DLSS 5 Takes a Fixed Slice of Every Frame No Setting Can Shrink

Independent on-versus-off testing puts a number on DLSS 5 Neural Rendering, NVIDIA commits to Ada support later, and AMD and Intel both ship game-support drivers.

Saya4 minute readDrivers & Software

DLSS 5 went live in NBA 2K27 on September 4, and the first independent numbers cover the measurement NVIDIA's own charts left out: what Neural Rendering costs with frame generation switched off. It is a fixed block of GPU time attached to output resolution, and almost nothing in the graphics menu moves it. NVIDIA has also changed its answer on RTX 40.

The neural pass costs a fixed slice of every frame

An RTX 5090 at 4K with DLAA and maximum settings runs the game at 76 FPS with Neural Rendering on against 192 without, a 60 percent drop, and dropping the preset to Low buys back 9 FPS, about 12 percent. That is what TechSpot's on-versus-off testing means by a fixed cost: the model post-processes the finished frame, so shrinking the rest of the render budget does not shorten the part that dominates.

An RTX 5080 shows the same pattern one tier down, dropping from 221 FPS to 102 at 1440p, a 54 percent cut, and from 185 to 53 at 4K. Upscaling does not rescue any of it, because the pass runs after Super Resolution rather than before it. An RTX 5060 Ti 16GB at 1440p falls from 129 FPS with DLAA to 46 with Neural Rendering, and switching to DLSS Quality recovers 4 FPS. Without the feature the same change is worth 28 percent. The measured pass times explain the shape of the whole stack: about 3.6 ms per frame on an RTX 5090 at 1440p rising to roughly 24 ms on an RTX 5050, and at 4K from about 7.9 ms to 53 ms. Anything above 16.7 ms cannot reach 60 FPS no matter how fast the rest of the frame renders, which rules out 4K on every RTX 50 card except the 5090 and, on a knife edge, the 5080. NVIDIA says optimization work continues, so these figures describe the launch build rather than a permanent ceiling.

What the model buys, and what it charges

The image holds up better than the pre-launch argument suggested. Hours of unguided play produced no warping faces, no corrupted UI or text, and consistent characters under fast camera motion, with the gains landing on tonemapping, contact shadows, subsurface scattering in skin, and light through hair and foliage, per HotHardware's image-quality pass. This is a post-process filter rather than a rendering technique: no extra light simulation happens, the model paints in what it has learned detail looks like, and roughly 70 percent of the subjective effect is reachable with ReShade shaders at a fraction of the cost. The remaining 30 percent is what conventional shaders cannot fake.

The trade lands in two places. Because the cost tracks output resolution, keeping the feature on usually means running the display below native, which throws away the fine detail the model was added to produce; distant jersey numbers and court lines are the casualty at 1440p on a 4K panel. A desktop RTX 5070 Ti also sat at its 300W limit for the entire test run, a FurMark-class load, because raster, RT and tensor units are all busy at once. ComputerBase's test reaches the same verdict, calling the effect on image quality, performance and power draw large in every direction.

Ada gets official support once Blackwell is tuned

NVIDIA now says DLSS 5 will officially reach GeForce RTX 40, after several earlier statements that the feature was RTX 50 only. The order of work is Blackwell first: the company claims a roughly fivefold speedup since March, expects further optimized models in the fall, and will start on Ada once RTX 50 performance is close to fully tuned, as ComputerBase reports the commitment. No date attaches to any of that, and no generation below Ada was mentioned.

The technical distance is short, which is why this reads as credible. DLSS 5 runs on FP8 rather than Blackwell's FP4, and Ada has native FP8, so the modders who got the leaked runtime working on an RTX 4090 only had to swap CUDA binaries compiled for the wrong architecture. AMD went through this with FSR 4 on RDNA 3, where an unofficial port preceded the version AMD had to build properly.

AMD and Intel both ship game-support drivers

Adrenalin 26.9.1 is an optional release adding support for Onimusha: Way of the Sword and The Blood of Dawnwalker, and its one stated fix is intermittent crashing in Starfield and Apex Legends on RX 9000 cards, per Guru3D's release listing. Two known issues are worth reading first: Smart Access Memory can end up disabled on Windows 10 after the update, and AI Bundle components fail to install where GitHub or Hugging Face access is restricted.

Intel's 32.0.101.8992 covers the same two titles and fixes graphics corruption in Marathon under DirectX 12 with anisotropic filtering, flickering in the DirectX 9 build of Assassin's Creed Revelations, and full-screen corruption in Chapter 8 of 007 First Light, across Arc A and B series and Core Ultra integrated graphics, per Guru3D's driver page. It ships as a generic package, so installing it on an OEM machine can replace a manufacturer-customized driver and the display or power behavior that came with it.

For anyone deciding whether to leave Neural Rendering on, the number to check is the pass time at the monitor's native output resolution, because that is the only variable that moves the answer.

Sources