Last week, AMD launched its latest graphics hardware, featuring the innovative RDNA 4 architecture. This new technology markedly enhances machine learning performance and paves the way for FSR 4, a new machine learning (ML)-based upscaler that rivals Nvidia’s DLSS in quality. Following the launch, AMD disclosed that FSR 4 was developed in collaboration with Sony as part of Project Amethyst. Mark Cerny, the lead system architect for PlayStation, mentioned in an interview with Digital Foundry that they aim to incorporate a similar upscaling solution on the PS5 Pro for titles slated for 2026, representing the next evolution of their PSSR technology.
Cerny explained that the neural network and training methods used in FSR 4’s upscaler are the initial outcomes of the collaboration. He expressed pride in the advancements achieved, stating that FSR 4 employs a more sophisticated approach that surpasses the crispness provided by PSSR. Digital Foundry has made a video available for viewers to compare FSR 4 with previous versions like FSR 3.1, as well as Nvidia’s DLSS 3.7 and 4.0. Initial impressions suggest that FSR 4 is a substantial upgrade over FSR 3.1 and shows notable competitive qualities against Nvidia’s offerings.
Despite these advancements, it will take time for this new technology to be implemented on the PS5 Pro. Cerny noted that for 2025, developers are focused on integrating PSSR into their games while simultaneously beginning work on the new neural network for the PS5 Pro. They anticipate having an upscaler akin to FSR 4 ready for the 2026 games, although the implementation is complex and requires oversight. For context, the RX 9070 and RX 9070 XT, AMD’s recent GPUs, boast much higher hardware specifications than the PS5 Pro, representing a middle-tier price range.
There has been some indication that FSR 4 could incur a considerable computational cost, yet Cerny is optimistic that the PS5 Pro’s machine learning hardware can handle the demands without needing a major redesign. He highlighted that PS5 Pro’s peak performance reaches 300 8-bit TOPS, which aligns well with the specifications of the latest AMD graphics cards.
It’s essential to clarify that PS5 Pro’s hardware design is distinct from RDNA 4, necessitating what Cerny refers to as a “reimplementation” of the new upscaling network for the console. There’s a growing expectation for collaborative hardware features focused on machine learning that could enhance compatibility across different gaming devices.
Project Amethyst is recognized as a significant initiative that is expected to extend beyond the PS5 Pro’s lifecycle, likely influencing the next generation of gaming consoles as well. Cerny mentioned that the project has two primary objectives: the immediate goal of co-developing neural network architectures for game graphics and a long-term ambition to create an optimal hardware design for machine learning tasks in gaming, leveraging insights from both Sony and AMD’s expertise.
While this collaboration is firmly based in the gaming domain, Cerny emphasized that the technology being developed would have applications beyond PlayStation, fostering a broader capacity for machine learning across various platforms. The ultimate aim is to allow developers to transfer their code seamlessly across devices.
AMD’s GPUOpen has historically been the platform through which the company has disclosed its innovations, raising curiosity about how this will evolve with the era of machine learning. FSR 4 has already been discussed in connection with enhancing mobile device battery life. For Sony, Project Amethyst represents a shared commitment to mutual goals while recognizing that each entity might adapt the technologies for different purposes.
Anticipation is building regarding the developments in machine learning and ray tracing through the Amethyst project. AMD’s recent showcases have already teased innovative techniques, including path tracing and neural rendering technologies, hinting at future capabilities in graphics.