AMD FSR Redstone: The big AI throwdown or just hot air? An outlook on the biggest FSR update to date

With the official announcement of the “FSR Redstone” release for December 10, 2025, AMD is making it clear that it is no longer just about classic upscaling, but about the systematic integration of machine learning into all relevant render paths of the Radeon graphics pipeline. The goal is clear: the Radeon RX 9000 series based on RDNA 4 is to become the spearhead of a new AI-accelerated graphics era, visually higher quality, better performance and at the same time more open than the proprietary competition. However, as is so often the case with major technological announcements, caution is advised. Experience has shown that there is a world of difference between promises, reality and actual market penetration.

So what exactly is Redstone? According to AMD, it is the biggest update of the entire FSR software stack to date. Instead of incremental improvements, as was the case with FSR 2 and FSR 3 with relative regularity, there is now talk of a profound leap: Redstone integrates four machine-learning core technologies into the render architecture. These include Neural Radiance Caching (an AI-supported form of global illumination via light behavior modeling), ML Ray Regeneration (a neural replacement for classic ray tracing denoisers), ML Super Resolution (upscaling via AI training instead of heuristic filters) and ML Frame Generation, i.e. the insertion of synthetic intermediate frames to increase the frame rate. All of these modules should not function in isolation, but as an interlocking pipeline to increase performance, reduce power consumption and improve image quality.

From a technological point of view, this step makes sense and, you have to hand it to AMD, is the right one in terms of content. The rendering stack of classic GPUs is increasingly reaching its limits, especially in the context of ray tracing and highly dynamic worlds. Classic methods for denoization, lighting and interpolation either require immense computing power or deliver visually unsatisfactory results. With DLSS 3.5 and Ray Reconstruction, NVIDIA has already demonstrated how these bottlenecks can be alleviated using specialized ML paths. AMD is now following suit, in its own way, but with comparable goals. What is particularly noteworthy, however, is that while NVIDIA is tying its features to its own hardware (Tensor Cores, RTX library), AMD is signaling openness, at least in part. Initial indications suggest that Redstone could theoretically also run on GPUs from other providers, i.e. on current NVIDIA or Intel models. If this actually happens, it would be a strategic exclamation mark: AMD would position itself as a proponent of an open, platform-independent AI standard, which NVIDIA is precisely not doing with its strictly proprietary DLSS policy.

But, as always, there is a catch. Officially, FSR Redstone will initially only be offered for the RX 9000 series. Although AMD has not yet made any definitive statements regarding support for older GPUs such as RDNA 3 or 3.5, it is obvious that they want to use the launch to showcase the new hardware generation exclusively. This has a clear market logic, but is also a disappointment for owners of current GPUs who are also hoping for more performance and better image quality. What’s more, AI features such as ray regeneration or frame generation are not trivial. They require carefully prepared data, functioning render pipelines and engines that support these functions. Redstone may be ready at the driver and API level, but integration into real games is another matter. The first demo was used in Call of Duty: Black Ops 7, albeit under controlled conditions and heavily optimized code. Whether this will also work for more complex open-world titles or poorly optimized games remains to be seen.

It becomes even more critical when you consider that many FSR Redstone features are in competition with existing mechanics and could potentially lead to inconsistencies. Frame generation, for example, is a double-edged sword in principle: the generated frames are “fictitious” by definition and can mask input latencies or create artifacts, depending on the game genre and frame rate. Upscaling via ML is not a panacea either: if the source image provides too little information (e.g. with heavily compressed or reduced texture quality), even the best model cannot work miracles. And finally, the question arises: How stable and driver-side mature will Redstone really be from day 1? The past has shown that AMD has often had to contend with delays, incompatibilities and functional gaps during software rollouts, whether FidelityFX Super Resolution 1.0 or the first implementations of FSR 2.

The bottom line is an ambivalent picture. AMD FSR Redstone is without question a major technological step and possibly the beginning of a new era for Radeon users. The vision of anchoring machine learning not as a gimmick, but as a firm foundation of modern image generation, is bold and long-term. However, success will only come if AMD manages to make Redstone consistent, open, stable and available in as many games as possible, and not just on a single GPU generation. Until then, skepticism remains appropriate. Because as the saying goes: you can announce a lot of things. The decisive factor is what you actually see on the screen in the end.

Source: JackMHuynh via X