AMD exec hints at AI-powered upscaling

AMD FSR with AI in the works?

AI-powered upscaling from AMD, teased by Mark Papermaster.

AMD FidelityFX Super Resolution, released in 2021, came 2 years behind NVIDIA DLSS. The first generation of FSR was based on spatial upscaling technique, which enabled many gamers to take advantage of simple upscaling thanks to open-source, cross-vendor and cross-platform support. The FSR2 expanded on that support by introducing a more complex and higher quality temporal upscaling technique.

What FSR is not is an AI-powered algorithm, thus not requiring special AI cores to work efficiently. This is the main difference between DLSS and XeSS, which take advantage of AI to fix temporal artifacts. The DLSS requires Tensor cores, while XeSS has two versions for GPUs supporting DP4a instructions and its XMX-core based implementation that offers better quality and performance.

It appears that AMD may be exploring using AI-algorithms for their upscaling as well, as hinted by Mark Papermaster (AMD CTO):

2024 is a giant year for us because we spent so many years in our hardware and software capabilities for AI. We have just completed AI enabling our entire portfolio, so you know cloud, edge, PCs and our embedded devices and gaming devices. We are enabling our gaming devices to upscale using AI and 2024 is really a huge deployment year. So now the bedrocks there, the capabilities are there. I have talked to you about all the partners that. So 2024 is for us a huge deployment. […]

— Mark Papermaster, AMD CTO

As per Papermaster’s remarks, there’s a possibility that the company is considering integrating AI into the upscaling. This statement doesn’t explicitly mention FSR. However, given that Papermaster was discussing gaming, it directs attention toward FSR. Should FSR get AI support, this would put FSR right beside DLSS and XeSS.

In reality, since both competing technologies are closed-source, one does not know to what extent AI is actually used for. For all we know, AMD plans may also be geared toward using a machine learning algorithm that removes artifacts, like in DLSS2.

Source (skip to the last segment):

[No Priors: AI, Machine Learning, Tech, & Startups] No Priors Ep. 53 | With AMD CTO Mark Papermaster (1,408 views)

Many thanks to uzzi38 for the tip!