Thursday, August 19th 2021
Intel's DLSS-rivaling AI-accelerated Supersampling Tech is Named XeSS, Doubles 4K Performance
Intel plans to go full tilt with gaming graphics, with its newly announced Arc line of graphics processors designed for high-performance gaming. The top Arc "Alchemist" part meets all requirements for DirectX 12 Ultimate logo, including real-time raytracing. The company, during the technology's reveal, earlier this week, also said that it's working on an AI-accelerated supersampling technology. The company is calling it XeSS (Xe SuperSampling). It likely went with Xe in the name, as it possibly plans to extend the technology to even its Xe LP-based iGPUs and the entry-level Iris Xe MAX discrete GPU.
Intel claims that XeSS cuts down 4K frame render-times by half. By all accounts, 1440p appears to be the target use case of the top Arc "Alchemist" SKU. XeSS would make 4K possible (i.e., display resolution set at 4K, rendering at a lower resolution, with AI-accelerated supersampling restoring detail). The company revealed that XeSS will use a neural network-based temporal upscaling technology that incorporates motion vectors. In the rendering pipeline, XeSS sits before most post-processing stages, similar to AMD FSR.
While AMD's FSR technology is purely shader based, the Intel algorithm can either use XMX hardware units (new in Intel Xe HPG), or DP4a instructions (available on nearly all modern AMD and NVIDIA GPUs). XMX stands for Xe Matrix Extensions and is basically Intel's version of NVIDIA's Tensor Cores, to speed up matrix math, which is used in many AI-related tasks. The Intel XeSS SDK will be available this month, in open source, using XMX hardware, the DP4a version will be available "later this year".
Source:
VideoCardz
Intel claims that XeSS cuts down 4K frame render-times by half. By all accounts, 1440p appears to be the target use case of the top Arc "Alchemist" SKU. XeSS would make 4K possible (i.e., display resolution set at 4K, rendering at a lower resolution, with AI-accelerated supersampling restoring detail). The company revealed that XeSS will use a neural network-based temporal upscaling technology that incorporates motion vectors. In the rendering pipeline, XeSS sits before most post-processing stages, similar to AMD FSR.
While AMD's FSR technology is purely shader based, the Intel algorithm can either use XMX hardware units (new in Intel Xe HPG), or DP4a instructions (available on nearly all modern AMD and NVIDIA GPUs). XMX stands for Xe Matrix Extensions and is basically Intel's version of NVIDIA's Tensor Cores, to speed up matrix math, which is used in many AI-related tasks. The Intel XeSS SDK will be available this month, in open source, using XMX hardware, the DP4a version will be available "later this year".
46 Comments on Intel's DLSS-rivaling AI-accelerated Supersampling Tech is Named XeSS, Doubles 4K Performance
So is it blurry, temporal artefacts everywhere ?
We'll never know, Intel really needs to start showing tech instead of talking of tech 5 years down the line, what a screaming baby they've become.
I just want to see Products!
infact I dont even understand that comment because TAA IS temporal and that has artifacts in motion....
On Topic:
I would like to know what the difference is between this and Nvidia's implementation because in base terms it all sounds the same
As for artefact on temporal solution, it's way harder to do a good implementation but it's not impossible. But it's probably more something AAA studios will be able to do. And that is also one of the great point of FSR, not as good as temporal solution, but easy to work with and quick to implement for smaller studios.
So with XeSS and FSR, there are probably no reason that a game doesn't implement one or both of them in the future...
Temporal component with motion vectors is basically more contemporary method for upscaling (or reconstruction) that both DLSS and now XeSS are using. FSR will no doubt evolve to include the temporal component but today it does not have any of that.From what it looks like, this is what Intel intends to exploit. If they open up XeSS, that is quite likely going to succeed in becoming the "standard" way of upscaling quite fast.From what Intel disclosed, Xe-HPG seems to go pretty heavy on matrix math, even heavier than Nvidia. IIRC AMD is going to have something similar in RDNA3.
No doubt everything is doable with shaders but more purpose-built units are likely faster for these ML-based techniques.
They even made their own FXAA competitor you can force in the driver (it does almost nothing to edges, and it costs nearly as much performance.)
Because (unlike SMAA and FXAA) we will never be able to insert FSR, game devs are going to have to choose on or two techs to bother with!
And while TAA is, by definition, temporal, FSR isn't. So it isn't really a fair comparison in the first place, people and media comparing them and concluding 'yeah, they're comparable' are simply lying to themselves. Like that HU video where they compare a bunch of static screenshots while saying nothing about FSR's performance in motion where it's all but a blurry mess, better than nothing i guess, but not at all comparable to even first gen DLSS.I don't know what you were reading, but FSR is very much not temporal. It's a relatively simple downscaling-upscaling based on Lanczos, a popular opensource upscaling algorithm. There are no way to make it temporal while retaining it's ease of implementation by game developers, it's simply a post-process shader and was developed as such. Whether AMD would eventually come up with an actual alternative to DLSS is another question. I sure hope so, but first, they would need to add DL acceleration blocks to their hardware.
Its a reason not to fund proprietary bullshit, but it requires competitors in the market, and so far, all we had was two giants trying to invent the best sauce for the same plate of food. But its the same plate of food.
Now that there are three, it gets a whole lot harder to defend that idea. Three inventors of a wheel where two are destined to fail is quite a bit more risk than the odd 50% adjusted for market(ing) share. Devs won't go about supporting three technologies either. They want them fed to them or they're not happening. Another marketing idea that won't be working anymore, is the good old difference of 'AMD open, Nvidia closed'. Its no longer a real thing to stand out with from the 'crowd' when a third player deploys yet another policy. Three is a great number for us.
- If the deciding factor will be a financial incentive for the developers, nVidia will win,
- If ease of implementation, FSR will win.
- If performance, Intel has the upper hand, if their first party benchmarking is to be believed. But they will have to really make it into a polished product since they have zero market share and brand recognition as far as dedicated GPUs go.
Personally I think if AMD can create FSR 2.0 with improved quality and performance AND make it hardware agnostic, it will be a clear winner for developers and consumers. As much as I admire the complexity and elegance of AI, the mass market works on the KISS principle.