AMD Rumored to Be Building Neural Lighting for RDNA 5 as DLSS 5 Rival

AMD is reportedly developing a more advanced neural rendering technology for its next generation Radeon hardware that could compete directly with NVIDIA DLSS 5. The rumored feature is currently being referred to as Neural Lighting and may require RDNA 5 class hardware because of the computational load involved.

The information comes from hardware leaker Kepler_L2, who claimed during an AnandTech forum discussion that AMD is working on its own neural lighting implementation for future GPUs. When asked whether the technology would be exclusive to RDNA 5, the leaker suggested that it probably would because of how demanding the workload is. AMD has not announced Neural Lighting or formally confirmed RDNA 5 as the name of its next gaming architecture, so both points remain unconfirmed.

The idea would push AMD beyond the neural rendering technologies it already offers through FSR Redstone. AMD currently provides machine learning based Upscaling, Frame Generation, Ray Regeneration, and Radiance Caching. Ray Regeneration reconstructs cleaner ray traced information from noisy inputs, while Radiance Caching trains a neural model in real time to predict how indirect light moves through a scene and reduce the number of expensive path traced calculations required. AMD GPUOpen describes these technologies as part of a much broader transition toward neural graphics.

Neural Lighting could potentially move that strategy closer to what NVIDIA introduced with DLSS 5. NVIDIA’s latest neural rendering system goes beyond conventional upscaling and denoising by using neural processing to alter lighting, materials, skin, hair, and other visual properties while remaining guided by information from the underlying 3D scene.

AMD has already signaled that its research is heading further in this direction. GPUOpen now features work on generative global illumination, neural indirect lighting, neural texture compression, and other machine learning graphics techniques. AMD itself says its neural rendering roadmap extends beyond the currently released FSR Redstone features, although it has not detailed specific future FSR products.

One part of the leak needs additional caution. Kepler_L2 also discussed how an AMD implementation could potentially differ from NVIDIA by giving neural models deeper access to raw game data such as geometry and depth information and by separating lighting and shading models. Those comments were presented as ideas for how such a system could be designed, not confirmed information about AMD’s actual implementation.

The potential hardware requirement is also believable from a performance standpoint. FSR Ray Regeneration and Radiance Caching currently target RDNA 4 hardware, while AMD has already expanded less demanding FSR Upscaling features back to RDNA 3. Neural rendering workloads can scale far beyond conventional reconstruction, making additional matrix compute and architectural improvements increasingly important.

We recently saw just how demanding this can become when a community developer managed to get DLSS 5 Neural Rendering running on the Radeon RX 9070 XT. The experiment worked, but even after optimization the RX 9070 XT managed only around 33 FPS at 1080p in Cyberpunk 2077 with Neural Rendering active.

AMD developing something beyond FSR Redstone would make sense. The company already has the individual building blocks for machine learning based upscaling, frame generation, ray reconstruction, and global illumination.

A future Neural Lighting system could combine those ideas into something much more ambitious, but there is an important difference between AMD researching neural graphics and AMD having a finished DLSS 5 competitor ready for RDNA 5.

For now, Neural Lighting remains a rumor. What is confirmed is that AMD is steadily moving more of its graphics pipeline toward machine learning, making a deeper neural renderer increasingly plausible.

Would you rather see AMD build a full DLSS 5 style neural renderer, or focus first on improving ray tracing performance and FSR across existing Radeon GPUs?