitive forecast for every official implementation.
On an RTX 5090 Founders Edition, Cyberpunk 2077 power consumption reportedly rose from 482W to 551W when DLSS 5 was enabled. That is an increase of about 14%.
The same test produced a more dramatic response from MSI's RTX 5090 Lightning Z. Consumption climbed from 580W to 723W, a 25% increase.
Hogwarts Legacy created the largest proportional difference in that test group. The Lightning Z rose from 480W without DLSS 5 to 720W with it, producing the widely reported 50% increase.
In Control, the dual-connector card moved from 691W to 802W. The standard Founders Edition reached its configured power limit instead of matching the Lightning Z's unrestricted scaling.
These DLSS 5 tests also measured performance losses between 39% and 49% across the small game sample. Multi Frame Generation was disabled, allowing the tests to isolate the neural model's underlying cost more clearly.
An official NBA 2K27 test later reinforced the power pattern. The dual-connector Lightning Z reportedly averaged nearly 850W, drawing 48% more power than the Founders Edition for 22% higher performance.
That result does not mean every RTX 5090 will consume 850W. The Lightning Z has two 12V-2x6 connectors and a specialized firmware mode that allows far more power than a standard card.
Instead, it shows what the GB202 processor requests when electrical and firmware constraints loosen. Standard cards can hit their limits before the same workload reaches its unconstrained performance level.
The Nvidia DLSS 5 power draw issue therefore extends beyond the damaged MSI card. Several separate workloads show that neural rendering can push already demanding hardware into a different operating range.
That range matters because electrical failures depend on more than an average wattage number. Sustained current, contact resistance, current imbalance, cable condition, socket condition, and cooling around the connection all influence heat.
A 600W Connector Leaves Little Room for Bad Contacts
High total power does not automatically melt a connector, but it increases the consequences of an imperfect electrical path.
Nvidia lists the RTX 5090 Founders Edition with 575W of total graphics power. The company's RTX 5090 specifications also recommend an 850W minimum system power supply, with higher capacity potentially required for some configurations.
Total graphics power describes consumption by the card, not power passing exclusively through its cable. A PCIe x16 motherboard slot can supply part of the card's energy, while the external connector supplies most of it.
That distinction means a GPU-Z reading above 600W does not establish that the 16-pin plug alone carried more than 600W. It also does not reveal how evenly the load traveled across the connector's six positive and six ground contacts.
The 12V-2x6 connector is the revised form of the earlier 12VHPWR design. Its twelve larger terminals carry power, while four smaller sense contacts communicate connection and power information.
A correctly seated connection with healthy terminals should handle its specified load. Problems emerge when one or more contacts develop higher resistance or when current becomes concentrated through fewer electrical paths.
Electrical heating grows with both resistance and the square of current. A modest rise in resistance can therefore become serious when a contact carries unusually high current for an extended period.
Heat can soften the surrounding plastic and reduce mechanical stability. That deterioration can further weaken contact quality, creating a feedback loop of rising resistance and temperature.
This mechanism explains why connector failures can look highly localized. A card's total consumption may remain inside an overall limit while one terminal becomes much hotter than the others.
Software monitoring usually reports board power and connector voltage, not the temperature and current of every individual contact. A normal-looking total can therefore conceal an uneven distribution.
Conversely, a high board-power reading does not prove that any terminal exceeded its safe condition. Properly distributed current and a sound connection can support a heavy workload without melting.
That is the central limitation in Erek's report. Photographs document the aftermath, while GPU-Z documents the card's reported total behavior. Neither records per-pin current or temperature before the damage occurred.
The yellow-tipped cable adds relevant context but does not close that gap. It suggests that the owner had a visible guide for full insertion, yet it cannot rule out terminal wear, manufacturing variation, cable movement, or progressive damage.
The card's condition before the test also remains important. A connection already carrying elevated resistance could fail when any workload raises consumption, whether that workload involves DLSS 5, ray tracing, rendering, or another sustained GPU task.
DLSS 5 might therefore have been the final stressor without being the underlying defect. In that scenario, the software exposed a weak connection rather than creating one.
The opposite interpretation is also possible. The connector may have remained serviceable during ordinary loads, while the new workload pushed its thermal margin beyond a practical boundary.
Only a controlled inspection could distinguish those paths. Useful evidence would include the power supply model, exact cable configuration, socket condition, terminal resistance, firmware settings, and current distribution before failure.
A teardown could also determine whether damage began at one terminal, across several contacts, or elsewhere on the card. No such independent examination was publicly available when the report appeared.
The Connector History Makes One Anecdote Harder to Ignore
This claim attracts attention because the 16-pin connection already has a documented history of overheating, not because one forum post establishes a failure rate.
Reports involving melted 12VHPWR connectors followed the RTX 4090's introduction. Investigations frequently focused on incomplete insertion, cable bending, manufacturing tolerances, damaged contacts, and uneven current distribution.
The 12V-2x6 revision altered the connector geometry to improve connection sequencing and reduce operation with a partially inserted plug. It did not eliminate every variable affecting contact quality.
RTX 5090 owners have since reported additional damaged plugs and sockets across several card, cable, and power-supply combinations. These reports demonstrate that failures still occur, but they do not provide a reliable denominator.
Without the total number of cards, usage hours, cable types, installation conditions, and verified duplicate-free cases, online posts cannot establish prevalence. Viral photographs can make a rare failure look common, while unreported incidents can create the opposite distortion.
A separate historical event shows why hardware details matter. The United States Consumer Product Safety Commission announced a recall of CableMod's angled 12VHPWR adapters in February 2024.
The adapter recall covered about 25,300 units after 272 reports involving adapters becoming loose, overheating, or melting into GPUs. The affected component was a third-party angled adapter, not Nvidia's revised connector or the cable in Erek's system.
That difference prevents the recall from proving a general defect in every 16-pin installation. It does establish that connection quality within this power-delivery family can produce real fire and burn hazards.
The DLSS 5 episode introduces a new dimension to the history. Earlier debates concentrated on whether a connector was inserted properly or whether hardware distributed current safely.
Now software can materially alter the sustained electrical workload after installation. The physical connection remains unchanged, but the rendering pipeline can ask the card to spend much more time near its limit.
This is not unprecedented in computing. Stress tests, rendering software, and artificial-intelligence workloads have long exposed marginal power supplies, cooling systems, memory, and overclocks.
DLSS 5 is unusual because Nvidia presents it as a gaming feature rather than a specialist burn-in utility. NBA 2K27 players can activate it during normal consumer use.
NBA 2K27 was the only official DLSS 5 game available when the connector report surfaced. Other early demonstrations relied on leaked or community-modified integration paths.
That fact strengthens the relevance of Erek's account while leaving causation unresolved. The owner says the failure occurred in an officially supported title, not during an unsupported modification.
Official support does not guarantee identical behavior across every system. Add-in-board firmware, factory overclocks, game settings, drivers, power limits, and background conditions can change consumption.
Nvidia's marketing also combines DLSS 5 with upscaling and Multi Frame Generation in some performance comparisons. Multi Frame Generation creates several displayed frames from rendered inputs, allowing the headline frame rate to rise despite a lower underlying rendering rate.
For NBA 2K27, Nvidia showed an RTX 5090 producing 370 displayed frames per second at 4K with its feature stack, compared with 240 without DLSS 5. The higher result used six-times frame generation and a 1080p internal input.
That presentation emphasizes output smoothness, not the cost of executing the neural model. Independent measurements show why buyers also need native frame rate, latency, board power, and power-limit data.
The commercial pressure now falls on Nvidia and its board partners. They must show that an advertised gaming workload can run safely on standard single-connector products, not only on specialized dual-connector hardware.
Game developers face a related tradeoff. DLSS 5 can enrich lighting and material detail, but aggressive model settings can consume performance and electrical headroom that might otherwise support native rendering.
Power-supply and cable manufacturers also remain part of the system. A GPU, cable, connector, and PSU form one electrical path, so a failure cannot always be assigned from a photograph alone.
What the DLSS 5 Claim Still Does Not Prove
The evidence supports concern and further testing, but it does not support declaring DLSS 5 the confirmed cause of this melted connector.
First, the owner supplied software telemetry rather than calibrated electrical measurements. GPU-Z is useful for observing board behavior, but it cannot replace instrumentation on every power lead.
Second, no independent party witnessed the full sequence. The photographs show damage after the event, not the connector's seating depth, terminal resistance, or temperature before the failure.
Third, the relevant card differs from Nvidia's reference design. Board partners can ship different firmware, cooling systems, factory clocks, and power limits.
Nvidia notes that its published reference specifications do not necessarily describe every add-in-board model. The precise power behavior of MSI's Gaming Trio must therefore be evaluated as its own configuration.
Fourth, earlier DLSS 5 power tests used both community integrations and an unusual dual-connector card. They confirm that the model can consume substantial energy, but they do not recreate Erek's system.
Fifth, correlation alone cannot identify the initial fault. DLSS 5 and the connector failure occurred during the same session, yet the connection might already have been deteriorating.
A cautious interpretation fits all available evidence. The workload reportedly raised total consumption, a damaged connection was later found, and independent tests show comparable power increases.
The unsupported step would be claiming that the feature electrically overloaded a healthy connector. No published evidence has established that sequence.
It would also be premature to claim that all RTX 5090 owners face an imminent failure. Public reports offer no trustworthy rate across the installed base.
Nvidia had not published a technical response addressing this specific case when the report circulated. MSI likewise had not released a public failure analysis tied to the reported card.
Those absences leave several practical questions unanswered. It remains unclear whether the card will be inspected, whether warranty service will identify a cause, or whether logs beyond the published screenshot exist.
The most informative response would be reproducible testing on ordinary single-connector cards. Researchers should log total board power, connector voltage, per-wire current, contact temperature, and performance across controlled settings.
Tests should include a properly seated new cable, a connection subjected to repeated insertion cycles, and several compliant power supplies. That design could distinguish a software-driven load increase from a connector-specific defect.
Official game integrations also deserve separate evaluation. Nvidia's Streamline framework may reduce overhead compared with community injection tools, while developer tuning can select less expensive model configurations.
NBA 2K27 provides the first opportunity to test that difference at scale. Results from one official implementation cannot represent the entire future catalog, but they can establish an initial baseline.
Owners should avoid treating monitoring accessories as proof of safety. Some devices can expose abnormal current distribution or initiate a shutdown, but they introduce additional connections and must themselves be validated.
A burning smell, discoloration, unexplained crashes under load, or a connector that becomes unusually hot warrants immediate attention. Users should shut the system down and seek guidance from the card, cable, or power-supply manufacturer.
They should not pull apart a hot, energized connection. Damaged components should not return to service merely because the system can still boot.
Routine installation guidance still matters. The plug should be fully seated, the cable should have adequate clearance, and sharp force near the socket should be avoided.
Those precautions reduce known risks, but they do not resolve the broader design question. Consumer hardware should tolerate its supported workloads without depending on constant manual inspection.
Three Signals Will Decide Whether This Is an Outlier
The next evidence must connect workload, electrical behavior, and verified hardware condition instead of adding more isolated photographs.
The first signal is an independent examination of Erek's card and cable. A manufacturer or qualified repair laboratory could identify the origin of the thermal damage and determine whether one terminal failed first.
Evidence of poor contact, contamination, damaged plating, or an incompletely mated terminal would weaken the argument that DLSS 5 initiated the failure. A healthy-looking connection with load-related damage would strengthen concerns about available power margin.
The second signal is broader NBA 2K27 testing on standard RTX 5090 models. Reviewers need sustained measurements across reference and partner cards using official DLSS 5 support.
If ordinary single-connector cards repeatedly approach their limits without abnormal terminal temperatures, this incident will look more system-specific. Consistent heat or current imbalance would make the design tradeoff harder to dismiss.
The third signal is Nvidia's promised optimization work. The company has said that RTX 50-series model updates are expected later in the fall, before official support expands to RTX 40-series cards.
Lower execution time, reduced board power, or better performance per watt would relieve pressure on current hardware. Unchanged consumption would suggest that DLSS 5 is designed around capabilities and electrical headroom beyond many present cards.
These updates also matter for the planned RTX 40-series rollout. Older cards have different Tensor Core performance, frame-generation limits, power targets, and connector configurations.
Expanding compatibility without transparent power and latency data would repeat the current uncertainty across a much larger hardware population. Clear per-card guidance would help users distinguish supported operation from experimental settings.
The reported Nvidia DLSS 5 power draw should therefore be treated as a warning that demands measurement, not a verdict. The connector melted, the owner says the GPU sustained more than 600W, and separate testing confirms that DLSS 5 can sharply raise consumption.
What remains missing is the link between those facts. Nvidia, MSI, reviewers, and repair specialists now have a narrow question to answer: can a healthy single-connector RTX 5090 sustain official DLSS 5 workloads with adequate thermal margin?
Until controlled tests answer it, owners should watch power behavior, follow installation guidance, and avoid assuming either perfect safety or inevitable failure. The most useful next step is not another viral image. It is per-contact data from a standard card running the same official workload.
