The video examines AMD’s accidental release of the FSR4 INT8 upscaler version, testing its performance on RDNA2 GPUs like the RX 6700 and Xbox Series X, revealing that while it offers some quality improvements, it falls short in performance compared to FSR3 and native rendering, especially at 1440p. It concludes that despite the potential for broader hardware compatibility, FSR4 INT8’s practical use is limited by hardware support and its relatively high processing cost.
The video discusses AMD’s recent release of the FSR4 (FidelityFX Super Resolution 4) machine learning-based upscaler source code to GitHub. Interestingly, AMD inadvertently included a version of FSR4 that uses INT8 machine learning instructions instead of the FP8 version used in shipping products. This accidental release broadened the utility of FSR4, making it potentially usable on older GPUs, not just the latest RDNA4 architecture.
The presenter tests the FSR4 INT8 implementation on an RDNA2 GPU, specifically the RX 6700 10GB, which is comparable to the PlayStation 5’s graphics technology but supports INT8 instructions unlike the PS5. The Xbox Series X, also based on RDNA2, could theoretically run this technology, though no exact equivalent GPU exists. Performance comparisons show that while FSR4 offers a 28% performance uplift over native 4K rendering in performance mode, it is slower than both the Xbox Series S and FSR3, which achieve higher performance gains.
At 1440p resolution, FSR4 continues to lag behind competitors on RDNA2 hardware. The Xbox Series S is about 20% faster than native rendering, FSR4 only improves by 11%, and FSR3 outperforms FSR4 by roughly 30%. The video highlights that FSR4’s processing time is significantly higher, which impacts its ability to reach the target 60 frames per second frame time budget of 16.7 milliseconds. This raises questions about whether the quality improvements justify the performance cost.
The video also showcases a custom Alan Wake 2 benchmark demonstrating FSR4 performance mode at 1440p with low settings compared to FSR2 balance mode and native 720p rendering. Despite adding 2.3 milliseconds of render time, FSR4’s higher internal rendering resolution gives it a slight 4% performance advantage and noticeably better image quality, particularly in reducing flicker and improving temporal coherence in foliage and power lines.
Lastly, the video notes that the PlayStation 5 base model does not support INT8 machine learning instructions, limiting the practical use of the FSR4 INT8 version on that platform. Testing on an AMD Radeon RX 5700 XT, which also lacks machine learning support, showed that FSR4 INT8 cannot fallback to non-ML instructions, rendering it unusable on such GPUs. The presenter concludes that while the INT8 release is an interesting development, its real-world viability depends heavily on hardware support and optimization.