AMD FSR Redstone in detail – milestone and real counterpoint to NVIDIA’s DLSS?

1 - Introduction, three looks back and one forward 2 - ML Radiance Caching in Detail 3 - ML Ray Regeneration in Detail 4 - ML Upscaling in Detail 5 - ML Frame Generation in Detail 6 - Aktiviation of FSR4 in game or in the Adrenaline drivers 7 - Benchmarks and Metrics 8 - Quality Comparison and Conclusion

Both AMD Ray Regeneration and NVIDIA’s DLSS Ray Reconstruction pursue the goal of reconstructing ray tracing images with just a few samples per pixel in such a way that they largely approximate significantly higher-quality reference images. Both replace classic heuristic denoisers and instead use data-driven models based on extensive training sets. The basic idea is semantic reconstruction, in which the model uses contextual information from the scene to deduce what the physically correct result should look like.

AMD uses a fully neural model for ray regeneration that processes channel-rich input data. This includes normals, depth, specular and diffuse components as well as radiance and optional visibility data. The model is designed to actively estimate and fill in missing image information. It thus fills gaps in the ray tracing path and generates an image that is less affected by noise or incomplete samples. The structure follows a reconstructive pipeline that clearly focuses on understanding the content of the image.

NVIDIA also uses a neural model with DLSS Ray Reconstruction, but it is more deeply embedded in the existing DLSS architecture. Initially, processing was based on the NRD pipeline, which mainly worked heuristically. However, with the introduction of the Ray Reconstruction Module, the focus shifted more towards machine learning. The architecture is based on tensor cores and also processes all relevant G-buffer data from the scene. The approach remains modular, as upscaling and denoising come from two different blocks that complement each other well.

Integration into the render path

AMD integrates Ray Regeneration as a cohesive part of the Redstone suite. This creates a consistent pipeline that uses upscaling, temporal processing, ray tracing reconstruction and later also radiance caching together. The data runs directly after the actual ray tracing into the regeneration block before upscaling and frame generation are executed. The Redstone architecture therefore provides for clearly sequential processing.

NVIDIA integrates ray reconstruction via the DLSS interface of the respective engine. The individual models are executed as part of the RTX pipeline, which has already been expanded over the years. As a result, RTX titles benefit from a stable and proven infrastructure basis. However, the modules continue to work separately from the upscaling or temporal reprocessing functions, as DLSS uses different internal function blocks for resolution, motion and noise reduction.

Image quality in practical application

AMD shows a significant improvement in temporal stability with Ray Regeneration. The model visibly reduces the flicker and grain effects typical of classic denoisers. The representation of material properties and reflection gradients appears more consistent, as the model has been strongly trained to preserve structures. The examples from Call of Duty Black Ops 7 illustrate that the gain is particularly noticeable in scenes with high contrast ranges.

NVIDIA’s DLSS Ray Reconstruction also produces high-quality results in supported titles. The clear edge and material separations, which are characterized by the training on path tracing data sets, are particularly noticeable. In titles such as Cyberpunk 2077, there is a pronounced precision in the restoration of small details. NVIDIA has a very stable temporal performance due to its long experience in ML inference.

Performance and hardware utilization

AMD uses the new AI blocks in RDNA 4 for ray regeneration, which means that inference remains within a dedicated hardware path that does not compete with classic shader calculations. The presentation shows that the performance overhead remains very low. According to AMD, ray regeneration should not cause any significant losses. NVIDIA uses Tensor cores, which have been heavily optimized over several generations. As a result, the inference load of Ray Reconstruction remains low and independent of the other shader tasks. This leads to high efficiency, especially in path tracing scenes.compatibility and distribution

AMD provides Ray Regeneration exclusively for Radeon RX 9000. Widespread use will therefore depend on how quickly games integrate the Redstone suite. A first use case currently exists with Call of Duty Black Ops 7. NVIDIA already has an established DLSS infrastructure with many titles that use DLSS 2, 3 and RTX path tracing. This means that Ray Reconstruction can be distributed more quickly, as the necessary pipelines are already implemented and only require a version update.

Summary of the differences

With ray regeneration, AMD is pursuing an integrated approach that is part of an end-to-end ML pipeline and is closely linked to upscaling and frame generation. The focus is on temporal stability and consistent reconstruction of complex surfaces.

NVIDIA relies on a modular system with a mature hardware base and a broad market presence. Ray Reconstruction delivers precise and detailed results and benefits from many years of optimization of tensor-based ML paths.