NVIDIA DLSS 5 October Update Trades Shadow Detail for Speed: Preset A vs Preset B Explained
We have followed DLSS 5 from its first leaked DLL in August to its official launch, and the October update is the most consequential release yet. NVIDIA has shipped a new, lighter neural model that runs dramatically faster than the launch version, and the community quickly discovered that the extra speed came with visible tradeoffs.
The story broke in the PC hardware community on November 4, 2026, when GameGPU reported that enthusiasts comparing the old and new models side by side in Like a Dragon: Infinite Wealth found a clear reduction in generated shadows and reflections. Below we lay out what changed, the numbers the community measured, why some players welcome the tradeoff, and how you can pick the model you prefer.
A note on method: we have not run these DLL builds ourselves. Every figure below comes from published community tests and outlet reports, attributed as we go. Do not treat this as lab testing of our own.
What DLSS 5 actually is
DLSS 5 is the marketing name for NVIDIA’s 3D-Guided Neural Rendering. Unlike classic DLSS upscaling, which reconstructs a higher-resolution image from a lower-resolution one, DLSS 5 takes the game’s finished frame plus its motion vectors and has a neural model re-light the scene: skin gets subsurface scattering, fabrics get sheen, hair behaves like hair. NVIDIA announced it at GTC 2026 on March 16, and it officially shipped on September 3, 2026 in exactly one game: NBA 2K27, on PC, for GeForce RTX 50-series cards.
Before the official launch, hardware enthusiast Renan Maniero found an unfinished pre-release build of the library, nvngx_dlssnr.dll version 310.8.0.0, inside NBA 2K27’s early-access PC files on August 26, 2026. That 158 MB build, carrying roughly 148 million FP8 parameters, is what almost every early DLSS 5 test used. Community testing showed it roughly halved frame rates wherever it landed, which made the official October update such a surprise.
What the October update changed
NVIDIA shipped the updated model inside an NBA 2K27 patch on October 16, 2026. Maniero extracted the new build, version 310.9.16.0, the same day, and the community immediately began A/B testing. The headline changes, as reported by Tech Times and GameGPU, are:
- Model size cut by 24 percent, down to roughly 118 million parameters.
- 4K video memory use reduced by 20 percent, a real relief for cards at the edge of their VRAM budget.
- Neural pass roughly twice as fast, which is the change players feel most directly.
- Both models ship in one DLL. The library contains Preset A (the September model) and Preset B (the October model), so testers can switch between them without swapping files.
Community benchmarkers measured large gains. Tester GamingOptimized reported The Witcher 3 running up to 90 percent faster on the October preset than the September one. Community tester Detro measured Cyberpunk 2077 on an RTX 4080 SUPER with the neural pass cost dropping from 8.8 milliseconds to 5.4, and frame rates with 4x frame generation climbing from 108 to 150 FPS, as covered by The Daily Checkpoint. In a natively integrated scenario, PC Gamer tested NBA 2K27 itself on a laptop with NVIDIA’s RTX Spark chip and found roughly 25 percent better performance at 1440p, with 1080p average frame rates moving from 143 to 168 FPS.
Preset A vs Preset B, side by side
| Aspect | Preset A (September model, 310.8) | Preset B (October model, 310.9.16.0) |
|---|---|---|
| Model size | About 148 million FP8 parameters | About 118 million parameters |
| 4K VRAM use | Higher | About 20 percent lower |
| Neural pass speed | Baseline | About twice as fast |
| Generated shadows and reflections | Denser contact shadows, soft character reflections | Reduced; fewer generated contact shadows |
| Image character | Can add gloss to matte surfaces, halo glow around characters | Softer gradients, more natural in dark scenes, less haloing |
| Typical performance gain vs Preset A | Baseline | 10 to 40 percent depending on game and GPU (community reports) |
Why players are upset
The controversy centers on what the lighter model no longer draws. Using Like a Dragon: Infinite Wealth as the test scene, enthusiasts found that the September model automatically rendered dense contact shadows and soft reflections of character models on the road surface, creating a convincing imitation of full path tracing. After the October update, GameGPU reports, those generated effects virtually disappeared, and objects and characters in the frame once again look detached from the surrounding geometry.
For players who bought into DLSS 5 for its re-lighting promise, this feels like a quiet downgrade sold as an optimization. The fear is that the “faster” model simply does less of what made DLSS 5 interesting in the first place.
Why some players defend the change
The other side of the debate has a point of its own. As GameGPU notes, the old model did not have accurate material roughness data to work with, and it often guessed wrong: it added an unnatural gloss to matte surfaces and wrapped characters in an artificial studio glow. Tester Detro found that Preset B produces softer color gradients and a gentler lighting pass that looks more natural in dark scenes, reduces haloing around character silhouettes, and smooths specular reflections on metallic surfaces without washing out environmental detail.
There is also the practical math. The September model roughly halved frame rates in injected testing; a model that costs half as much per frame and still re-lights the scene is arguably the usable version of the technology, while the heavyweight model was a tech demo that happened to ship. GameGPU’s own earlier testing found Preset B pushing Shadow of the Tomb Raider to a stable 120 FPS on portable hardware, a scenario where Preset A simply could not run.
How to switch between the presets yourself
Despite Preset B becoming the default, NVIDIA kept both models in the DLL, and the community tooling already exposes the switch. Practical options:
- DLSS AiO Swapper lets you swap the library and choose the preset directly.
- OptiScaler and RenoDX can inject DLSS neural rendering into games without native support and expose the A/B toggle.
- The Lecram 10.8 addon adds preset switching for its users.
- None of this requires rolling back your graphics driver.
Two warnings before you experiment. First, do not swap DLLs in multiplayer titles: as Perexp Teamworks points out, modifying game files in titles protected by automated anti-cheat, such as Call of Duty or Fortnite, can trigger instant account penalties. Second, enthusiasts report that running the neural pass twice with Preset B at 85 or 90 percent model resolution scale restores much of the shadow depth with fewer halos, at performance similar to a single pass of the old model, but that this mode noticeably raises power consumption on flagship cards at the RTX 5080 level and above. That last tradeoff is worth knowing about on high-end builds, and our PC settings guide for maximum FPS covers how to balance settings against power and thermals.
What this means going forward
NVIDIA has made Preset B the mainline model for future releases, which means the controversy is not going away. The next big test is Hogwarts Legacy, which is scheduled to get official DLSS 5 integration; our Hogwarts Legacy DLSS 5 update guide tracks that rollout. If the lighter model is what ships in every new game, the Preset A look becomes something you have to opt into with community tools rather than something games give you by default.
The broader context helps too. NVIDIA is expanding the DLSS ecosystem rapidly: the company just added DLSS 4.5 and Reflex support to a new batch of games, which we covered in our DLSS 4.5 and Reflex games roundup, and War Thunder is getting DLSS 5 neural rendering mid-November, as we reported in our War Thunder DLSS 5 coverage. The October model is the foundation all of that will be built on.
Our take
This is a genuine tradeoff, not a scandal and not a triumph. Preset B gives up some of the September model’s most impressive re-lighting tricks and, in return, becomes fast enough to actually use, with 10 to 40 percent better performance and 20 percent lower 4K VRAM use in community testing. Players who judge with their eyes in slow, dark scenes will miss Preset A. Players who judge with their frame counter will take Preset B without hesitation. The good news is that NVIDIA did not delete the old model: as long as both presets ship in the DLL, the choice stays with you.
