NVIDIA DLSS 5 NBA 2K27 Performance
For most of its existence, NVIDIA DLSS has been associated with a fairly simple proposition: use machine learning to make demanding graphics more practical. Super Resolution reconstructs a higher-resolution image from a lower-resolution render, Frame Generation inserts additional frames between conventionally rendered ones, and Ray Reconstruction uses a neural network to replace hand-tuned denoisers in ray-traced workloads. DLSS 5 takes the technology in a noticeably different direction. Making its debut with NBA 2K27, it introduces what NVIDIA calls 3D-Guided Neural Rendering, a neural graphics system designed primarily to improve the appearance of the rendered image rather than simply increase frame rates. It works as a later stage of the graphics pipeline, using the game’s existing rendered frame and engine data as the foundation upon which its neural model enhances lighting and materials.
Basically, DLSS 5 is not being asked to independently generate a scene. Geometry, textures and the underlying artistic composition continue to come from the game engine. Instead, the model interprets that information and can introduce richer material interactions, including more convincing skin subsurface scattering, light transmission through hair and foliage, deeper contact shadows and improvements to global illumination. In practical terms, NVIDIA is beginning to use AI not just to reconstruct pixels that would otherwise have been expensive to render, but to influence how those pixels ultimately look.
As we mentioned previously, DLSS 5 adds a different category of workload: 3D-Guided Neural Rendering, which is designed to enhance lighting and material detail in the final image rather than primarily improving performance. The system operates towards the end of the rendering pipeline. The game engine still produces the underlying frame, including its geometry, textures and lighting buffers, and that rendered frame becomes the fixed foundation for the neural model. NVIDIA describes the engine frame as determining what must remain unchanged, while developers retain control over what the neural renderer is allowed to enhance.
DLSS 5 takes information including frame colour and motion vectors as inputs, while the model is trained to recognise engine data such as surface albedo(how much sunlight a surface reflects), detailed lighting and surface normals. It uses this information to analyse objects, materials, spatial relationships and light sources before modifying how those materials and lighting interactions appear in the finished image. And that is where the “3D-Guided” part becomes important. This is not a text-to-image model generating a new frame from a prompt, nor is it simply applying the same visual filter to everything on screen. The neural renderer remains anchored to the game’s existing 3D scene and can enhance specific effects such as skin subsurface scattering, light transmission through hair and foliage, contact shadows and global illumination.
NVIDIA has also designed the system around temporal stability. Unlike video-generation models that process batches of frames, DLSS 5 works on a one-frame-in, one-frame-out basis and uses motion vectors supplied by the game engine to maintain consistency as objects and cameras move. The aim is to avoid artefacts such as shimmering, swimming and temporal drift. Obviously, there’s a higher compute overhead when you deal with a single frame at a time but you have to start somewhere. And the benefit of AI-powered rendering engines is that it will get smarter and more efficient over time.
The neural model itself runs locally on the Tensor Cores of GeForce RTX 50 Series GPUs. Crucially, 3D-Guided Neural Rendering is an independent part of the DLSS stack, which means developers can use it alongside Super Resolution, Multi Frame Generation and Ray Reconstruction rather than replacing them.
That changes the usual DLSS equation. Some parts of the DLSS stack are intended to reduce the cost of rendering or increase displayed frame rates, while DLSS 5 Neural Rendering can spend additional GPU compute to improve the visual quality of the final image. This comes to fore when looking at our RTX 5090 telemetry, because the sharp increase in power draw and utilisation is consistent with DLSS 5 adding another substantial neural processing stage rather than simply reconstructing the same image more efficiently. However, it should be stated that this is the beginning of DLSS 5 being introduced, things will smoothen over time and we expect the power draw to become leaner over time.
NBA 2K27 is a particularly suitable environment in which to demonstrate this approach. Sports games contain large numbers of human characters, recognisable faces, complicated skin and hair rendering, reflective surfaces and strong directional lighting. Camera work also moves constantly between wide gameplay views and close-ups of athletes, making subtle material and lighting differences relatively easy to spot.
Visual Concepts, the game developer, has consequently concentrated its implementation on elements where neural rendering can make a visible difference without changing the identity of the underlying assets. DLSS 5 is used to improve how light interacts with skin, ears, hair and clothing while retaining the facial geometry scanned from real athletes. NVIDIA highlights natural light transmission through ears, warmer skin under arena lighting, finer hair definition and stronger contact shadows around the neck and jersey as examples.
Similar changes are visible elsewhere in the game. We can see how subsurface scattering across the face, more precise contact shadows beneath the nose and chin, and stronger ambient occlusion across folds in clothing seem real-er. Courtside spectators and staff are also affected, rather than the feature being restricted exclusively to star players. So spectators no longer seem as afterthoughts, each person seems to have their own identity, although not as clear as the players.
The important point is that these are not higher-resolution textures being swapped in when DLSS 5 is enabled. The AI model is interpreting the rendered scene and changing how lighting and material interactions are presented.
One of the obvious problems with applying generative AI techniques to real-time graphics is consistency. A system that subtly changes a face, material or lighting treatment from one frame to the next could look impressive in screenshots but become distracting as soon as the scene moves.
NVIDIA’s answer is to tightly anchor DLSS 5 to engine data. The system takes information including colour and motion vectors and is trained to recognise data such as surface albedo, lighting and normals. NVIDIA says this lets it identify objects, materials, spatial relationships and light sources while keeping the generated result grounded in the original frame.
Temporal behaviour is equally important. Rather than operating like a video-generation system that considers sequences of frames in batches, DLSS 5 follows a one-frame-in, one-frame-out model and uses motion vectors supplied by the game engine. NVIDIA says this is intended to prevent problems such as shimmering, swimming and temporal drift as objects and cameras move.
A neural model can enhance the way light travels through an athlete’s ear or how shadows form around the face, but it cannot be allowed to reinterpret the actual facial proportions that Visual Concepts has captured.
For players, DLSS 5 can simply appear as an on/off option. Developers see a much more complex system behind it. Studios can choose between different neural models with different parameter weightings, and those choices do not necessarily need to remain constant throughout an entire game. One model can be used for a particular type of environment and another for a different scene or cutscene. NVIDIA also provides Structure Intensity and Tone Intensity controls, which affect different aspects of the final image.
Structure Intensity governs higher-frequency details including ambient occlusion, reflections and subsurface scattering, while Tone Intensity is intended for broader lighting and colour response. Together, they allow developers to decide how strongly the neural renderer alters the finished image.
Semantic AI masking provides another layer of control by recognising categories of objects in a scene, allowing developers to treat characters differently from backgrounds. Engine-level masking goes further by letting artists isolate specific props or groups of assets, such as foliage, glassware or water droplets, and apply neural lighting adjustments selectively.
This is arguably one of the more important aspects of DLSS 5. A universal “make this more photorealistic” filter would be of limited use in games where stylisation is deliberate. Giving artists control over where neural rendering is applied makes the technology part of the rendering and art pipeline rather than a post-processing effect imposed indiscriminately across everything on screen. There was a lot of hullabaloo when DLSS 5 was first introduced around the loss of artistic intent. That’s quite an important facet of maintaining an artist’s creative vision across the development journey. So NVIDIA providing this level of control should allay artists’ concerns. Unfortunately, as is the case with most AI implementation across industries, the management executives are likely to use this to shave off a few more jobs along the production pipeline.
DLSS 5 also does not make traditional rendering techniques obsolete. The quality of the result remains tied to the quality of the information that enters the neural model. NVIDIA says 3D-Guided Neural Rendering can noticeably improve rasterised graphics, but providing richer source information through ray tracing or path tracing leads to more accurate results. In effect, better lighting information gives the network a stronger basis from which to infer how materials and illumination should appear.
That makes DLSS 5 complementary to ray tracing rather than a replacement for it. Ray tracing can provide more physically meaningful source information, while neural rendering can use that information to produce additional visual detail that would otherwise be expensive to model explicitly. The cost of doing that becomes particularly interesting once GPU telemetry is examined. All this goes onto say that you’ll still continue to need better artists and more artists to ensure better output.
To get an initial idea of how demanding DLSS 5 is, we monitored a GeForce RTX 5090 under identical workload conditions with the feature disabled and enabled. Six sample sequences were collected with DLSS 5 off, immediately followed by six sample sequences of the same scenes with it switched on. These numbers should not be treated as a final judgement on the computational cost of DLSS 5. The technology is at an early stage, and the neural model, game integration and drivers could all be further optimised before broader deployment. The relatively short sample window also makes some parameters, particularly fan response and bursty bus activity, unsuitable for definitive conclusions.
Even with those caveats, the change in GPU behaviour was substantial.
| Metric | DLSS 5 off | DLSS 5 on | Change |
| GPU power | 166.6 W | 314.1 W | +88.5% |
| GPU power as % of TDP | 28.9% | 54.4% | +25.4 pp |
| GPU temperature | 59.3°C | 67.2°C | +7.8°C |
| GPU hot spot | 64.6°C | 75.8°C | +11.2°C |
| Memory junction | 66.3°C | 69.7°C | +3.3°C |
| GPU core load | 31.3% | 56.8% | +25.5 pp |
| D3D usage | 30.1% | 53.1% | +23.0 pp |
| Memory controller load | 3.3% | 8.8% | +5.5 pp |
| Reported core clock | 2,348 MHz | 2,789 MHz | +18.7% |
| Effective core clock | 2,380 MHz | 2,776 MHz | +16.6% |
| Memory clock | 1,750 MHz | 1,750 MHz | No change |
| Core voltage | 0.960 V | 1.045 V | +8.7% |
| VRAM usage | 34.3% | 34.3% | No change |
The clearest difference was power consumption. Average GPU power increased from 166.6 W with DLSS 5 disabled to 314.1 W with it enabled, a rise of 147.5 W or approximately 88.5 per cent. Total GPU power as a proportion of TDP increased correspondingly from 28.9 per cent to 54.4 per cent. There was still plenty of headroom on the RTX 5090, whose rated power limit is 575 W, so the observation here is not that DLSS 5 pushed the card anywhere close to a power ceiling. Rather, neural rendering turned what had been a relatively modest graphics workload into one that demanded considerably more sustained GPU activity.
GPU core utilisation rose from an average of 31.3 per cent to 56.8 per cent when DLSS 5 was enabled, while D3D usage increased from 30.1 per cent to 53.1 per cent. Those counters cannot tell us precisely how much of the additional workload belongs to Tensor Core inference versus other parts of the rendering pipeline, so they should not be interpreted as a direct measurement of DLSS 5’s neural workload alone.
They do, however, show that the feature materially changes how heavily the GPU is being exercised. NVIDIA describes the DLSS 5 model as a specialised real-time neural network running locally on GeForce RTX 50 Series Tensor Cores, with the technology able to operate independently alongside Super Resolution, Multi Frame Generation and Ray Reconstruction. Memory-controller load also rose sharply in relative terms, moving from 3.3 per cent to 8.8 per cent. The 165 per cent increase sounds dramatic, although the low absolute values provide more useful context. VRAM allocation remained completely unchanged at 34.3 per cent, and the memory clock stayed fixed at approximately 1,750 MHz.
That combination suggests that this particular workload is considerably more compute-intensive than capacity-intensive. DLSS 5 was moving additional data through the memory subsystem, but it was not placing enough pressure on GDDR7 bandwidth or VRAM allocation to force a different memory operating state. Core clocks told a similar story. The reported GPU clock increased from an average of 2,348 MHz with neural rendering disabled to roughly 2,789 MHz with it enabled, while effective clock rose from 2,380 MHz to 2,776 MHz. More importantly, the DLSS 5-on measurements displayed far less clock variation, indicating that the heavier workload was keeping the RTX 5090 in a sustained high-boost state rather than allowing clocks to rise and fall opportunistically.
Core voltage increased alongside it, from roughly 0.96 V to 1.045 V. Taken together with the power figures, the data indicates that DLSS 5 was not simply making use of some otherwise idle AI hardware without affecting the rest of the GPU’s operating behaviour. The card moved into a meaningfully higher performance and power state.
An extra 147 W of average GPU power naturally produced more heat. Average core temperature increased from 59.3°C to 67.2°C, while hot-spot temperature climbed from 64.6°C to 75.8°C. The highest hot-spot reading recorded during the DLSS 5 measurements was 79.7°C. That 11.2°C average increase at the hot spot is worth noting, but the absolute figures remained well within normal operating territory for the RTX 5090. There was no indication within this sample that DLSS 5 was forcing the card into thermal throttling or otherwise compromising sustained operation.
Memory temperatures moved much less. Memory junction temperature increased from 66.3°C to 69.7°C, while average memory-chip temperature went from 64.8°C to 67.3°C. That modest change is consistent with the unchanged memory frequency and VRAM utilisation seen elsewhere in the telemetry. The cooling response was more unusual. Despite the much higher power draw and temperatures, fan speeds barely changed. Fan 1 averaged 1,421.5 RPM with DLSS 5 disabled and 1,416.8 RPM with it enabled, while Fan 2 moved from 1,419.5 RPM to 1,417.5 RPM. Reported fan percentage remained essentially flat at 43 per cent.
The short measurement window makes it difficult to interpret this definitively. The GPU may simply have remained within a flat part of its fan curve, or the fan-control loop may take longer to react than our sampling period allowed. Sustained testing over several minutes would provide a much better picture of the eventual equilibrium between temperature, acoustics and fan speed.
The larger significance of these results is that DLSS 5 alters the familiar relationship between DLSS and GPU performance. Historically, the most visible DLSS technologies have existed largely to reduce rendering cost or increase perceived performance. Super Resolution lets the GPU render fewer native pixels, while Frame Generation and Multi Frame Generation increase displayed frame rates without requiring every frame to be conventionally rendered. 3D-Guided Neural Rendering deliberately spends GPU resources instead. The objective is not to make the same image cheaper, but to use additional AI compute to make the final image more sophisticated.
That does not mean overall performance necessarily has to fall. DLSS 5 sits alongside the rest of the DLSS suite, and NBA 2K27 can combine Neural Rendering with Super Resolution and Multi Frame Generation. NVIDIA’s own performance numbers put the RTX 5090 at up to 370 FPS at 4K with the Ultra preset and ray tracing when the broader DLSS stack is enabled. At 2560 × 1440, NVIDIA quotes up to 590 FPS on the RTX 5090, 410 FPS on the RTX 5080, 350 FPS on the RTX 5070 Ti and 260 FPS on the RTX 5070. These are NVIDIA-provided figures rather than our own benchmark results, but they illustrate how the different parts of the DLSS stack can work in opposite directions. Some recover performance by reducing the amount of conventional rendering required, while Neural Rendering can spend part of that available compute budget on image quality. That could become an increasingly important way of thinking about GPU performance. The question may no longer be only how many conventionally rendered pixels a graphics card can produce, but how intelligently its available compute can be divided between traditional rendering and neural processing.
The first implementation of any substantial rendering technology rarely tells the entire story, and DLSS 5 is unlikely to be an exception. Our early RTX 5090 telemetry shows that the neural renderer carries a meaningful computational cost in its current form. An almost 89 per cent increase in average GPU power draw is difficult to overlook, even when the card being tested has enough headroom to accommodate it comfortably. Optimisation will therefore be worth watching closely. Future versions of the model may require less compute, developers may learn to apply neural rendering more selectively, and later GPU architectures could potentially execute this kind of inference more efficiently. DLSS 5’s developer controls already make selective application possible, which gives studios considerable scope to balance visual improvement against computational cost.
The wider significance lies in the rendering model NVIDIA is proposing. DLSS 5 can operate alongside rasterisation, ray tracing and path tracing rather than replacing any one of them, while developers retain control over which objects and scenes receive neural enhancement. The result is a hybrid pipeline in which traditional graphics techniques establish the scene and provide reliable geometric and lighting information, while neural processing becomes another stage responsible for aspects of the final presentation. DLSS originally demonstrated that AI could reconstruct pixels a GPU had not rendered natively. Frame Generation extended that idea to entire intermediate frames. With DLSS 5, NVIDIA is applying neural inference deeper into the appearance of the scene itself, particularly the way materials and light interact.
NBA 2K27 is therefore interesting for reasons that extend beyond basketball. It offers the first practical look at what happens when neural networks become part of the artistic rendering process rather than simply a tool for recovering performance. Our initial RTX 5090 measurements also make one thing clear: at least in these early days, that added intelligence comes with a very measurable compute, power and thermal cost.