Nvidia DLSS vs AMD FSR

Which AI upscaling technique has the edge?

- ZAK STOREY

4K GAMING is a problem. It’s always been a challenge to render at that high a resolution. If anything, the res arrived long before any graphics card or console could touch it. The problem is simple: compared to 1080p, 4K represents a 300 percent increase in the number of pixels on display at any given time. We effectively quadrupled the total workload overnight by shifting to a 4K resolution. Not only that, but even in 2024, eliminating jaggies is as important as it ever was. Antialiasing efforts similarly are incredibly graphically intensive. Jaggies occur because games render textures in pixels, and because pixels are square, you end up with artifacts creating a jagged texture, often along the edge of a model.

We ended up in a challenging situation where the resolution and refresh rates increased, and although pixel density had shot up, jaggies were still a thing, and the methods we were using to eliminate them via anti-aliasing likewise cost even more performance when processed at 4K.

To get around this, Nvidia began to implement DLSS in its first-generation RTX graphics cards. With the added benefit of Tensor cores in its cards, it could finally leverage machine learning and AI supersampling (predominantly done on its own supercomputers to create models for the tensor cores to utilize) to effectively eliminate that jaggie problem and upscale textures. Initially, DLSS 1.0 focused mainly on anti-aliasing. It was only with DLSS 2.0 that we received the supersampling as well. In that case, instead of rendering a game at 4K by enabling DLSS, the GPU would drop native resolution and upscale that back to 4K, using AI to identify the patterns based on the models provided by Nvidia’s own supercomputers, and simultaneously improve frame rates in the process.