Upscaling has rapidly become one of the most competitive and strategically important technologies in modern graphics. With the cost of native rendering at high resolutions continuing to rise, particularly in ray-traced workloads, vendors are increasingly reliant on advanced reconstruction algorithms to deliver playable performance without compromising visual fidelity. NVIDIA has long positioned DLSS as a cornerstone of its GeForce ecosystem, Intel is iterating rapidly on XeSS, and AMD has pursued an open and hardware-agnostic approach with FidelityFX Super Resolution (FSR).
The forthcoming iteration, FSR 4, represents the most significant architectural leap for AMD’s upscaling solution since its inception. Unlike its predecessors, FSR 4 introduces AI-assisted inference, bringing AMD’s strategy closer to NVIDIA’s deep learning approach. According to reporting from Moore’s Law is Dead (MLID), AMD is now allocating additional resources to ensure FSR 4 support across RDNA 3 and RDNA 3.5 hardware, while definitively ruling out official compatibility with RDNA 2-class GPUs.
The shift marks a strategic narrowing of support in exchange for higher quality, deeper architectural integration, and positioning against competing solutions.
FSR 4 Development: Current Status
Following the release of the FidelityFX SDK 2.0 earlier this year, AMD officially introduced hooks for FSR 4. However, implementation across hardware has not yet reached parity with competing solutions. While FSR 2 is broadly supported across multiple GPU generations (including NVIDIA and Intel hardware), FSR 4 is taking a more constrained approach, relying on new architectural features introduced with RDNA 3.
According to MLID’s reporting, AMD has internally prioritised three development goals for FSR 4:
-
Integration into RDNA 3 discrete GPUs (Radeon RX 7000 series)
-
Deployment across RDNA 3.5 APUs and mobile solutions
-
Porting to Sony’s upcoming PlayStation 5 Pro, powered by a semi-custom RDNA 3.75 design
These goals represent the bulk of AMD’s active graphics roadmap, covering desktops, handhelds, notebooks, and consoles.
RDNA 2 Excluded from Support
One of the most significant revelations is AMD’s decision not to backport FSR 4 to RDNA 2 or earlier GPUs. Architecturally, this exclusion is not arbitrary. RDNA 3 introduced dedicated AI acceleration blocks, which AMD refers to as AI Matrix Cores. These blocks offload inference workloads, enabling neural networks to run with significantly lower overhead compared to general-purpose shader execution.
In contrast, RDNA 2 lacks such hardware. Running FSR 4 entirely on shaders would impose heavy performance penalties, undermining the objective of an upscaler designed to increase effective framerates. While hacked community versions of FSR 4 have been demonstrated in unstable or “janky” forms on RDNA 2, AMD has little incentive to invest resources into official support when the outcome would deliver a poor user experience and dilute brand perception.
Thus, FSR 4 will mark a clean architectural cutoff, unlike FSR 2, which was deliberately designed for maximum backward compatibility.
Technical Context: What Makes FSR 4 Different?
From Spatial to Temporal to AI-Driven
-
FSR 1 was a purely spatial upscaler, effectively applying image sharpening and scaling without temporal data. Performance overhead was minimal, but visual quality was limited.
-
FSR 2 introduced temporal accumulation, motion vectors, and advanced anti-aliasing. It significantly improved image reconstruction, bringing it closer to DLSS 2, but it remained entirely shader-based.
-
FSR 3 layered on frame generation, but the upscaler itself remained fundamentally the same as FSR 2. Frame interpolation was useful but inconsistent, often hampered by latency and artifacts.
-
FSR 4 represents the first step into AI-assisted reconstruction, using trained neural networks to improve detail recovery, stabilise temporal history, and reduce artefacts.
The Role of AI Accelerators
RDNA 3 integrates specialised AI Matrix Cores that can accelerate 8-bit and 16-bit matrix operations. These cores are designed to handle the type of tensor math required for neural inference. In the context of FSR 4, they allow:
-
Faster reconstruction passes at lower cost compared to general compute shaders.
-
More complex neural models which improve detail recovery in fine textures such as foliage, hair, and high-frequency surfaces.
-
Reduced motion instability, as AI weighting better preserves edges and reduces ghosting.
Without these accelerators, running the same models on RDNA 2-class hardware would significantly cut into performance, often to the point where enabling FSR 4 could result in lower framerates than native rendering.
Moore’s Law is Dead on FSR 4 Development
“But when it comes to FSR4 on RDNA3, it is coming. It hasn’t been cancelled. I was told that the SDK is just taking a ton of resources. And what I would infer is that the priority for AMD has been to improve FSR 4, prepare FSR 4 Redstone for deployment, and port FSR 4 to PS5 Pro. So they are putting that as a priority over getting it working on older graphics cards. Um, but the good news is that they are kind of finishing up some of the other work, it sounds like. And so they are putting more people on the RDNA 3 port team of FSR 4 right now. And it’s just a lot of work.”
“They’re basically having to rewrite the code for FSR 4 like and that takes time. And so they’re going to do it, though, because they want PS5 Pro to get it, which has basically RDNA 3.75 if you will. Um, and they definitely want their RDNA 3 cards. And again, I think most importantly, all of their Strix Halo laptops, their Hot Point mini PCs, their Strix and Phoenix laptops that are out there, they want all of those to eventually have FSR 4 support. And so, it’s going to come, but it’s probably going to come around quarter 1 or at the end of quarter 4. So, that’s what it is. It’s just taking a lot of work. They prioritized other stuff until now. And late quarter 4 or quarter 1 is when RDNA3 and RDNA 3.5 and PS5 Pro should get FSR 4. Now, the only thing I will say is this is there is a little bit of bad news on this one, and that is that I do not believe it’s coming to RDNA 2. Now, obviously, we’ve seen some hacked versions of FSR4 running on RDNA 2 in a janky way. So, it sounds like they could plausibly port it to RDNA 2 in a nerfed way if they wanted to. I was directly told that they’re not doing that, or at least no resources are going towards that right now. So, if it ever came to RDNA 2, it’d probably be a very long time from now and not perform remotely as well as you’d probably want it to. And so, you know, there you go.“
Competitive Context: NVIDIA DLSS and Intel XeSS
NVIDIA DLSS 3.5
NVIDIA’s DLSS has consistently set the pace in AI-driven upscaling. With DLSS 3.5, NVIDIA now integrates Ray Reconstruction—a neural model that replaces manually tuned denoisers with AI inference. This not only improves ray-traced workloads but also showcases how deeply NVIDIA has embedded AI acceleration into its rendering pipeline.
Key differences relative to AMD’s FSR 4:
-
DLSS relies exclusively on proprietary tensor cores and requires RTX hardware.
-
DLSS models are trained on NVIDIA’s in-house supercomputers, giving them an advantage in data quality and refinement.
-
FSR 4, while AI-assisted, retains a broader focus on cross-platform flexibility and open SDK integration.
In short, NVIDIA is pursuing maximum quality through proprietary hardware and closed ecosystems, while AMD is balancing openness with newer hardware dependencies.
Intel XeSS 1.3
Intel’s XeSS occupies a middle ground. It supports both XMX hardware (on Arc GPUs) and DP4a shader-based inference (on other GPUs). This allows XeSS to scale across architectures, though quality varies depending on the execution path.
Relative to FSR 4, XeSS demonstrates a more hardware-agnostic approach to AI, though performance suffers outside of Intel’s own GPUs. AMD, by contrast, is explicitly optimising for its own accelerators in RDNA 3 and beyond, reducing backwards compatibility in exchange for performance guarantees.
Implications for Consoles and Mobile
PlayStation 5 Pro
Sony’s upcoming PlayStation 5 Pro is reported to integrate a semi-custom “RDNA 3.75” design, effectively bridging RDNA 3 and RDNA 4. Ensuring FSR 4 support at launch is a strategic priority, as it allows Sony to leverage reconstruction as a first-class rendering feature across its games catalogue.
Handhelds and Laptops
With RDNA 3.5 APUs forming the backbone of AMD’s mobile roadmap, FSR 4 becomes a critical feature for handhelds such as ASUS’ ROG Ally, Strix Halo-powered notebooks, and compact mini-PCs. These devices face the sharpest trade-offs between power consumption and performance, making high-quality upscaling a fundamental requirement rather than an optional enhancement.
Outlook and Timeline
Based on current development trajectories, FSR 4 is expected to arrive on RDNA 3, RDNA 3.5, and PS5 Pro platforms in late Q4 2025 or early Q1 2026. The rollout will be staged, with SDK refinement preceding game-level adoption.
For RDNA 2 users, no official support is planned. While experimental implementations may continue to surface in the community, AMD is unlikely to allocate engineering resources toward legacy hardware at this stage.
Conclusion
FSR 4 represents AMD’s most ambitious step yet in upscaling technology. By integrating AI-assisted reconstruction and leaning into RDNA 3’s architectural strengths, AMD is seeking to close the quality gap with NVIDIA’s DLSS while still maintaining broader platform flexibility than Intel’s XeSS.
The trade-off is clear: higher performance and visual quality on newer hardware, at the expense of backward compatibility. For AMD, this aligns with its long-term roadmap, particularly as it seeks to consolidate FSR as a ubiquitous solution across desktops, laptops, handhelds, and consoles.
As of now, AMD’s allocation of resources signals a decisive move: FSR 4 is coming to RDNA 3, RDNA 3.5, and PS5 Pro — but RDNA 2 is being left behind.
Source – Moore’s Law is Dead, YouTube
Leave a Reply Cancel reply