First of all, it should be clarified that this is an incomplete patent, because it is not possible for the moment to access the images which accompany it and which are essential for a full understanding of it. this. Remember that a patent first describes the components of a technology or service, then explains how it works through the relationship between the components. What he does through numbered references to his drawings.
We are therefore waiting for a full version of the AMD patent to be able to discuss it in more detail.
New AMD patent
A new patent appeared in the United States Patent Office titled GAMING SUPER RESOLUTION assigned to AMD and dated November 2019, just a year before the official announcement of its FidelityFX Super Resolution. What has led many to believe that it is the AMD Fidelity FX Super Resolution patent, which promises to increase the resolution and / or speed of games on the RX 6000 and therefore its overall performance.
However, the patent would not be related in principle to the FidelityFX super resolution because it uses a different methodology to achieve more resolution. Let’s not forget that AMD has not yet given official details on the FidelityFX Super Resolution. Officially, its representatives claim that its code is in a very early version and in the hands of developers.
No obstante la narrativa de la patent no corresponde con la del FidelityFX Super Resolution, especial por el hecho that the patent menciona el uso de unidades del tipo Tensor Cores in a hardware of AMD, las cuales its unpublished hasta el momento in the GPU of the brand. Although we can’t forget that FidelityFX is based on DirectML, the algorithms of which can work on different types of hardware, but with different results and efficiency for everyone.
Tensor Cores on AMD GPUs
However, in the AMD patent we found a language referring to elements that are not found in an AMD RX 6000, as for example in the following paragraph:
The devices and methods include a network architecture for Super Game Resolution (GSR), which efficiently super-resolve images from convolutional pathway and generalizable. GSR architecture uses condensation and the combination of linear and non-linear operations to speed up the process to viable levels for games. The GSR renders low quality images and creates high quality images to achieve high frame rates.
GSR is defined here as new hardware, the term convolution makes us think of neural networks of this type, something the RX 6000 does in that they don’t have systolic arrays, which NVIDIA Tensors are. Cores. .
The GSR network makes more precise and precise approximations than conventional super-resolution techniques by training the weights of the convolutional layers with a corpus of images.
AMD has therefore patented a super resolution system via AI, which works the same as NVIDIA’s DLSS, so it is not intended for the RX 6000, but for hardware with Tensor type units, which are described in the patent under the name of GSR.
Are these GSR units part of the new features that we will see in RDNA 3? Everything points to yes, because the lack of Tensor cores in AMD causes a huge disadvantage in several areas compared to NVIDIA, not only when it comes to DLSS but also with techniques such as Denoising for Ray Tracing.