Nvidia told the press that DLSS 5 could hit 797 frames per second on a GeForce RTX 5090. What it did not put in the headline: roughly 664 of those frames are generated, not rendered. An analysis published by XenoSpectrum on September 1, 2026 divided Nvidia’s own displayed FPS figures by six, the multiplier tied to DLSS 5’s Multi Frame Generation, and arrived at an estimated real rendered frame rate of about 133 FPS. That’s still fast, but it is nothing like the number printed on the slide.
The gap between Nvidia’s marketing FPS and the frame rate a GPU is actually computing has become the center of a wider controversy since DLSS 5 launched alongside NBA 2K27 on September 3, 2026. Independent outlets including Club386, XenoSpectrum, and Les Numériques have spent the past two weeks running their own numbers, and a separate thread of the story, that the RTX 5090 Founders Edition is hitting its 575W power limit under DLSS 5 neural rendering, adds a second layer to the same problem: bigger numbers on the box, bigger numbers on the power meter.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
The Number Nvidia Put on the Box
Nvidia’s own NBA 2K27 demo table, shown to press and later reproduced by XenoSpectrum, lists an RTX 5090 hitting 797 displayed FPS at 1080p with the full DLSS 5 suite active, including 6x Multi Frame Generation. An RTX 5080 is listed at 602 FPS, an RTX 5070 Ti at 530 FPS, an RTX 5070 at 386 FPS, an RTX 5060 Ti at 325 FPS, and an RTX 5060 at 258 FPS, all at 1080p. At 1440p, the RTX 5090 figure drops to 594 FPS. Nvidia says the underlying DLSS 5 models now process frames five times faster than the prior generation, which is part of how it is justifying such steep multipliers.
Those numbers were already covered when DLSS 5 launched in NBA 2K27 earlier this month. What has changed in the days since is scrutiny of what those FPS figures actually represent, and it turns out the headline number and the frame rate your eyes and your input latency actually experience are two very different things.
How Multi Frame Generation Inflates the Count
DLSS 5’s headline feature is 6x Multi Frame Generation, a step up from the 4x mode Nvidia shipped with the RTX 50 series at launch. In practice, MFG renders a base frame conventionally, then uses a neural network to generate five additional interpolated frames for every one that’s actually computed from game state. That means for every real frame, the counter displayed on screen adds six frames to the tally, even though only one of them reflects fresh input.
That’s a meaningful distinction for anyone judging performance by a frame counter. A displayed FPS of 797 sounds like a GPU rendering nearly 800 unique images a second. What it actually means, according to XenoSpectrum’s analysis, is closer to 133 rendered frames a second, with the rest synthesized between them. The visual smoothness may still be real, but the responsiveness, and the underlying rendering workload, is not six times higher than a native frame rate. It is roughly in line with what a high-end GPU has always delivered at native settings.
The Divide-by-Six Math Behind the Real Frame Rate
XenoSpectrum’s method is straightforward: take Nvidia’s own displayed FPS number from the NBA 2K27 demo table and divide by six, since six is the number of frames MFG outputs per rendered frame. The outlet published the resulting estimate as a way to separate marketing output from computational reality.
Nvidia displayed FPS ÷ 6 (MFG multiplier) = estimated rendered FPS
Example: RTX 5090 at 1080p → 797 ÷ 6 ≈ 133 rendered FPS
Example: RTX 5070 at 1080p → 386 ÷ 6 ≈ 64 rendered FPS
Applied across Nvidia’s full table, XenoSpectrum estimates rendered frame rates land in a range of roughly 43 to 133 FPS at 1080p, 44 to 99 FPS at 1440p, and 39 to 62 FPS at 4K, depending on the card. Those numbers are respectable for a demanding neural rendering workload, but they are a fraction of the figures that appeared in Nvidia’s own launch materials and in early press coverage of the 370 FPS and 797 FPS claims.
Rendered FPS Across the RTX 50 Stack
The table below lines up Nvidia’s displayed 1080p figures against XenoSpectrum’s divide-by-six estimate of actual rendered frame rate for each RTX 50-series desktop GPU tested in the NBA 2K27 demo.
| GPU | Nvidia Displayed FPS (1080p, 6x MFG) | Estimated Rendered FPS | Approx. Reduction |
|---|---|---|---|
| RTX 5090 | 797 | ~133 | ~83% |
| RTX 5080 | 602 | ~100 | ~83% |
| RTX 5070 Ti | 530 | ~88 | ~83% |
| RTX 5070 | 386 | ~64 | ~83% |
| RTX 5060 Ti | 325 | ~54 | ~83% |
| RTX 5060 | 258 | ~43 | ~83% |
The reduction is roughly consistent across the stack because it is a function of the MFG multiplier itself rather than any single GPU’s performance ceiling. That consistency is part of why XenoSpectrum’s analysis has spread quickly among PC hardware readers: it applies to any DLSS 5 title using 6x Multi Frame Generation, not just NBA 2K27.
RTX 5090 Founders Edition Hits Its 575W Ceiling
The FPS-math story landed the same week as a separate finding: an unofficial DLSS 5 benchmark reported by XenoSpectrum on September 3 and 4 shows the RTX 5090 Founders Edition reaching its configured 575W power limit while running DLSS 5 neural rendering. That’s the card’s board power ceiling, meaning the GPU is running into a hard cap rather than simply drawing more power on average.
Neural rendering adds an inference workload on top of traditional shading and ray tracing, and that extra compute has to draw power from somewhere. When a flagship card is already bumping against its power limit, it typically means the GPU is throttling clocks to stay within that ceiling, which can quietly erode some of the performance gains DLSS 5’s frame generation is supposed to deliver in the first place. This is a distinct issue from the frame-counting question above, but the two combine into the same broader complaint from reviewers: DLSS 5’s biggest numbers come with the biggest costs, in both rendered-frame reality and electricity.
Why an 802W Custom Board Doesn’t Solve the Problem
XenoSpectrum’s power analysis also compared the 575W Founders Edition against a custom-design RTX 5090 board rated in the 802W class, the same figure that anchored the site’s earlier DLSS 5 performance and power benchmarking report. The higher-wattage board does not simply scale performance up in proportion to its extra power headroom, and the outlet’s analysis specifically cautions that connector count and rated power alone don’t explain the performance gap between the two designs. That points to board-level power delivery and thermal engineering, not just how many watts a card is allowed to pull, as the deciding factor in how well a given RTX 5090 handles DLSS 5’s neural rendering load.
For buyers, that’s a more complicated shopping decision than “buy the highest-wattage card.” Two RTX 5090 boards with different power limits and cooler designs can behave differently under the exact same DLSS 5 workload, which is not something a spec sheet communicates clearly.
Independent Game Benchmarks Back Up the Pattern
Outside of Nvidia’s own NBA 2K27 demo, independent testers have been running DLSS 5 in shipping games since a leaked build began circulating in late August. French outlet Les Numériques tested Control and Cyberpunk 2077 on RTX 5070 Ti and RTX 5070 hardware, and the pattern matches what the frame-generation math predicts: enabling DLSS 5 neural rendering roughly halves the frame rate a card would otherwise produce.
| Game / GPU | Native or Prior Mode FPS | DLSS 5 Neural Rendering FPS | Drop |
|---|---|---|---|
| Control, RTX 5070 Ti, 4K | 71 | 35 | ~51% |
| Control, RTX 5090, 4K | 91 | 50 | ~45% |
| Cyberpunk 2077, RTX 5070, 1440p, rasterization | 138 | 68 | ~51% |
| Cyberpunk 2077, RTX 5070, 1440p, RT Ultra | 111 | 55 | ~50% |
| Cyberpunk 2077, RTX 5070, 1440p, path tracing | 71 | 45 | ~37% |
Those figures, published by Les Numériques and echoed in Club386’s own coverage, are separate from the MFG multiplier discussion above, they reflect the base rendering cost of DLSS 5’s neural rendering models before any frame generation is layered on top. Combined with the divide-by-six math, the picture for players is consistent: DLSS 5 does more computational work per frame, and it costs frame rate before Multi Frame Generation ever starts padding the display counter back up.
Club386’s Modded DLSS 5 Test: 90 FPS to 40 FPS
Club386 got hands-on with an early, leaked build of DLSS 5 running in multiple games on August 28, ahead of the official launch. Using a Core i7-14700F paired with a GeForce RTX 5080, the outlet recorded frame rates falling from roughly 90 FPS to about 40 FPS once DLSS 5 was switched on, describing the result as more than halving frame rates. Club386’s launch-week coverage, published September 3, put the same conclusion in plainer terms: turning on DLSS 5 can practically halve frame rates on a GeForce RTX 5090, while also requiring more watts from the wall.
That wording, chosen independently by Club386’s editors, has effectively become the shorthand the rest of the PC hardware press has used to describe DLSS 5’s early reception, and it lines up with both the Les Numériques game testing and the XenoSpectrum frame-math analysis above.
The Skyrim Mod Stress Test
One of the more unusual data points came from a modder who got an unreleased build of DLSS 5 running inside Skyrim, a 2011 engine never designed with neural rendering in mind. XenoSpectrum covered the mod on August 29: in the game’s character creation screen, frame rates ran between roughly 80 and 120 FPS with DLSS neural rendering disabled, then dropped to around 60 FPS once it was switched on. Outdoor footage, a heavier rendering scenario, showed frame rates in the 45 to 50 FPS range with DLSS 5 active.
The Skyrim numbers matter because they show the same roughly 30 to 50 percent performance tax showing up in a fourteen-year-old game engine, not just in showcase titles like NBA 2K27, Control, or Cyberpunk 2077. That suggests the cost is tied to DLSS 5’s neural rendering pipeline itself rather than any specific engine’s optimization work.
Nvidia’s Own 50-60% Admission
Notably, Nvidia has not disputed the scale of the performance cost. In a press briefing covered by XenoSpectrum, the company told reporters that DLSS 5 neural rendering alone can reduce frame rate by roughly 50 to 60 percent, depending on the game, scene, and resolution, before any frame generation is applied to recover perceived smoothness. That figure is Nvidia’s own characterization of the technology’s base rendering cost, and it is broadly consistent with what Club386, Les Numériques, and the Skyrim mod test all measured independently.
In other words, the controversy isn’t really about whether DLSS 5 is expensive to run, Nvidia has said as much itself. It’s about whether the marketing frame rates built on top of that cost, inflated sixfold by Multi Frame Generation, give buyers an honest picture of what their GPU is actually doing.
DLSS 5 vs DLSS 4.5 Ray Reconstruction
The contrast with Nvidia’s prior release makes the DLSS 5 story sharper. Club386 published hands-on testing of DLSS 4.5 Ray Reconstruction on an RTX 4090 on August 25, and found that the swap to a transformer-based model meaningfully improved image quality over earlier versions while having little performance impact. That’s the opposite trade-off from DLSS 5: a clear quality win with a small frame rate cost, versus a frame rate cost Nvidia itself puts at 50 to 60 percent, offset on paper by a frame counter that is padded with synthesized frames.
The two releases are related but distinct: DLSS 4.5 Ray Reconstruction improves how ray-traced lighting is denoised, while DLSS 5’s neural rendering and Multi Frame Generation attempt something more ambitious, generating a larger share of the final image with AI rather than traditional rasterization or ray tracing. The bigger ambition appears to carry a bigger performance bill.
Where AMD’s FSR Fits Into the Picture
DLSS 5 is exclusive to Nvidia’s RTX 50-series desktop and laptop GPUs, plus GeForce NOW’s RTX 5080 cloud instances, according to Nvidia’s own launch details, which were covered in Tech Insider’s report on the NBA 2K27 launch. RTX 40-series and RTX 30-series cards do not have official DLSS 5 support. AMD’s FSR upscaling and frame generation, by contrast, remains vendor-agnostic and runs across a wider range of hardware, including Nvidia and Intel GPUs.
Detailed, apples-to-apples FSR-versus-DLSS-5 benchmark numbers in the same scenes have not yet been widely published, since most of the last month’s testing has focused on DLSS 5 in isolation. But the broader positioning is already clear from the coverage so far: FSR has generally been framed by reviewers as a lighter-weight approach than DLSS 5’s neural rendering pipeline, at the cost of the image-quality gains Nvidia is promising with its heavier compute model.
Market Impact for RTX 50-Series Buyers and AIB Partners
The timing compounds an already difficult market. RTX 50-series pricing has climbed sharply this year, a trend Tech Insider covered when Nvidia RTX prices jumped as much as 39 percent amid broader AI-chip demand pressure. Buyers now paying a premium for an RTX 5090 or RTX 5080 are the same buyers being told that the marquee feature meant to justify that premium, DLSS 5’s frame generation, produces headline numbers that overstate real rendering performance by roughly six times, and that flagship cards can run into their own power ceilings while doing it.
For add-in board partners, the 575W-versus-802W-class comparison raises a design question that will likely show up in marketing for next-generation custom cards: cooling and power delivery, not just headline wattage, now directly affect how well a card handles Nvidia’s own showcase feature. That’s a more technical selling point than board partners have typically needed to make, and it suggests DLSS 5 could become a differentiator between otherwise similar custom RTX 5090 designs.
Historical Context: From DLSS 1 to Neural Rendering
Nvidia has billed DLSS 5 as its biggest graphics breakthrough since it introduced real-time ray tracing in 2018, and the SIGGRAPH announcement covered by XenoSpectrum in July described a fall 2026 rollout with support planned for titles including Assassin’s Creed Shadows, Hogwarts Legacy, Resident Evil Requiem, and Starfield. Each prior DLSS generation added a new pillar, upscaling, then frame generation, then ray reconstruction, and each carried its own adjustment period before reviewers and players settled on how to judge it. DLSS 5’s neural rendering and per-object AI control represent the most ambitious version of that pattern yet, replacing a larger share of traditional rendering with learned models rather than just cleaning up or interpolating around it.
What’s different this time is the size of the gap between the displayed number and the underlying workload. Earlier frame generation modes, at 2x or 4x multipliers, already drew scrutiny over input latency and interpolation artifacts. A 6x multiplier widens that gap enough that outlets are now doing the arithmetic in public rather than taking the on-screen counter at face value, and the comparison Tech Insider previously drew in RTX 4090 versus RTX 5080 ray tracing testing, where DLSS 5 support itself became a dividing line between generations, only sharpens that scrutiny.
What Happens Next
Based on the pattern of coverage over the past two weeks, several outcomes look likely in the near term.
- Reviewers will increasingly report rendered FPS alongside displayed FPS for any DLSS 5 title using Multi Frame Generation, rather than quoting Nvidia’s multiplied figures on their own.
- Expect closer scrutiny of RTX 5090 board designs specifically, given the 575W-versus-802W-class power gap XenoSpectrum flagged, as more independent reviewers test additional custom cards.
- Game-by-game variance will keep widening: path tracing titles like Cyberpunk 2077 already show a smaller relative FPS hit than rasterization or standard ray tracing modes in Les Numériques’ testing, and that pattern is likely to hold as more games add DLSS 5 support this fall.
- Pressure will build on Nvidia to publish rendered-frame figures directly, rather than leaving outlets to reverse-engineer them from the MFG multiplier, especially as more RTX 5060 and RTX 5070-class cards reach players who are more sensitive to the underlying power and frame rate cost.
- AMD’s FSR is likely to lean on its cross-vendor compatibility and lighter compute cost as a marketing contrast once direct FSR-versus-DLSS-5 benchmarks in shared scenes start appearing.
Frequently Asked Questions
Why is Nvidia DLSS 5’s real FPS so much lower than the advertised number?
Nvidia’s displayed FPS figures include frames generated by 6x Multi Frame Generation, meaning five of every six frames shown are synthesized rather than freshly rendered. XenoSpectrum’s analysis divides the displayed figure by six to estimate the actual rendered frame rate, which lands far below the marketing number.
Does DLSS 5 really cut performance by 50 to 60 percent?
Yes, according to Nvidia itself. The company told press in a briefing covered by XenoSpectrum that DLSS 5 neural rendering alone can reduce frame rate by roughly 50 to 60 percent before frame generation is applied, a figure that lines up with independent testing from Club386 and Les Numériques.
Why is the RTX 5090 hitting its power limit with DLSS 5 on?
An unofficial benchmark reported by XenoSpectrum on September 3 and 4 showed the RTX 5090 Founders Edition reaching its 575W configured power ceiling while running DLSS 5’s neural rendering workload, which adds an inference cost on top of standard shading and ray tracing.
Which GPUs support DLSS 5?
DLSS 5 is limited to GeForce RTX 50-series desktop and laptop GPUs, plus GeForce NOW’s RTX 5080 cloud instances at launch. RTX 40-series and RTX 30-series cards do not have official support.
Is DLSS 5 the same as DLSS 4.5 Ray Reconstruction?
No. DLSS 4.5 Ray Reconstruction, which Club386 tested on an RTX 4090, focuses on improving denoising for ray-traced lighting with little performance cost. DLSS 5 goes further, using neural rendering and per-object AI control to generate a larger share of the final image, which carries the much larger performance cost described above.
Does the performance hit happen in every game, or just showcase titles like NBA 2K27?
Independent testing suggests it’s not limited to showcase titles. A modder running an unreleased build of DLSS 5 inside the fourteen-year-old Skyrim engine measured a similar roughly 30 to 50 percent frame rate drop, and separate tests in Control and Cyberpunk 2077 by Les Numériques showed reductions in the same range.
Is Multi Frame Generation the same thing as fake frames?
Multi Frame Generation uses a neural network to interpolate additional frames between conventionally rendered ones, rather than computing each one from fresh game state. Whether to call those frames “fake” is a matter of framing, but they do not carry new input data the way a rendered frame does, which is why outlets are now distinguishing displayed FPS from rendered FPS.
Should I upgrade to an RTX 5090 or RTX 5080 for DLSS 5?
That depends on whether the visual smoothness of Multi Frame Generation matters more to you than raw rendered frame rate and power draw. Buyers focused on competitive responsiveness or running near a card’s power ceiling should weigh the rendered-FPS estimates and the RTX 5090 Founders Edition’s 575W ceiling alongside Nvidia’s headline figures before deciding.
Related Coverage
- Nvidia DLSS 5 Launches Sept. 3 in NBA 2K27: 370 FPS Claim [2026]
- Nvidia DLSS 5 Tanks FPS 48%, Guzzles 802W [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]
