AMD is bringing FSR 4.1 compatibility to older Radeon hardware, with an expected July release for RDNA 3 (RX 7000) cards and a 2027 rollout for RDNA 2 (RX 6000) series. Software leads Andrej Zdravkovic and Terry Makedon confirmed the company is re-engineering AI models to work across different mathematical formats, such as FP8 and INT8, to ensure performance across various GPU architectures.
How will FSR 4.1 function on different Radeon architectures?
The technical implementation of FSR 4.1 varies depending on the generation of the graphics card. While the upcoming Radeon RX 9000 series, based on the RDNA 4 architecture, utilizes FP8 (8-bit floating point) for AI processing, older cards lack this specific hardware capability.
To bridge this gap, AMD is adapting the software model for the RDNA 3 architecture found in RX 7000 series cards. According to Andrej Zdravkovic, AMD software manager, and Terry Makedon, senior software director, these cards will instead use INT8 (8-bit integer) operations. This conversion allows the upscaling technology to function on hardware not originally designed for FP8 math.
AMD stated that the visual quality should remain consistent across these different mathematical approaches. Users with RX 7000 series hardware should not see a degradation in image quality compared to the RDNA 4 implementation, despite the underlying model being technically different.
Why is the RDNA 2 support delayed until 2027?
The timeline for RDNA 2 (RX 6000 series) support is significantly longer due to the absence of dedicated AI accelerators in those older chips. Unlike RDNA 3 and RDNA 4, the RX 6000 series must rely on standard Stream Processors to handle the AI workload required by FSR 4.1.
This creates a resource competition. Because the upscaling algorithm must share the GPU’s primary calculation units with the actual game rendering, AMD must optimize the code to prevent the upscaling itself from slowing down the game. AMD’s engineers are working to ensure the algorithm remains efficient enough that the performance gain from upscaling outweighs the cost of running the AI model on general-purpose shaders.
| Architecture | Series Example | AI Math Method | Expected Release |
|---|---|---|---|
| RDNA 4 | RX 9000 | FP8 | Initial Launch |
| RDNA 3 | RX 7000 | INT8 | July |
| RDNA 2 | RX 6000 | Stream Processors | 2027 |
What is the AMD model training pipeline?
AMD uses a tiered approach to develop and refine these AI-driven upscaling models. The process begins with high-level training on Instinct MI accelerators, which are professional-grade solutions designed specifically for large-scale AI and data center computations.
Once the initial model is trained, AMD moves the development to Radeon Pro hardware. This stage focuses on refining the models for real-world graphical environments. The company utilizes the ROCm platform to unify development across its various hardware tiers, from professional Instinct accelerators to consumer Radeon cards.
Before a wide release, AMD performs testing across a massive variety of PC configurations. This includes testing how the algorithm interacts with different CPUs, motherboards, memory capacities, power supplies, and driver versions to ensure stability in diverse user environments.
Frequently Asked Questions
When can RX 7000 users expect FSR 4.1?
AMD is targeting a July release for the RDNA 3-based Radeon RX 7000 series.
Will FSR 4.1 work on my RX 6000 series card?
Yes, but AMD has indicated that support for RDNA 2 cards is not expected until 2027.
Does FSR 4.1 require AI-specific hardware?
While RDNA 4 uses dedicated AI formats (FP8), AMD is adapting the technology to work on RDNA 3 via INT8 and on RDNA 2 using standard Stream Processors.
What do you think about AMD’s strategy to extend the life of older GPUs? Let us know in the comments below, or subscribe to our newsletter for more hardware deep dives.
