Sony confirms PS5 Pro uses RDNA 2.x GPU and cooperates with AMD Amethyst to develop more artificial intelligence resources

Mark Cerny, chief architect of Sony Interactive Entertainment (SIE, Sony Interactive Entertainment), revealed more hardware details of the next-generation console PlayStation 5 Pro (PS5 Pro), including the performance-upgraded AMD RDNA 2.x architecture GPU and the codename of the cooperation with AMD. Amethyst’s machine learning/artificial intelligence project.

The focus of PS5 Pro performance improvement is still focused on the three pillars of more powerful GPU, advanced ray tracing and AI-driven PSSR super-resolution function.

The GPU uses the customized AMD RDNA 2 architecture, which Sony named RDNA 2.x. It has some advanced features introduced only in the RDNA 3 architecture, and its scale has changed from the 18 sets of WGP (36 sets of CU) of the original PS5 built-in GPU. , improved by 67% to 30 groups of WGP (60 groups of CU).

Sony emphasized that the reason why they do not want to upgrade the architecture to RDNA 3 is because they feel that PS5 and PS5 Pro are products of the same generation, and they do not want different versions of games to be released for the two models. Existing games do not need to specially adjust the rendering and shading methods, and they can all use GPU resources smoothly.

The precise specifications and configuration of the memory are not listed. It is only mentioned that PS5 Pro still uses GDDR6, but the bandwidth has increased by 28% from 448 GB/s of PS5 to 576 GB/s, and the memory capacity allocated to games has increased. 1GB or more.

With a 67% increase in GPU size and a 28% increase in memory bandwidth, the overall game performance of PS5 Pro can be improved by approximately 45% compared to PS5. For example, the original PS5 takes about 16 ms to render one frame, but the PS5 Pro only takes 11 ms to complete it.

In terms of ray tracing, the RDNA 2.x GPU specially designed for PS5 Pro can support the BVH8 acceleration structure, and the number of intersections between each cyclic polygon and the ray border is increased from 1 to 2.

Sony also added and removed the stacking that originally divided the PS5 shading code into two pieces and integrated them into the PS5 Pro, effectively simplifying the shading process and improving performance.

Therefore, compared to PS5, PS5 Pro has a slight improvement in shadow and plane reflection performance, but a significant improvement in curved surface reflection and diffusion performance.

PSSR is the abbreviation of PlayStation Spectral Super Resolution. It uses convolutional neural network (CNN) to expand the lower-resolution selected images to high-resolution, or generate frames (filling frames) to improve the smoothness of the image.

Compared with PCs, which mostly use a fixed rendering resolution and a fixed upscaling magnification ratio, PS5 Pro adopts a combination mode of dynamically adjusting the rendering resolution and magnification ratio to provide a fixed and smooth game screen.

Finally, Mark Cerny pointed out that Sony will work closely with AMD on the “Amethyst” project to develop a hardware architecture more suitable for machine learning and artificial intelligence, as well as more convolutional neural networks for games. This project is not exclusive to Sony, but allows more developers to access resources from it.