The US Patent Applications Office shows a new AMD patent with might be their alternative to NVIDIA's DLSS. It was filed under the name Gaming Super Resolution (GSR)”, and that logically would be the artist formerly known as FidelityFX Super Resolution (FSR).
One of the big features lacking on the current Radeon RX 6000 lineup is the lack of some sort of DLSS alternative that can really boost performance, especially in raytracing. it seems AMD is now hard at work at the tech, however it would be a way different implementation than DLSS running over its Tensor cores (AMD does not have any dedicated hardware for this in their GPUs). The patent entry however describes how it works. GSR would rely on interference to scale images. Additionally, AMD engineers claim that conventional neural networks are generalizable and are trained to operate without significant knowledge of the immediate problem. The company also claimed that this deep learning approach does not take into account aspects of the original image, which can cause colors or details to be lost in the final image. Their new solution would.
AMD's answer uses linear and non-linear upscaling technology, the combination will preserve and improve the accuracy of the image. It is also said that the technology will “create high-quality image approximations and achieve high framerates”.
“The GSR architecture employs image condensation and a combination of linear and non-linear operations to accelerate the process to levels viable for gaming. GSR renders images at a low-quality scale to create high-quality image approximations and achieve high frame rates.
High-quality reference images are approximated by applying a specific configuration of convolutional layers and trigger functions to a low-quality reference image. The GSR network approaches more generalized problems with greater precision and efficiency than conventional super-resolution techniques, training the weights of the convolutional layers with a corpus of images ”. For the moment we'll have to do with this leaked patent info. This new feature might be out for a while as it sounds like software houses will need to implement support for the technology as well.
The patent was submitted in November 2019 and was public earlier this week.