AMD today released FidelityFX Super Resolution, a feature that significantly improves game performance with minimal trade-offs to image quality. The idea isn't to dial down the eye-candy of your game, such as textures, geometry detail, effects, etc., but to render the game at a lower resolution and upscale the output back to your desired resolution while reconstructing details so it doesn't feel like a low-resolution image that's simply been stretched to fit. The computational cost involved in restoring details to the upscaled image is less than rendering the game at the higher resolution, and so the net result is increased framerates.
NVIDIA pioneered such a technique two years ago, in 2018, using machine learning. NVIDIA's Deep Learning Super-Sampling (DLSS) was introduced as one of two killer features of its GeForce RTX graphics cards (the other being real-time raytracing). The updated DLSS 2.0 in particular gained much popularity with gamers as it provided them with a practically free 50-100% frame-rate boost, which came in handy with games that use raytracing. AMD has finally caught up and introduced their version of super-sampling with FidelityFX Super Resolution (FSR), which we review today.
At their core, both FSR and DLSS aim to do the same thing—enable rendering at lower resolution with minimal image-quality loss to eke out performance, but how they go about it vastly differs. NVIDIA DLSS 2.0 is a pre-trained AI network that takes in the lower-resolution frames and motion vectors (information that suggests which way the image is about to change) to generates a higher-resolution image based on those. On the user-side, only inference happens, all training is done at NVIDIA's labs and is not game specific. In the older 1.0 version of DLSS, NVIDIA used a pre-trained network that was optimized per-game, but didn't take into account motion vectors to reconstruct the details. A drawback with DLSS is the hardware dependency of the Convolutional Autoencoder AI. Within NVIDIA's product stack, only the GeForce RTX 20-series, RTX 30-series, and Quadro RTX-series graphics cards support DLSS, so this ability is limited to fewer products than FSR, which can work on any somewhat recent graphics card.
FidelityFX Super Resolution, on the other hand, is a technology that is integrated with the game's own rendering pipeline. Game developers can simply transplant FSR code, which will be open sourced at AMD's GPUOpen site in July (according to AMD). The technology doesn't require specialized hardware, and so AMD enabled the feature for all its Radeon GPUs dating back to "Polaris" (RX 400 series), as well as NVIDIA GeForce GPUs as far back as the GTX 10-series "Pascal." AMD's strategy is to repeat FreeSync, where it came out with a simpler alternative to G-SYNC, that went on to become popular due to its simplicity and royalty-free nature. On the next page, we will dive into the details of how FSR works, and how it can be configured.
In this review, we will analyze the image quality and performance impact of the various modes of AMD FidelityFX Super Resolution and also show you how it affects performance on various graphics cards, including NVIDIA GeForce models.