Radiance Caching
AMD's FSR Radiance Caching is a new technology aimed at reducing the performance cost of ray-traced or path-traced lighting and global illumination in games. Breaking the term down first: radiance refers to the amount and color of light leaving a point on a surface in a given direction, while caching loosely means reusing information instead of recalculating it every frame.
The practical goal is to reduce the number of rays that need to be traced by using a learned approximation of the lighting environment, allowing paths to terminate early while still capturing indirect illumination at an acceptable quality level.
Earlier descriptions of FSR Radiance Caching were ambiguous, particularly around how much of the system is actually "neural" and when learning occurs. At events like Computex, AMD referred to the feature as Neural Radiance Caching and emphasized that it continually learns. Later reviewer materials softened that language and focused instead on "learning" and "intelligent prediction", while retaining the name FSR Radiance Caching.
I looked up the now-live SDK developer documentation to verify what is really happening. This page has been reworded accordingly. Radiance Caching is implemented as a fully online machine learning system that trains continuously at runtime. There is no offline or pre-shipping training step, no baked data bundled with the game, and no one-time compilation or learning phase at first launch. Instead, the neural model starts from default parameters and is updated every frame using data generated by the game's path tracer.
Training data is sourced directly from camera paths traced into the scene. As rays intersect polygons, samples are emitted that capture position, normal, view direction, albedo and roughness, paired with an estimate of the radiance leaving the surface along that ray. These samples are used to update the cache while, in parallel, the same model is queried to provide radiance predictions used to terminate other paths early.
This setup places Radiance Caching clearly in the category of runtime lighting algorithms, not precomputed lighting techniques like baked lightmaps. The cache itself is not a lookup table or set of stored probes, but the learned weights of a neural network that approximates the radiance function in the current scene.
At runtime, Radiance Caching isn't used right away when a ray first hits something. Instead, rays are allowed to bounce around the scene until they've spread out enough that fine detail is no longer critical. At that point, the renderer asks the cache for a lighting estimate and stops tracing that ray any further. This "final gather" style approach helps hide errors from the cache and preserves hard-to-replace details like small geometry features, self-shadowing, and reflections. Rays that hit mirror-like surfaces are traced further, since they're harder to approximate reliably.
When the cache returns a lighting result, it isn't used on its own. The prediction is scaled by the same weight the path would normally have accumulated and then blended back into the path tracer's output. In practice, the neural cache replaces the most expensive deeper bounces, while keeping the math consistent with traditional ray tracing.
The documentation also makes it clear that this isn't a free win. Because the system trains in real time using noisy ray tracing samples, it can suffer from flickering or instability, especially in difficult lighting situations. To control this, developers can adjust how fast the model learns and how aggressively its predictions are smoothed over time. AMD also recommends feeding the cache better data in the first place, using techniques like smarter light sampling (ReSTIR), path guiding, or slightly roughening training rays to reduce noise.
This also explains why Radiance Caching is not supported on older RDNA generations (this was unclear even in AMD's own reviewer's guide). If it were pre-trained offline, shipped with the game and inference-only, then older GPUs should be able to handle this just fine, but the SDK makes it clear that training is a core part of the workload, executed every frame. Efficient matrix math, bandwidth, and support for high-performance machine learning backends are fundamental to the feature's design, it probably uses INT8 as well.
Radiance Caching is now available to game developers as a technical preview, with early implementations expected in 2026. AMD has confirmed that Warhammer: Darktide will be among the first titles to ship with the feature.
NVIDIA employs a range of techniques that are also described as radiance caching in the broader sense, but AMD's implementation differs in emphasis. Rather than storing radiance samples in probes, grids, or reservoirs, FSR Radiance Caching uses a continuously trained neural model to predict radiance directly. In that sense, the "cache" is not a conventional data structure, but the learned behavior of the network itself.