Nvidia’s DLSS 5 was supposed to be the headline feature of the RTX 50 series’ second act. Instead, the first independent benchmarks are showing a much rougher picture: frame rates cut nearly in half and power draw climbing toward 800 watts on a single graphics card. Two outlets running separate test rigs, Club386 and XenoSpectrum, published matching conclusions within a day of each other this week, and the overlap between two unrelated labs is what turned this from a rumor into a story.
Club386 tested two RTX 5090 cards, the Founders Edition and the MSI RTX 5090 Lightning Z, and found that flipping on DLSS 5’s neural rendering pipeline could cut frame rates by up to 48%, while power draw on the Lightning Z jumped by as much as 50%. XenoSpectrum’s separate PCAT-based power measurements landed in the same territory: the Founders Edition hit its 575-watt power limit the moment DLSS 5 was switched on, and the Lightning Z pulled as much as 802 watts from the wall in one test run. For a feature Nvidia has marketed as the future of real-time rendering, that is a steep bill in both frames and electricity.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Club386 and XenoSpectrum actually measured
Both outlets ran their tests on RTX 5090-class hardware rather than lower-tier cards, since the 5090 is currently the only desktop GPU with enough headroom to run DLSS 5’s neural rendering models without immediately bottoming out. Club386’s testing showed frame rates falling by as much as 48% once DLSS 5 was enabled, a figure the outlet described as capable of practically halving performance on a card that already costs well over $3,000 at current street prices. XenoSpectrum’s numbers, gathered independently using a Power Capture Analysis Tool rig rather than software-reported wattage, told a consistent story: the RTX 5090 Founders Edition average frame rate in Control dropped from 87.5 fps to 50.9 fps with DLSS 5 active, a 42% decline, while its measured power draw rose from 563 watts to 575 watts, exactly its configured board power limit.
The Lightning Z, which ships with a higher power ceiling and beefier cooling than the Founders Edition, told a slightly different story. In the same Control test, XenoSpectrum recorded its frame rate falling from 107.0 fps to 61.6 fps, also a 42% loss, but its power draw climbed from 691 watts to 802 watts, a jump the outlet measured at roughly 13% higher board power translating into a system pulling more than 800 watts through a single graphics card. Neither card was throttling in the traditional thermal sense; both were running into power limits, which is a distinct and arguably more concerning problem for anyone budgeting a power supply around a single high-end GPU.
The power draw problem: 802 watts and climbing
The power story is arguably the bigger surprise here, since Nvidia’s own marketing around DLSS 5 has focused almost entirely on frame rate multipliers rather than what the neural rendering pipeline costs in watts. XenoSpectrum’s data shows that enabling DLSS 5 did not simply redistribute existing power budget from rasterization to neural inference; it pushed total board power higher across the board, with increases ranging from roughly 13% on the Lightning Z in Control up to a reported 50% figure from Club386’s own MSI Lightning Z testing when averaged across other titles. That range depends heavily on which game is being tested and how close to its power ceiling the card already sits without DLSS 5 enabled.
An RTX 5090 system pulling more than 800 watts on the GPU alone starts to run into real-world constraints. A single card at that draw, combined with a modern high-end CPU, pushes total system power comfortably past 1,000 watts, which is the point where PSU headroom, case airflow, and even home circuit capacity start to matter in ways they didn’t for previous GPU generations. It also reopens a conversation the industry had hoped was closed: whether the 12V-2×6 power connector standard used on RTX 50-series cards has enough safety margin for sustained draw at these levels, a question XenoSpectrum’s report raises directly by noting that connector count alone does not explain the gap between the Founders Edition and Lightning Z designs.
Frame rate losses by game: Cyberpunk 2077, Control, Hogwarts Legacy
XenoSpectrum ran its unofficial DLSS 5 comparison across three games at 4K without Multi Frame Generation layered on top, isolating the cost of neural rendering itself rather than the combined effect of frame generation. The results were consistent enough across titles that they read less like an outlier and more like a baseline. In Cyberpunk 2077, average frame rates fell by roughly 39% to 42% depending on the card, while power draw rose between 14% and 25%. In Hogwarts Legacy, the Founders Edition dropped from 121.1 fps to 62.8 fps, a loss of about 48%, while the Lightning Z fell from 124.2 fps to 72.8 fps, a roughly 41% decline. Across the three titles, XenoSpectrum calculated a “performance retention” ratio, meaning frame rate with DLSS 5 divided by frame rate without it, landing between 52% and 61% depending on the card and game.
| Game | Card | FPS without DLSS 5 | FPS with DLSS 5 | Performance retained |
|---|---|---|---|---|
| Control | RTX 5090 Founders Edition | 87.5 | 50.9 | ~58% |
| Control | RTX 5090 Lightning Z | 107.0 | 61.6 | ~58% |
| Hogwarts Legacy | RTX 5090 Founders Edition | 121.1 | 62.8 | ~52% |
| Hogwarts Legacy | RTX 5090 Lightning Z | 124.2 | 72.8 | ~59% |
| Cyberpunk 2077 (avg. of tested cards) | RTX 5090 (both variants) | Varies by card | Varies by card | ~58%–61% |
What stands out in that table is not just the size of the drop but its consistency. A performance retention band clustered between roughly 52% and 61% across three genre-different games, tested by an independent lab using its own measurement rig, is a strong signal that this is a characteristic of the current DLSS 5 build rather than a bug specific to one title or one driver revision.
NBA 2K27 and the Multi Frame Generation math problem
Nvidia’s own launch benchmarks for DLSS 5, published alongside its NBA 2K27 launch showcase, used a combination of DLSS 5 and what the company calls 6x Multi Frame Generation to advertise frame rates as high as 797 fps at 1080p on an RTX 5090. XenoSpectrum’s analysis of those same marketing figures pointed out that dividing the displayed frame rate by the six-frame multiplier implies an actual rendered frame rate closer to 133 fps on the RTX 5090, roughly 100 fps on the RTX 5080, and around 88 fps on the RTX 5070 Ti. In other words, a large share of what shows up on the frame counter with Multi Frame Generation active is synthetic, interpolated output rather than frames the GPU actually rendered from scratch.
That distinction matters because it lines up almost exactly with the independent power and performance data from Club386 and XenoSpectrum. According to XenoSpectrum, Nvidia’s own briefing, reportedly attended by outlets including Wccftech and PCWorld, acknowledged that DLSS 5’s neural rendering alone can reduce raw frame rate by roughly 50% to 60% before Multi Frame Generation is layered on top to inflate the displayed number. The marketing benchmark and the independent lab benchmark are, in effect, describing the same underlying cost using two very different framings.
Why DLSS 5’s neural rendering is so demanding
DLSS 5 is not simply an upscaler with a version bump. Nvidia unveiled it at SIGGRAPH with three separate neural models and per-object rendering control, a shift from earlier DLSS generations that applied a single upscaling pass across the full frame. That per-object approach is what enables finer image quality control, but it also means the GPU’s tensor cores are running substantially more neural inference work per frame than DLSS 4.5 required. Every object in a scene that opts into neural rendering adds its own inference pass, and on current RTX 50-series silicon, that additional compute load is what shows up as both the frame rate drop and the power draw spike measured by Club386 and XenoSpectrum.
Club386’s own DLSS archive coverage put it bluntly in an earlier note, warning RTX 50-series owners testing early DLSS 5 builds that they “should prepare for your frame rate to tumble.” That warning, published before this week’s fuller power and performance data set, now reads less like a caveat and more like an accurate preview of what independent testing has since confirmed.
RTX 40 and RTX 30 series: locked out, and for good reason
Nvidia has not officially extended DLSS 5 support to RTX 40-series or RTX 30-series cards at launch, and the leaked and modded builds that have surfaced on older hardware help explain why. A leaked DLSS 5 implementation tested on RTX 30-series silicon, reported by Videocardz, showed render latency increasing from around 29 milliseconds to 3,326 milliseconds once the neural rendering path was active, with the standalone neural rendering test requiring roughly 1.2 seconds per evaluation. In practical terms, that means frame rates collapsing toward the low single digits, a result Videocardz described as performance dropping to roughly 1 fps in the worst cases.
Separately, Wccftech’s own hands-on testing of an unofficial DLSS 5 build across ten modern games found that even RTX 5090 hardware, the newest and most capable GPU Nvidia sells, loses roughly half its performance running the experimental implementation. In outdoor scenes, Wccftech measured frame rates falling from a range of 155 to 250 fps down to roughly 118 to 125 fps, while one indoor scene dropped from around 270 fps to about 132 fps. If the flagship card built for this feature is losing half its frame rate, the case for withholding official support from three-year-old RTX 30-series hardware becomes considerably easier to understand. Readers who already tested the earlier RTX 40-series rollout, covered in our report on DLSS 5’s patched RTX 4000 support, saw a smaller but still meaningful frame rate hit even on Ada Lovelace-generation cards.
How DLSS 5 compares to AMD FSR 4 and Sony PSSR 2
Nvidia is not the only company betting on neural or AI-assisted upscaling and frame generation right now, and the competitive backdrop makes this week’s benchmarks more consequential than a single bad test result would otherwise be. AMD’s FSR 4 runs on dedicated AI accelerators built into RDNA 4 hardware and has generally been reported as less compute-hungry per frame than Nvidia’s neural rendering approach, though AMD has not shipped anything with DLSS 5’s per-object neural rendering ambitions. Sony’s PSSR 2, used on PS5 Pro, takes a console-first approach where Sony controls both the hardware and the power envelope, sidestepping the desktop GPU power-limit problem entirely since the console’s thermal and power budget is fixed by design rather than left to a graphics card vendor and a motherboard’s PCIe slot.
| Technology | Vendor | Rendering approach | Hardware requirement | Current status |
|---|---|---|---|---|
| DLSS 5 | Nvidia | Per-object neural rendering, 3 models, optional Multi Frame Generation | RTX 50 series at launch; RTX 40/30 unofficial only | Early builds, independently benchmarked this week |
| FSR 4 | AMD | AI-accelerated upscaling on dedicated RDNA 4 hardware blocks | Radeon RX 9000 series | Shipping, narrower feature scope than DLSS 5 |
| PSSR 2 | Sony | Console-tuned AI upscaling within a fixed power envelope | PlayStation 5 Pro | Shipping, fixed hardware target |
The strategic difference is worth underlining. Sony can tune PSSR 2 to a single, known thermal and power ceiling because it controls the whole console. AMD’s FSR 4 runs on purpose-built accelerator hardware rather than general tensor cores repurposed for an increasingly ambitious neural rendering pipeline. Nvidia, by contrast, is asking DLSS 5 to scale across a wide range of desktop cards and power supplies while simultaneously expanding what the neural model is responsible for rendering, and this week’s data suggests that ambition is currently outrunning what the RTX 50-series’ power budget was designed to absorb gracefully.
From DLSS 1 to DLSS 5: a pattern of rising compute costs
Nvidia’s upscaling technology has followed a fairly consistent arc since DLSS 1 launched alongside the RTX 20 series: each generation adds capability, and each generation asks the tensor cores to do more work per frame in exchange for better image quality or higher displayed frame rates. DLSS 2 introduced temporal accumulation, DLSS 3 added the first generation of Frame Generation, and DLSS 4.5, which Nvidia previously validated on RTX 4090 hardware, refined the Super Resolution model without the scale of neural rendering DLSS 5 now attempts. What is different about DLSS 5 is the scope of the jump: moving from a frame-level upscaling pass to per-object neural rendering with three distinct models is a considerably larger increase in inference workload than any prior DLSS revision, and the RTX 5090’s own power benchmarks are the first hard evidence of what that jump actually costs.
Market impact: what this means for RTX 5090 buyers and PSU makers
For consumers who already paid a premium for RTX 5090 hardware, this week’s benchmarks complicate the value proposition in a way that goes beyond frame rate charts. A card whose flagship feature can push system power well past 1,000 watts changes the calculus for power supply selection, case airflow, and even electricity cost for anyone running the GPU for extended sessions with DLSS 5 enabled. Power supply manufacturers, who have already had to respond to rising GPU and AI chip power draw across recent RTX generations, now have a fresh data point to plan around: a single consumer GPU pulling more than 800 watts under a feature Nvidia intends to promote as a selling point rather than an edge case.
There is also a competitive dimension. If DLSS 5’s neural rendering costs roughly half of a card’s frame rate and pushes power draw toward its ceiling, buyers evaluating an RTX 5090 against an RTX 5080 without DLSS 5 headroom or against AMD’s next Radeon generation have a more complicated decision than Nvidia’s marketing benchmarks alone would suggest. The gap between displayed frame rate with Multi Frame Generation active and actual rendered frame rate, as XenoSpectrum’s NBA 2K27 analysis illustrated, is exactly the kind of detail that shapes how reviewers and buyers weigh a $3,000-plus purchase.
The power supply and connector question resurfaces
The 12V-2×6 connector standard, the successor to the 12VHPWR connector that drew scrutiny during earlier RTX 40-series generations over melting reports, is once again part of the conversation. XenoSpectrum’s report specifically noted that connector count alone does not explain the roughly 227-watt gap it measured between the Founders Edition hitting its 575-watt limit and the Lightning Z pulling 802 watts in the same test, pointing instead to differences in each card’s configured power limit and board design. That distinction matters for anyone trying to understand whether DLSS 5 itself is the risk factor or whether it is simply the first widely used feature demanding enough to push well-designed cards up against limits that were always technically present but rarely exercised in day-to-day gaming.
Example PCAT-style power log (illustrative of reported test conditions)
Card: RTX 5090 Founders Edition | Game: Control | Resolution: 4K
DLSS 5 OFF -> avg 87.5 fps | board power 563 W
DLSS 5 ON -> avg 50.9 fps | board power 575 W (power limit reached)
Card: RTX 5090 Lightning Z | Game: Control | Resolution: 4K
DLSS 5 OFF -> avg 107.0 fps | board power 691 W
DLSS 5 ON -> avg 61.6 fps | board power 802 WWhat Nvidia has (and hasn’t) said
Nvidia has not published its own independently verifiable power and performance breakdown for DLSS 5 at the level of detail Club386 and XenoSpectrum provided this week; its public benchmarks so far have centered on displayed frame rate figures boosted by Multi Frame Generation, such as the NBA 2K27 numbers used at launch. According to XenoSpectrum’s account of a briefing attended by other outlets, Nvidia has privately acknowledged that DLSS 5’s neural rendering alone can cut frame rate by roughly 50% to 60%, which is broadly consistent with what independent testers have now measured directly. Nvidia has also not detailed exactly which GPU generations and quality settings will be able to run DLSS 5 at real time performance through 4K when the feature ships more broadly this fall, leaving official RTX 40 and RTX 30-series support as an open question rather than a confirmed roadmap item.
How reviewers are framing the trade-off
The reviewer consensus emerging this week is less about whether DLSS 5 looks good, which multiple outlets have found genuinely improved in specific scenes, and more about whether the performance and power cost is currently justified for most players. Club386’s warning that RTX 50-series owners should expect their frame rate to fall, paired with XenoSpectrum’s harder power numbers and Wccftech’s independent frame-rate testing across ten games, forms a rare case of three separate outlets converging on the same conclusion within the same week: DLSS 5 in its current form is a significant performance and power trade-off, not a free image-quality upgrade layered on top of existing frame rates the way earlier DLSS generations were pitched.
5 predictions for DLSS 5’s next six months
- Nvidia will likely ship driver and model optimizations before DLSS 5’s full fall release that narrow the performance gap measured this week, similar to how earlier DLSS versions improved after launch-week criticism.
- Expect Nvidia to lean harder on Multi Frame Generation multipliers in marketing rather than raw neural rendering frame rates, since the displayed-fps approach used in the NBA 2K27 benchmarks avoids highlighting the underlying rendered frame rate drop.
- Power supply vendors will likely begin marketing units explicitly rated for sustained 1,000-watt-plus total system draw as a response to RTX 5090 DLSS 5 configurations, following the same pattern seen after earlier high-draw GPU generations.
- Official RTX 40-series DLSS 5 support, if it arrives, will probably ship with a heavier performance penalty than RTX 50-series hardware sees, given the frame rate collapse already observed on leaked RTX 30-series builds.
- Reviewers and outlets will likely push Nvidia toward publishing its own power-draw benchmarks alongside frame rate claims, given how much attention the Club386 and XenoSpectrum power figures have drawn this week.
Should you turn on DLSS 5 right now?
Based on this week’s independent testing, DLSS 5 in its current, early form is best treated as an image-quality option to evaluate scene by scene rather than a feature to leave on by default. Players running an RTX 5090 with headroom in their power supply and cooling can experiment with it in single-player, non-competitive titles where a drop from roughly 100 fps to 50 or 60 fps is still comfortably playable. Anyone running a card closer to its rated power limit already, or relying on a power supply sized for pre-DLSS-5 RTX 50-series draw estimates, should treat the roughly 40% to 48% frame rate hit and power draw approaching 800 watts as a real constraint rather than a worst-case outlier, since two independent labs measured very similar numbers using different test methodologies.
Frequently asked questions
How much does DLSS 5 reduce frame rates on the RTX 5090?
Independent testing from Club386 and XenoSpectrum found frame rate drops ranging from roughly 39% to 48% depending on the game, with XenoSpectrum calculating overall performance retention between about 52% and 61% across Cyberpunk 2077, Control, and Hogwarts Legacy.
How much extra power does DLSS 5 use?
XenoSpectrum measured the RTX 5090 Founders Edition hitting its 575-watt power limit with DLSS 5 enabled, while the MSI RTX 5090 Lightning Z pulled up to 802 watts in the same test conditions, an increase Club386 separately measured at up to 50% higher power draw on the Lightning Z.
Does DLSS 5 work on RTX 40-series or RTX 30-series GPUs?
Not officially at launch. Nvidia has confirmed support across the RTX 50 series, while leaked or modded builds tested on RTX 30-series hardware by Videocardz showed render latency spiking from around 29 milliseconds to over 3,300 milliseconds, with frame rates collapsing toward roughly 1 fps in the worst cases.
Why does Nvidia’s own NBA 2K27 benchmark show much higher frame rates?
Nvidia’s marketing figures combine DLSS 5 with 6x Multi Frame Generation, which multiplies the displayed frame count with interpolated frames. XenoSpectrum’s analysis estimated the actual rendered frame rate behind Nvidia’s 797 fps RTX 5090 figure at roughly 133 fps once the multiplier is accounted for.
Is the RTX 5090 Founders Edition or a partner card like the Lightning Z better for DLSS 5?
XenoSpectrum found the Lightning Z maintained a frame rate lead of roughly 15% to 21% over the Founders Edition with DLSS 5 enabled, thanks to its higher configured power limit, though it also drew significantly more power in the process.
How does DLSS 5 compare to AMD’s FSR 4?
FSR 4 runs on dedicated AI accelerator hardware built into AMD’s RDNA 4 architecture with a narrower feature scope than DLSS 5’s per-object neural rendering approach. Direct head-to-head power and performance benchmarks between the two have not yet been published.
When does DLSS 5 officially launch?
Nvidia has said DLSS 5 will run in real time up to 4K and ship this fall, though it has not specified which GPU generations or quality settings will achieve that target at launch.
Should I upgrade my power supply before using DLSS 5?
If your RTX 5090 system is currently running close to your power supply’s rated capacity, the power draw increases measured this week, up to 802 watts on the GPU alone in some tests, are worth planning around before relying on DLSS 5 for extended sessions.
Related Coverage
- Nvidia DLSS 5: 797 FPS Claim Is Really 133 FPS [2026]
- Nvidia DLSS 5 Launches Sept. 3 in NBA 2K27: 370 FPS Claim [2026]
- Intel Crescent Island AI GPU Slips, Nvidia Holds 80% [2026]
- Nvidia GB300 Ships as Data Center Revenue Hits $89B [2026]
- Nvidia Invests $3.5B in MediaTek, 90% of Bond Sale [2026]
