Alongside a sneak peek at their forthcoming Xe-HPG structure, the opposite huge reveal immediately from Intel’s client graphics group comes from the software program facet of the enterprise. Along with making ready Intel’s software program stack for the 2022 launch of the primary Arc merchandise, the group has additionally been exhausting at work at their very own tackle trendy, neural net-driven picture upscaling strategies. The product of that analysis is Xe Super Sampling, or XeSS, which Intel is pitching as the perfect resolution but for prime picture high quality and low processing value picture upscaling.
As briefly hinted at by Intel initially of this week with the announcement of their Arc video card model, the corporate has been creating their very own tackle picture upscaling. As it seems, they’re truly fairly far alongside, so for immediately they’re not simply saying XeSS, however they’re displaying off footage of the expertise as nicely. Even higher, the preliminary model of the SDK will probably be delivery to recreation builders later this month.
XeSS (pronounced “ex-ee-ess-ess”) is, at a excessive degree, a mixture spatial and temporal AI picture upscaling approach, which makes use of educated neural networks to combine each picture and movement information with a view to produce a superior, larger decision picture. This is a area of analysis that has seen a substantial amount of analysis within the final half-decade, and was dropped at the forefront of the buyer area a few years in the past by NVIDIA with their DLSS expertise. Intel’s XeSS expertise, in flip, is designed to handle related use circumstances, and from a technical perspective finally ends up trying lots like NVIDIA’s present DLSS 2.x expertise.
As with NVIDIA and AMD, Intel is seeking to have their cake and eat it too with respect to graphics rendering efficiency. 4K screens are more and more low cost and plentiful, however the type of efficiency wanted to natively render at 4K in trendy AAA video games is outdoors the attain of all however the most costly discrete video playing cards. Ultimately seeking to discover methods to drive these 4K screens with extra modest video playing cards and with out the standard drop in picture high quality, this has led current analysis into sensible picture upscaling strategies, and in the end DLSS, FSR, and now XeSS.
In selecting their method, Intel appears to have gone in an analogous course as NVIDIA’s second try at DLSS. Which is to say, they’re utilizing a mixture of spatial information (neighboring pixels) and temporal information (movement vectors from earlier frames) to feed a (seemingly generic) neural community that has been pre-trained to upscale frames from video video games. Like many different features of immediately’s GPU-related bulletins, Intel isn’t going into too a lot element right here. So there are many excellent questions on how XeSS handles ghosting, aliasing, and different artifacts that may come up from these upscaling options. With that mentioned, what Intel is promising isn’t one thing that’s out of their attain in the event that they’ve actually finished their homework.
Meanwhile, given using a neural community to deal with components of the upscaling course of, it ought to come as no shock that XeSS is designed to leverage Intel’s new XMX matrix math models, that are making their debut within the Xe-HPG graphics structure. As we noticed in our sneak peek there, Intel is baking fairly a little bit of matrix math efficiency into their {hardware}, and the corporate is little question fascinated by placing it to good use. Neural network-based picture upscaling strategies stay among the finest methods to make use of that {hardware} in a gaming context, because the workload maps nicely to those systolic arrays, and their excessive efficiency retains the general hit to border rendering instances small.
With that mentioned, Intel has gone one step additional and can also be creating a model of XeSS that doesn’t require devoted matrix…
