AMD Neural Lighting versus DLSS 5: Can RDNA 5 finally catch up with NVIDIA?

AMD would be preparing its own response to NVIDIA's neural rendering. The name that is circulating is Neural Lighting , a technology associated with future RDNA 5 cards and capable of reconstructing part of the lighting with neural networks. Said like that, it almost sounds like a product announcement. It is not one.

The primary source is a message from Kepler_L2 published on the AnandTech forum on September 15 and 16, 2026. Its formula is short: AMD would have its version of "Neural Lighting" from DLSS 5 on the next generation. It adds that the calculation would be "pretty heavy," so probably too heavy for RDNA 4. TweakTown, Club386, VideoCardz, and TechPowerUp then picked up this discussion. AMD has confirmed neither this name, nor an RDNA 5 exclusivity, nor even the product.

There is still some solid evidence behind the rumor. AMD published two papers in 2026 dedicated to indirect neural lighting. This is the right way to read this leak: a credible hint placed on a real research track, but not a technical sheet hidden in a drawer.

In the vocabulary used by Kepler_L2, neural rendering encompasses shader operations replaced or assisted by a neural network. Neural Lighting would more precisely target indirect light, shadows, and complex interactions between materials. We are therefore talking about something other than a simple FSR that enlarges a rendered image in low definition.

• The name Neural Lighting comes from the leaker, not from a trademark or an AMD presentation.
• The association with RDNA 5 is based on the assumed computing cost and on rumors of more powerful AI hardware.
• The idea of a deep integration into the engine comes partly from Kepler_L2's personal analysis. He does not present it as a confirmed AMD specification.
• The 2027 or 2028 calendar remains unofficial. AMD has not announced a Radeon "RX 10000" nor a launch date.

This last nuance matters a lot. Several articles have transformed the technical preference of the leaker, with access to geometry, BVH, materials, and depth buffers, into an already decided AMD architecture. We know nothing about it. The concept is relevant. Its integration into a driver, FSR, or a future developer kit remains unknown.

The best argument in favor of the rumor does not come from another leaker. It comes from GPUOpen. At SIGGRAPH 2026, AMD researchers presented a lightweight model for indirect lighting reconstruction. It combines camera-visible buffers with reflective shadow maps rendered from light sources. The result: the network also receives information about off-screen surfaces, those that a purely screen-space method often misses.

Click to enlarge

The prototype has 2.19 million parameters and requires 432 GFLOPs. In the published test, it runs in 45.56 ms at 512 × 512 on an Instinct MI250 accelerator, in FP32 and without inference optimization. In other words: it is a proof of research, not yet a real-time gaming feature ready to check in Adrenalin . At 60 frames per second, the entire frame must fit within 16.7 ms. The model alone far exceeds this budget in this configuration.

The fact remains that the method exactly addresses the problem mentioned in the leak: to reconstruct a plausible indirect light without launching enough rays to melt the framerate. The intellectual bridge is obvious. The product bridge, however, remains to be built.

AMD has also released for ECCV 2026 a "generative illumination" model with latent diffusion in a single step. It starts from direct lighting, normals, albedo, roughness, materials, and some light cues. A temporal decoder reuses the previous image with the motion vectors to limit flickering.

Click to enlarge

It is clearly more structured than an image generator running free. The geometry and properties of materials serve as safeguards. AMD also acknowledges that classic diffusion models can hallucinate or overexpose the scene, then shows how its control inputs reduce these flaws.

But again, no shortcut. This paper does not mention RDNA 5 and does not name anything Neural Lighting. It shows that AMD has been working on the subject for several years. It does not prove that a consumer Radeon will execute this model as is.

NVIDIA has already crossed the stage of commercial demonstration. DLSS 5 takes the color and motion vectors of each frame, then applies a model trained to understand faces, hair, fabrics, translucent skin, and lighting conditions. Developers have intensity, colorimetry, and mask settings. NVIDIA refers to "3D-Guided Neural Rendering" because the result must remain tied to the 3D content and stable from one image to the next.

Click to enlarge

The change is significant: DLSS is no longer just a performance tool. It retouches lighting and material rendering. This is precisely what divides. When the result respects the studio's intention, the image can achieve stunning realism. When the model imposes its idea of "natural," a face changes, the skin becomes waxy, or a deliberately stylized atmosphere gets smoothed out. That's where it gets stuck.

Kepler_L2 proposes to separate the lighting model from the model responsible for complex surfaces and to eliminate the color grading step. This is an interesting technical critique, not proof that AMD has already chosen this division.

In terms of ideas, yes. AMD is not starting from scratch and its GPUOpen work is serious. Providing the model with scene information rather than asking it to guess each surface from the final image can reduce hallucinations. The separation between light and materials can also give artists more control. On paper, this is even the approach that many players would prefer.

AMD's problem has never been solely the algorithm. It requires hardware capable of running the network without slaughtering the FPS, an easy-to-integrate SDK, tested profiles on dozens of games, and strong relationships with studios. NVIDIA has been pushing Streamline and DLSS since 2018. Eight years of tools, documentation, and field feedback do not disappear just because a research paper is clever.

And the calendar stings. DLSS 5 is already here in 2026. If RDNA 5 arrives at the end of 2027 or in 2028, AMD will have to aim for the version of NVIDIA available at that time, not the one from today. Catching up to a moving target remains the real challenge.

Running DLSS 5 on a current Radeon with mods gives an idea of the computing wall, even if this hack obviously does not measure an optimized AMD implementation. On an RX 9070 XT, reported tests fall below 20 FPS in certain scenarios. This is not proof that RDNA 4 is incapable of neural lighting. It is proof that you cannot carry a heavy NVIDIA model on another architecture without paying a hefty price.

RDNA 5 is precisely expected with more matrix capabilities and dedicated hardware for more ambitious ray tracing. These elements also remain in the rumor column. AMD could offer a full version on the new generation and a reduced mode on RDNA 4, just as NVIDIA extends certain functions to the RTX 40. Or cut it off completely. We do not have the answer.

For the buyer, the conclusion is simple: do not buy a 2026 card imagining that it will receive Neural Lighting later . No commitment guarantees it. A RX 9070 XT must be chosen for its performance, its VRAM, and the features it has today.

The rumors surrounding a unified UDNA architecture place the next generation AMD at the heart of future Radeon but also consoles. Sony is already working with AMD on Project Amethyst, with a clear focus on machine learning and ray tracing. Microsoft is also communicating about neural rendering for its next hardware.

If a common base equips PC, PlayStation and Xbox, studios will have a good reason to integrate compatible buffers and rendering paths. This is probably AMD's best card. Not because consoles guarantee technical success, but because they can turn a minority Radeon option into a cross-platform feature planned from the engine's design.

Be careful though: neither Sony nor Microsoft has confirmed Neural Lighting. Mixing their official work with the name that came from a forum would make a nice story, but bad information.

AMD owns the research bricks. NVIDIA already owns the product, the games, and the ecosystem. This is the balance of power, without smoke and mirrors. Neural Lighting could become a more data-respectful response to 3D and artistic direction. For now, it is mainly an unofficial name attached to two excellent scientific papers.

RDNA 5 can catch up with NVIDIA on a specific feature. Catching up the entire DLSS stack in 2027 or 2028 will be much harder. AMD will need to deliver good performance from the first driver, several solid games, and simple integration. A lab demo at 512 × 512 will not be enough.

The leak deserves to be followed. Not to be taken at face value.

No. The name and its association with RDNA 5 come from Kepler_L2 on the AnandTech forum. AMD has not released any product statement regarding this.

We do not know. The technology could join a future version of FSR, another SDK, or remain a research project. The name FSR 5 is not confirmed in this context.

DLSS 5 is an official NVIDIA product that enhances lighting, materials, and characters based on the rendered frame and temporal data. Neural Lighting would aim more at global illumination and could leverage deeper 3D data, but this architecture remains hypothetical.

Nothing indicates it. Kepler_L2 thinks that the load will be heavy enough to reserve the function for RDNA 5. AMD has not provided any compatibility list.

Leaks talk about 2027, with a risk of slipping to 2028. AMD has announced neither date, nor commercial name, nor corresponding Radeon range.

The primary discussion is on the forum AnandTech . The two AMD works are available on GPUOpen : Lightweight attention-based indirect illumination and Temporally stable generative illumination . The official description of DLSS 5 has just been NVIDIA . The leak has been cross-referenced with TweakTown and Club386.