CES 2025: AMD Announces RDNA 4, Machine Learning-Powered FSR 4; More Info Coming Later This Quarter

mmorpg.com:

AMD has officially announced its next GPU architecture, RDNA 4, today at CES, and while the company didn’t spend much time talking about what this means (more to come later this quarter), during a pre-brief ahead of today’s press conference, the company broke down a little bit more about what gamers can expect from the upcoming Radeon boards. 

Building off of the last generation RDNA 3 architecture, AMD states that RDNA 4 will come with what they are calling “optimized compute units,” tuned to drive improvements in IPC and frequency compared to RDNA 3. Additionally, with its third generation versions of its Raytracing Accelerators, AMD states that RDNA 4 users should see improved RT performance per CU compared to previous generations. 

Additionally, with AI taking center stage in so many tasks nowadays, especially with machine learning-driven AI reconstruction such as Nvidia’s Deep Learning Super Sampling (DLSS) and Intel’s Xe Super Sampling (XeSS), AMD’s RDNA 4 cards are built with AI in mind. This includes its second generation of AI Accelerators, as well as helping to power AMD’s own version of ML-powered super sampling, FidelityFX Super Resolution 4 (FSR 4).

FSR 4 was built with RDNA 4 in mind, taking advantage of the gains made in the architecture’s AI compute performance over RDNA 3. Unlike FSR 3, which was platform agnostic in how it was implemented, it looks like FSR 4 will be exclusive to the upcoming RDNA 4 cards, according to the materials sent to the press ahead of today’s announcements. 

Related:

AMD RX 7900 XTX and RX 7900 XT Review

While we don’t know much about how this works, whether it will be on par with today’s competitors, especially Nvidia’s DLSS, which has been in the market since their 20-series GPUs, and even how it will integrate potential frame generation technologies, I will admit it’s exciting to finally see AMD implementing machine learning for this process. FSR 3’s upscaling has always been hit or miss, with some implementations giving