Most contemporary upscalers are based on temporally fed algorithms. This applies to NVIDIA DLSS 2.0, AMD FSR 2.0, Intel XeSS as well as Unreal’s TSR upsampler. The core difference between them is that FSR 2.0 and TSR don’t rely on dedicated hardware units to run them at appreciable speeds. The reason is that they don’t leverage a neural network to improve the image quality with each training session. The difference in image quality between the two kinds of upscalers is still hard to spot. Furthermore, it allows FSR 2.0 to run on virtually any modern GPU, significantly expanding the accessibility.
Intel plans on adopting a hybrid approach for XeSS. The Arc GPUs will use the XMX matrix cores to boost upscaling while mixed-precision compute will be adopted for NVIDIA and AMD systems. If a recent LLVM repository update is to be believed, then AMD is working on a similar approach.
The RDNA 3 architecture will seemingly support WMMA (Wave Matrix Multiply Accumulate) instructions for matrix multiplication, much like NVIDIA’s Tensor cores and Intel’s XMX units. Now, since matrix multiplication instructions are too long for standard vector units, AMD is likely to leverage dedicated machine learning units (possibly from its Xilinx portfolio) to accelerate them.
For a gaming GPU, the only practical use of machine learning is with respect to neural network-driven image upscaling. FSR 2.0 might get an update later next year adding a neural network to further fine-tune the temporal upscaler, or it could simply be introduced with FSR 3.0. From a technical perspective, this will put FSR, DLSS, and XeSS on equal footing.