Intel AI-based XeSS upscaling seems impressive at first glance, will work on all competitor hardware including integrated GPUs
Alongside detailing the Arc Xe-HPG DG2 Alchemist architecture, Intel also spoke about its own implementation of an AI-based supersampling technique. Intel calls this XeSS, and initial results shown by Intel already seem quite impressive.
Intel's Lisa Pearce said that Arc DG2 Alchemist has been developed from the ground-up to support the complete Microsoft DirectX12 Ultimate feature set. Lisa also alluded to upcoming UI features such as live streaming, performance tuning, and AI-assisted virtual cameras.
Lisa then spoke about render quality trade-offs between image quality and performance. She noted that traditional upscaling techniques fall short for use with high-quality graphics and when computationally intensive elements such as ray tracing are involved. Intel's solution to achieve both high performance and high quality will be coming in the form of XeSS.
In the case of Nvidia's DLSS 2.0 implementation, the game developer uses Nvidia's library to input data such as a lower resolution of the rendered frame, motion vectors, depth buffer, and jitter for each frame. This information is then fed to the convolutional autoencoder that upscales the image to the target resolution. The AI math is done by Tensor cores present on Nvidia GPUs from the Turing generation onwards. Intel's XeSS also works on a similar principle, though it will likely employ different means to achieve a similar end result.
According to Intel, XeSS also uses jitter and motion information along with history of the image and other rasterization and lighting parameters. It then subjects them to neural network processing that is done by the XMX or matrix engines in each Xe-core. The resulting image is then subjected to post-processing before being displayed on the screen.
Lisa demoed a real-time rendering of a 4K Unreal Engine demo that is actually rendered at 1080p by the engine and upscaled by XeSS. The demo looks impressive for starters with not much of a perceivable quality loss. That being said, though up to a 2x performance boost can be expected, it is not without some additional cost in terms of frame render times.
AMD's FidelityFX Super Resolution (FSR) is not AI-based unlike Nvidia and Intel, but it was praised for its ability to run on almost all GPU hardware including competing products from Nvidia. Now, Intel is also taking a more open approach with XeSS. The demo showed by Lisa was run on an Arc Xe-HPG Alchemist GPU with the upscaling being performed by the XMX engines. For competitor hardware, Intel XeSS will use the widely available DP4a instruction set. This will result in some quality/performance trade-offs but does broaden XeSS's applicability to all kinds of GPUs including integrated GPUs.
The XeSS SDK that leverages the XMX engines in Arc Alchemist GPUs will be made available for game developers this month while the XeSS DP4a SDK will be available later this year.
Intel Architecture Day 2021 Keynote