GeForce RTX 4090 DLSS 4.5 vs DLSS 5: a baseline and outlook for game teams
The GeForce RTX 4090 remains a common high-end QA reference for 4K path-traced builds. Its current stack is DLSS 4.5, including transformer-based Super Resolution and Ray Reconstruction plus Frame Generation.
NVIDIA has not announced DLSS 5 support for this card, so any comparison has to stop at measured DLSS 4.5 behavior and treat the next version as an open planning question.
The useful evidence comes from public measurements on the 4090 itself. They show what the card can deliver in shipping games and give teams a baseline they can repeat if NVIDIA later documents DLSS 5 for Ada Lovelace.
What ships in the DLSS 4.5 stack
DLSS 4.5 combines several techniques behind one runtime and model-delivery pipeline. Each solves a different rendering cost:
- Super Resolution reconstructs a higher-resolution image from a lower-resolution render. It can replace or supplement a game’s TAAU path and provides the base for the other DLSS features.
- Ray Reconstruction uses spatiotemporal samples and a guide buffer as a learned replacement for traditional denoisers. It can reduce the sample count needed for stable path-traced lighting.
- Frame Generation uses Ada Lovelace’s Optical Flow Accelerator to insert a frame between two rendered frames, increasing perceived frame rate without adding the same rasterization load.
The RTX 4090 has 16,384 CUDA cores, 128 third-generation RT cores, 512 fourth-generation Tensor cores, and 24 GB of GDDR6X on a 384-bit bus.
That hardware gives 4K ray tracing enough headroom for the neural stages to fit inside the frame budget.
The Super Resolution and Ray Reconstruction networks in DLSS 4.5 use a transformer backbone. In practice, the newer models sharpen hair, foliage, and thin geometry while holding together better in motion-heavy scenes. Studios do not need to rebuild their renderer to adopt them. A new model file can move an existing integration from the older CNN route to the transformer route as a content-only update.
DLSS 5 remains unannounced for the RTX 4090
No public roadmap names the 4090 for DLSS 5 launch support. There is also no official basis for a frame-rate claim, release window, or new API description on this card. Until NVIDIA publishes documentation, SDK headers, or a supported driver, DLSS 4.5 is the shipping ceiling.
The uncertainty matters because the card still appears in studios and enthusiast systems. Live-service producers need current integrations to survive for years, while graphics programmers and technical artists need to know whether future quality modes require separate rendering paths.
Ship against 4.5, keep vendor calls behind a small abstraction, and avoid hard-coding a model name or runtime version that must survive several driver generations. That approach prepares the renderer for an update without promising one.
Measured 4K results from shipping games
These are the only FPS figures used here. They come from direct measurements on the GPU’s review page, so the hardware is exact and the test method is public.
| Game and version | Resolution | Settings | Average FPS |
|---|---|---|---|
| Cyberpunk 2077 1.6 | 3840×2160 | Ray Tracing Ultra Preset, DLSS off | 43.0 |
| Cyberpunk 2077 | 3840×2160 | Ray Tracing Overdrive + DLSS 4 Transformer Quality + Frame Generation 2x | 89.5 |
| Alan Wake 2 | 3840×2160 | High Preset + Ultra Ray Tracing + Quality DLSS + Frame Generation 2x | 103.7 |
Cyberpunk 2077 1.6 averages 43.0 FPS at native 4K with the Ray Tracing Ultra preset and DLSS disabled. That is playable, but it is not a comfortable high-refresh result.
Ray Tracing Overdrive with DLSS 4 Transformer Quality and Frame Generation 2x reaches 89.5 FPS. Alan Wake 2 reaches 103.7 FPS with its High preset, Ultra Ray Tracing, Quality DLSS, and Frame Generation 2x, leaving clear room above a 60 Hz target.
The jump comes from reconstruction and generated frames rather than more raw shading. The 4090 can feed those stages at 4K, but a future DLSS version would still need to improve the result without breaking the existing integration.
Known behavior on one side, open questions on the other
DLSS 4.5 can be tested now. DLSS 5 cannot. The comparison below keeps that difference visible instead of turning expectations into specifications.
| Layer | DLSS 4.5 shipping behavior on RTX 4090 | DLSS 5 outlook (subject to official documentation) |
|---|---|---|
| Super Resolution model | Transformer backbone for sharper geometry and motion | A newer model class may retain the same integration surface, but no specific gain is public. |
| Ray Reconstruction | Replaces legacy denoisers with a learned spatio-temporal reconstructor | It may continue replacing hand-tuned denoiser chains; sample-budget changes have not been announced. |
| Frame Generation | Optical-flow interpolation, 2x path, with multi-frame variants on supported titles | An optical-flow and motion-vector path is plausible, but latency and quality targets are not public. |
| Game integration surface | SDK plus plugin chain, model files delivered through the NVIDIA runtime | Keeping the same SDK surface would preserve existing 4.5 integrations, but NVIDIA has not confirmed it. |
| Driver and runtime | Distributed through NVIDIA app and GeForce Game Ready Driver | Requirements will be clear only after a public release. |
The left column is a production target. The right column is a release-note checklist. For the surrounding hardware context, the RTX 40 series overview places Ada Lovelace within the wider RTX family.
Integration details that matter on this card
Most teams use NVIDIA’s SDK, an Unreal Engine or Unity plugin, or an equivalent route in a proprietary engine.
The hard work is supplying clean inputs, mapping presets into the game’s own menu, and keeping base frame pacing stable.
- Provide the rendered color buffer, depth, motion vectors, and exposure or HDR scale. Missing or low-precision motion vectors are a common source of ghosting and edge artifacts.
- Map Quality, Balanced, Performance, and Ultra Performance to the project’s existing graphics vocabulary and their respective internal render scales.
- Keep the base frame time steady when Frame Generation inserts an image. A high displayed FPS figure cannot hide stutter in the rendered frames.
- Check materials with strong roughness gradients and bright specular highlights. They make denoiser bias easier to spot when Ray Reconstruction is disabled.
Set an absolute minimum base frame time for the heaviest 4K path-traced scene. The measured 89.5 FPS in Cyberpunk 2077 is effectively a 90 FPS reference, while Alan Wake 2 reaches 103.7 FPS with Quality DLSS and Frame Generation 2x. Neither result replaces an internal worst-case capture.
Do not claim support before these checks pass
A public DLSS 5 line needs vendor evidence. No launch support or card-specific support has been announced, so a producer should require all of the following before marketing changes its copy:
- The feature appears in NVIDIA’s SDK changelog or release notes.
- A public DLSS 5 SDK or header set is available through the registered-developer channel.
- The current Game Ready Driver exposes the runtime path on the 4090.
- The driver release notes name the feature for this hardware.
- NVIDIA permits the title to appear under a DLSS 5 row on its supported-games page.
If one check fails, describe only the shipping DLSS 4.5 behavior. The Notebookcheck RTX 4090 benchmarks page is the direct public reference for the measurements used here.
How the 4K frame budget is divided
At native 4K, Cyberpunk 2077 1.6 reaches 43.0 FPS with Ray Tracing Ultra and DLSS off. That result covers the raster and ray-tracing load. The 89.5 FPS result with the transformer upscaler and Frame Generation 2x shows what happens when the internal raster and ray load is reduced and neural reconstruction is added.
A later neural tier would probably rebalance the same work rather than add an unrelated fourth block. It might push more reconstruction into the learned model or improve the generated frame while retaining optical flow. Neither outcome is confirmed for DLSS 5. They are engineering possibilities, not product claims.
Capture a baseline before the next runtime arrives
- Measure the heaviest 4K ray-traced scene with DLSS off. Cyberpunk 2077 1.6 at 43 FPS is a public reference for a demanding native workload.
- Repeat the scene with DLSS 4 Quality and Frame Generation off, recording GPU time, frame-time variance, and tensor-pass cost.
- Enable Frame Generation 2x. Cyberpunk 2077 at 89.5 FPS and Alan Wake 2 at 103.7 FPS provide upper reference points for current shipped games.
- Record VRAM with the Frame Generation buffers allocated. The 24 GB capacity is rarely the bottleneck, but the 384-bit bus and its bandwidth limit how many full-resolution G-buffers a heavier model can use.
- Log the SDK version, model revision, and render-pass connections so the later comparison uses the same integration.
Rerun that capture if DLSS 5 ships for the card. The difference between two controlled profiles is far more useful than a result from a rumor or preview build. If internal profiling is unavailable, the studio’s game development services page describes how that external work can be scoped.
The RTX 4090’s place in a 2026 QA rack
A typical 2026 matrix includes a current flagship, a previous flagship, a current midrange card, and an older or lower-tier GPU. The 4090 often fills the previous-flagship slot. Its 24 GB framebuffer can hold 4K G-buffers and neural-network intermediates, while its mature DLSS 4.5 integration and long measurement history make regressions easier to identify.
Keeping it in that rack is also a bet that Ada’s neural hardware remains relevant for another 12 to 24 months. The sensible way to manage that bet is to use the current SDK behind an abstraction and accept model or runtime updates through the standard NVIDIA delivery channel. The exact behavior of anything branded DLSS 5 still depends on future documentation.
Claims that do not match the evidence
- “DLSS 5 is already running on the 4090.” It is not supported by a launch announcement or public card-specific documentation.
- “Frame Generation doubles every GPU’s frame rate.” It can roughly double displayed output from a stable base, but inconsistent rendered frame times remain visible as judder.
- “The transformer model is only marketing.” It changes the Super Resolution and Ray Reconstruction networks and produces measurable improvements in motion stability.
- “A new DLSS feature will make the 4090 obsolete.” Ada has the Tensor and optical-flow hardware required by the current stack. A later tier may be additive, but NVIDIA has not confirmed how DLSS 5 maps to this silicon.
Where official confirmation will appear
- NVIDIA Developer Zone SDK changelogs and release notes list the runtime and minimum driver.
- Game Ready Driver notes identify supported titles and features.
- NVIDIA’s supported-games and supported-features pages change when a feature becomes official for a title.
- Registered-developer emails and Developer Zone announcements provide the lead time studios need for integration work.
These sources turn a prediction into a versioned statement that a production team can test.
Shipping DLSS 4.5 without closing the upgrade path
- Use the current SDK and transformer models for Super Resolution and Ray Reconstruction.
- Expose Quality, Balanced, Performance, and Ultra Performance using wording consistent with the rest of the graphics menu.
- Check a demanding 4K scene with Frame Generation 2x. Use 89.5 FPS in Cyberpunk 2077 and 103.7 FPS in Alan Wake 2 as public reference points, not universal targets.
- Keep motion vectors, depth, and exposure at full precision.
- Test Frame Generation on stable and variable base frame times. A native 43 FPS workload is the sort of case that benefits strongly from reconstruction and generated frames.
- Put the vendor integration behind a small internal interface so a model or SDK revision does not require a renderer rewrite.
- Review NVIDIA’s official channels when updates are documented, not when they leak.
What a later DLSS release would need to preserve
None of these conditions is confirmed for DLSS 5. They describe the compatibility qualities that would make a new runtime useful to existing 4090 owners:
- The existing SDK surface remains valid.
- Drivers continue to deliver the runtime through the standard channel.
- Model-file swaps allow content-only updates without a new game build.
- Existing transformer models remain compatible so unpatched titles can still receive runtime improvements.
- Documentation names every new quality preset and required input vector.
Those expectations follow earlier DLSS version transitions, but they are not a promise from NVIDIA.
Limits of the available measurements
All FPS figures here come from one RTX 4090 review page. Cyberpunk 2077 and Alan Wake 2 cover demanding 4K ray-traced AAA workloads, but they do not represent every game, scene, or preset. Studios should repeat the same kinds of captures on their own worst-case content.
There is no public DLSS 5 measurement for this card because support has not been announced. Every statement about that side of the comparison is an expectation, test condition, or production constraint rather than a measured result.
Common developer questions
Can the GeForce RTX 4090 run DLSS 5 today?
No official support has been announced. DLSS 4.5 is the current shipping tier, with transformer-based Super Resolution and Ray Reconstruction plus the existing Frame Generation path.
Which features make up DLSS 4.5 on this GPU?
It includes transformer-based Super Resolution, learned Ray Reconstruction, and optical-flow Frame Generation. They use the same SDK and model-delivery channel.
What do the public 4K results show?
Cyberpunk 2077 1.6 averages 43.0 FPS at 3840×2160 with Ray Tracing Ultra and DLSS off. Cyberpunk 2077 reaches 89.5 FPS with Ray Tracing Overdrive, DLSS 4 Transformer Quality, and Frame Generation 2x. Alan Wake 2 reaches 103.7 FPS at 3840×2160 with High settings, Ultra Ray Tracing, Quality DLSS, and Frame Generation 2x. No other FPS figures are used here.
Should a 2026 release wait for DLSS 5?
No. Ship the supported DLSS 4.5 path and keep the integration replaceable. Waiting for an unannounced feature adds schedule risk without a confirmed gain.
Will the current integration survive a DLSS 5 release?
NVIDIA’s eventual SDK changelog will decide that. Reusing the SDK surface and model-delivery route is a reasonable expectation, but it is not confirmed.
Does Frame Generation work the same way on newer cards?
On the 4090 it uses Ada Lovelace’s Optical Flow Accelerator. The SDK and driver control the feature, while the card’s base 4K frame time determines the result. The 89.5 FPS and 103.7 FPS measurements show the upper end of current 2x behavior in the two tested titles.
What should a team profile before a future update?
Capture the heaviest 4K ray-traced scene with DLSS off, DLSS 4 Quality without generated frames, and Frame Generation 2x. Record GPU time, frame-time variance, tensor cost, and VRAM use.
Which sources should teams monitor?
Use NVIDIA Developer Zone SDK notes, Game Ready Driver notes, supported-games and supported-features pages, and registered-developer announcements. They are the safe basis for production planning.








Leave a Reply