Only one game currently has verified, model-specific DLSS 5 figures, and they apply to one GPU. DLSS 5 is scheduled to launch on September 3, 2026 at 21:00 Pacific. The confirmed launch figures come from NBA 2K27 on a desktop GeForce RTX 5090 with the full DLSS stack active. Keep that narrow test boundary in mind when judging any number published the next day.
A benchmark log must separate the desktop GeForce RTX 5090, the wider RTX 50 Series, and the DLSS 5 feature supplied through the game and driver. Treating them as interchangeable makes a launch result impossible to compare.
GeForce RTX 5090 DLSS 5 benchmark: the launch context
The verified scope covers only the desktop GeForce RTX 5090. DLSS 5 is officially supported as an RTX 50 Series feature, with release gated to NVIDIA’s stated time of September 3, 2026 at 21:00 PT. Results collected before then are preliminary. Results published after release still apply only to the desktop card, not laptop variants, Max-Q parts, the GeForce RTX 5080, the GeForce RTX 5070 Ti, or another model in the family.
For a valid launch-day comparison, check the exact SKU, post-launch driver, DLSS 5 runtime, and game patch. The title must genuinely enable DLSS 5 on day one, which currently means NBA 2K27 plus any other game NVIDIA names in its launch materials. A mismatch means you’re looking at different hardware, a different feature, or an older feature generation.
NVIDIA’s published desktop RTX 5090 figures
The launch brief lists only two verified, model-specific results, both from NBA 2K27. They cover 4K and 1440p at the Ultra preset, with ray tracing and the full DLSS stack active. On this generation, that stack includes DLSS 5. NVIDIA qualifies both results as “up to” figures, not guaranteed minimums.
These are results from a specific run, not a sustained floor that every desktop card will hold. The gap is also substantial: 370 FPS at 4K versus 590 FPS at 1440p, or about 60 percent. That relationship is a useful sanity check. If an independent run reverses it or shows a much smaller gap, raw resolution probably isn’t the only limit and the GPU is unlikely to be acting alone.
NBA 2K27 carries a heavy CPU and animation workload, while the reported results use ray tracing, DLSS Super Resolution, DLSS Ray Reconstruction, and DLSS 5 frame generation together. The 590 FPS result at 1440p indicates what the system can deliver when the engine is not the main limit. At 4K, the 370 FPS result shows pixel shading placing more pressure on the neural pipeline. A launch test should reproduce a similar relationship, though not necessarily the exact frame rates.
| Game | Resolution | Settings profile | Reported figure | Qualifier |
|---|---|---|---|---|
| NBA 2K27 | 3840 x 2160 (4K) | Ultra preset, ray tracing on, full DLSS stack including DLSS 5 | 370 FPS | up to |
| NBA 2K27 | 2560 x 1440 (1440p) | Ultra preset, ray tracing on, full DLSS stack including DLSS 5 | 590 FPS | up to |
DLSS 5 within the RTX 50 Series stack
DLSS 5 isn’t a self-contained graphics toggle. It sits within a stack built across several generations, with each layer doing a different job. A benchmark that merely says “DLSS on” leaves too much unspecified.
- DLSS Super Resolution reconstructs a higher-resolution image from a lower internal render. It has shipped on multiple generations and is the foundation of the stack.
- DLSS Ray Reconstruction replaces hand-tuned denoisers in ray-traced passes with a neural network. It is the layer most affected by scene complexity and ray-traced effect count.
- DLSS Frame Generation inserts synthesized frames between rendered ones, increasing perceived frame rate without raising the underlying render cost.
- DLSS 5, on RTX 50 Series hardware, extends that pipeline with additional neural rendering work. On a desktop GeForce RTX 5090, it activates when the game, the driver, and the runtime all support it.
The verified line for “Ultra preset, ray tracing, full DLSS stack” therefore describes several active systems. Turning off any layer can measurably change the result. Isolating DLSS 5 requires the same scene with and without that layer while every other setting remains fixed, and a shipping game with one combined toggle may not make that possible.
What changes at the September 3 release time
Until 2026-09-03 21:00 PT, public material describes DLSS 5 as scheduled for launch. RTX 50 Series support is official, but performance data, driver packages, and the complete list of enabled games remain gated to the release time.
A screenshot from a pre-release build hitting an internal frame-rate target isn’t a public benchmark. Its configuration may already be outdated, the hardware may not be the desktop card, and readers cannot yet use the feature. Before release, the only figures worth quoting are NVIDIA’s two NBA 2K27 results shown above.
| Phase | Feature state | Public figures | What a benchmark can claim |
|---|---|---|---|
| Pre-launch (before 2026-09-03 21:00 PT) | DLSS 5 is scheduled to launch | None officially published for any specific SKU | No model-specific FPS for desktop GeForce RTX 5090 with DLSS 5 |
| Launch (2026-09-03 21:00 PT) | DLSS 5 launches for the RTX 50 Series | NBA 2K27 figures at 4K and 1440p for desktop GeForce RTX 5090 | Reproduction of those figures under matching settings, or honest variance with reasons |
| Post-launch (after the release time) | DLSS 5 is available as a feature on RTX 50 Series | Same NBA 2K27 figures, with more titles expected as they patch in | Per-title comparisons only after a title’s official DLSS 5 patch is confirmed |
Reading the NBA 2K27 results
The 590 FPS result at 1440p, rather than a hypothetical 700 FPS, and the smaller-than-half drop to 370 FPS at 4K point to the same issue: the bottleneck moves as resolution changes.
At 1440p with the full DLSS stack on the desktop GeForce RTX 5090, the GPU is rarely the binding constraint in NBA 2K27. The game’s CPU and animation work still has to complete for every frame. At 590 FPS, the result is closer to the limit of the rest of the pipeline than the card itself. A slower CPU or storage subsystem can therefore reduce the 1440p result even when the GPU matches.
At 4K, each rendered pixel covers four 1080p pixels and the back end of the neural stack has more work to do. The drop from 590 to 370 FPS, about 37 percent, fits a workload that is partly GPU-bound at 1440p and more GPU-bound at 4K, yet still not fully saturated. A launch test with nearly identical 4K and 1440p results is likely constrained elsewhere, often by frame-time variance, VSync behavior, or an engine frame-rate cap.
Neither result covers minimum frame rate, 1 percent lows, power draw, noise, or scenes with especially heavy ray-traced reflections. Those measurements have to be captured separately before anyone can judge whether 370 or 590 FPS actually feels smooth.
Building a launch-day test plan
Start by trying to reproduce NVIDIA’s NBA 2K27 figures as a sanity check. Then widen the test to other games, resolutions, and DLSS configurations. Both approaches are worthwhile if the configuration is documented and held constant.
Use a system that won’t distort the result
At this performance level, an unlisted test system makes the result nearly meaningless. NVIDIA’s NBA 2K27 setup did not bottleneck the GPU at 4K or 1440p, and the ratio between those resolutions is the best clue to the platform’s headroom.
- A recent high-clock desktop CPU, because 590 FPS at 1440p in a sports title with this much animation work leaves almost no slack for a weak CPU.
- At least 32 GB of system memory at a speed that matches the platform, because frame pacing at high frame rates is sensitive to memory bandwidth and latency.
- NVMe storage, because modern sports titles stream a great deal of data per frame and a slow drive shows up as frame-time spikes long before it shows up as a lower average FPS.
- A case or test bench with sufficient airflow, because the desktop GeForce RTX 5090 is a high-power part and thermal throttling will silently cut sustained frame rates well before it shows up in a single run.
- A known-good monitor or capture path, because a high-frame-rate run is wasted if the display or the capture device cannot accept that signal cleanly.
Treat any result without its CPU, memory speed, storage, and cooling setup as preliminary. Otherwise, there’s no way to tell whether 370 FPS at 4K reflects a GPU limit or a system limit.
Lock the driver, BIOS, and OS baseline
The driver and DLSS 5 runtime must be the versions NVIDIA publishes on September 3, 2026 at 21:00 PT or later. An earlier build doesn’t count as a DLSS 5 run, even if the menu exposes a toggle. Record the driver build, DLSS runtime, card VBIOS, and Windows build so another tester can match the setup.
Windows settings matter too. Game Mode, Hardware-Accelerated GPU Scheduling, Variable Refresh Rate behavior, and MPO (Multi-Plane Overlay) paths on Windows 11 interact with frame generation and capture overlays. A different combination from NVIDIA’s NBA 2K27 setup prevents a strict comparison, even when the average FPS looks close.
Lock the in-game settings
To compare against the verified figures, use the Ultra preset, enable ray tracing, and select the full DLSS stack. Then record three settings that often introduce large differences.
- Frame cap. If the engine has a built-in cap, it will show up as the published number regardless of the GPU. Many sports titles ship with a cap that needs to be raised or disabled before a high-refresh test is meaningful.
- VSync. VSync off in the engine, plus a VRR display that is set to its maximum refresh rate, is usually the cleanest way to capture the GPU’s natural output. Any other combination is measuring the sync path, not the GPU.
- DLSS mode. The DLSS preset must be at the same quality or performance mode the published run used. Mixing “Quality” and “Performance” presets between two test runs will produce differences that have nothing to do with the GPU SKU.
Without all three settings, results from otherwise identical hardware still aren’t directly comparable.
Capture more than average FPS
Average FPS is easy to publish, but it hides most pacing problems. A useful launch-day test for the desktop GeForce RTX 5090 running DLSS 5 needs several other measurements to show how the game actually behaves.
Frame-time variance shows whether 370 FPS at 4K feels as fast as it sounds. Tight variance and a strong 1 percent low produce a smoother result; a similar average with loose variance feels fast but uneven. Frame generation makes that distinction more obvious because synthesized frames inherit the timing and jitter of the base frames.
| Metric | What it actually measures | Why it matters for DLSS 5 |
|---|---|---|
| 1 percent low frame time | The worst 1 percent of frame intervals | DLSS 5 frame generation smooths averages but not necessarily worst-case intervals, so a high 1 percent low is the strongest signal of a real-feeling run |
| Frame-time variance | The spread of frame intervals around the median | A 590 FPS run with high variance can feel worse than a slower run with low variance, and DLSS 5 amplifies that perception |
| GPU power draw | Average and peak watts at the PCIe power connectors | The desktop GeForce RTX 5090 is a high-power part, and a frame-per-watt comparison is the only fair way to talk about efficiency |
| GPU temperature | Junction or hotspot temperature over a 10 to 15 minute run | Thermal throttling on a poorly cooled card will silently cut the back half of any long benchmark |
| CPU utilization per core | Per-core load during the same run | Identifies whether a low number is a GPU limit or a CPU limit, which is the central question at 1440p with DLSS 5 |
Why RTX 5090 results don’t transfer to other models
The desktop GeForce RTX 5090 sits at the top of the consumer RTX 50 Series. DLSS 5 support extends across the family, but performance does not. RTX 5090 results cannot support claims about the GeForce RTX 5080, GeForce RTX 5070 Ti, laptop parts, or Max-Q variants because their shader counts, memory subsystems, and power limits differ.
The launch brief therefore assigns its two verified figures specifically to the desktop GeForce RTX 5090 without translating them to other models. Even a partner overclock or a different cooler design creates a separate test, despite the same model name on the box.
What day-one testing cannot establish
One day of testing cannot answer every question about DLSS 5.
- It cannot tell you how DLSS 5 behaves in games that have not yet shipped a patch that enables the feature. The list of day-one titles is the list, and anything outside it is a different feature or a different test.
- It cannot tell you how DLSS 5 compares to a hypothetical alternative, because no competing neural rendering stack is currently shipping on this hardware under the same conditions.
- It cannot tell you how the feature will age, because long-term driver work, game patches, and OS changes will all move the numbers.
- It cannot tell you whether the perceived smoothness matches the headline FPS, because that question lives in frame-time data, not in averages, and it depends on the display as much as on the GPU.
Any launch result is a snapshot, not a final verdict. Repeat it across several driver versions, game patches, and at least one or two more titles after they confirm DLSS 5 support.
Checking the verified sources
Two linked sources define the available evidence. The Wikipedia entry gives family-level context for the desktop RTX 5090, while NVIDIA’s research page documents the DLSS 5 launch and NBA 2K27 performance figures. They describe the same launch event from different perspectives.
Wikipedia confirms that the desktop RTX 5090 belongs to the RTX 50 Series generation with DLSS 5 support. NVIDIA ties the headline figures to one game, two resolutions, and a defined settings profile. Together, those sources establish the GPU model and feature generation a comparable test must use.
To verify the result on another bench, confirm the desktop RTX 5090, NBA 2K27, the Ultra preset with ray tracing, the full DLSS stack including DLSS 5, and post-launch driver and runtime versions. If one condition differs, report the measurement on its own and state the difference instead of comparing it directly.
Launch-day publishing checklist
Before quoting a result from the first 24 hours after release, check all seven items.
- The GPU is specified as the desktop GeForce RTX 5090, not a laptop, Max-Q, Ti, or SUPER variant.
- The driver and DLSS runtime are post-launch versions tied to September 3, 2026 at 21:00 PT or later.
- The game is one whose patch is confirmed to enable DLSS 5 on day one.
- The settings are stated in full: preset, ray tracing state, DLSS mode, frame cap, and VSync state.
- Frame-time data is included alongside average FPS, not just a single headline number.
- The system configuration is listed, including CPU, memory, storage, cooling, and display refresh rate.
- The number is presented as a measured value, with the qualifier “up to” if it is being compared to the published figures.
Items one through four establish whether the comparison is valid. Items five through seven determine whether it is informative. Developers and test engineers can work with a result that meets all seven; one that meets only the first three is little more than a curiosity.
Using the launch figures as a baseline
After release on September 3, 2026, DLSS 5 becomes an official RTX 50 Series feature on the desktop GeForce RTX 5090 and the rest of the family. The verified figures define one game, two resolutions, and one settings profile, giving independent testers a common baseline.
For game developers, that baseline turns a marketing claim into a measurement that can be checked. Future NBA 2K27 patches, driver updates, and newly supported games may move the numbers. The 370 FPS at 4K and 590 FPS at 1440p results remain the initial reference for tracking those changes.
Frequently asked questions
What is the only verified number for a GeForce RTX 5090 DLSS 5 benchmark right now?
Right now, the only verified, model-specific numbers in the launch brief are the NBA 2K27 figures: 370 FPS at 4K and 590 FPS at 1440p, both on the Ultra preset with ray tracing on and the full DLSS stack active, including DLSS 5. Both are qualified as “up to” values, which means they are specific run results rather than guaranteed minimums.
Does the desktop GeForce RTX 5090 officially support DLSS 5?
Yes. DLSS 5 is officially supported as an RTX 50 Series product, and the desktop GeForce RTX 5090 is part of that series. The launch is gated to September 3, 2026 at 21:00 PT, and the publicly verified performance numbers are tied to that release time.
Can I use the NBA 2K27 figures to claim performance in other games?
No. The 370 FPS and 590 FPS figures are tied to NBA 2K27 at a specific preset and a specific DLSS configuration. Other games have different engine costs, different ray-traced workloads, and different DLSS 5 integration, so a GeForce RTX 5090 DLSS 5 benchmark in a different title will produce different numbers that cannot be inferred from the NBA 2K27 results.
Are the published figures the same for laptop or Max-Q variants of the RTX 5090?
No. The verified numbers are for the desktop GeForce RTX 5090 only. Laptop, Max-Q, Ti, SUPER, and other neighboring variants have different power envelopes and different shader counts, and any number from those parts is a separate test that should not be conflated with the desktop figures.
Why is the 1440p figure so much higher than the 4K figure in NBA 2K27?
At 1440p with the full DLSS stack, the desktop GeForce RTX 5090 is rarely the bottleneck in NBA 2K27, and the 590 FPS value is closer to the limit of the rest of the system than to the limit of the GPU. At 4K, the GPU does more work per frame and the number drops to 370 FPS, which fits a workload that was partially GPU-bound at 1440p and is more GPU-bound at 4K.
Do the NBA 2K27 numbers include DLSS 5 frame generation, or only the older DLSS layers?
The settings line in the verified table reads “full DLSS stack including DLSS 5,” which means DLSS Super Resolution, DLSS Ray Reconstruction, DLSS Frame Generation, and the DLSS 5 layer are all active at once. A test that wants to isolate DLSS 5 itself has to compare the same scene with and without the DLSS 5 layer while keeping the rest of the stack constant.
What is the minimum system I need to reproduce the 590 FPS figure at 1440p?
To approach the 590 FPS number at 1440p, the system has to be roughly in the same class as the one NVIDIA used, which means a recent high-clock desktop CPU, at least 32 GB of system memory, fast NVMe storage, and enough cooling headroom that the desktop GeForce RTX 5090 does not throttle. A slower CPU or a slower storage subsystem will reduce the 1440p number even with the same GPU.
Is a GeForce RTX 5090 DLSS 5 benchmark from before the release time valid?
Pre-release numbers for the desktop GeForce RTX 5090 with DLSS 5 are preliminary by definition, because the feature and the matching driver are gated to the release time. A pre-release figure cannot be compared directly to the post-launch NBA 2K27 results, and it should be reported as a separate data point rather than as a GeForce RTX 5090 DLSS 5 benchmark in the public sense.
Why does headline FPS not tell the full story for DLSS 5?
DLSS 5 frame generation raises averages but doesn’t remove worst-case frame intervals, so a 590 FPS run with high frame-time variance can feel less smooth than a slower run with tight variance. A serious benchmark reports 1 percent lows and frame-time data alongside the average instead of relying on one figure.
How should I treat a launch-day result that disagrees with the published figures?
A launch-day result that is lower than the published figures is normal and expected, because the published numbers are “up to” values from a specific run, not a floor. A result that is meaningfully higher, or that lands at a very different ratio between 4K and 1440p, is a sign that something other than the GPU is bottlenecking the test, and the system configuration should be checked before the number is trusted.
Leave a Reply