Thursday, April 4th 2024
Intel Releases XeSS 1.3, Improves FPS Across Presets with New Resolution Scaling, Improved Upscalers
Intel on Wednesday released the XeSS 1.3 performance enhancement, which works with Intel Arc "Alchemist" discrete GPUs, and Intel Arc iGPUs powering the Core Ultra "Meteor Lake" processors. The new super sampling technology brings several under-the-hood improvements to the upscaler, which improves image quality at a given resolution. Intel leveraged this improved upscaler to rework the resolution-scale of each performance preset, thereby improving performance per preset; while also introducing new presets at both ends of the resolution scale. The company released the XeSS 1.3 SDK on GitHub, so developers can begin exploring the tech and implementing it on their games.
The XeSS 1.3 update is predicated on an improved upscaler. Intel says that it has updated the AI models with new optimizations, and additional pre-training, particularly with difficult to upscale elements (such as meshes, as in textures with a lot of alpha pixels). The updated upscaler offers better reconstruction of detail, better AA, less ghosting, and improved temporal stability. Intel then used this up change the resolution scale across all its presets as detailed in the table below. It introduced the new Ultra Performance preset that does a 3.0x resolution scale, something that didn't exist in the previous versions of XeSS. On the other end of the spectrum is Native AA, a mode that has zero upscaling, but just the full application of the upscaler as a varnish—this is essentially Intel's take on DLAA.
An increase in resolution scale implies that the game is being rendered at an even lower resolution, for the upscaler to work its magic with. The lower render resolution translates to performance improvements, which Intel presented in the tables above, for an Arc A750 discrete GPU, and a Core Ultra 7 155H mobile processor iGPU. Intel has internally tested XeSS 1.3 with a number of popular titles, at a given preset; and all its new presets (new resolution scales) with Cyberpunk 2077.
The XeSS 1.3 update is predicated on an improved upscaler. Intel says that it has updated the AI models with new optimizations, and additional pre-training, particularly with difficult to upscale elements (such as meshes, as in textures with a lot of alpha pixels). The updated upscaler offers better reconstruction of detail, better AA, less ghosting, and improved temporal stability. Intel then used this up change the resolution scale across all its presets as detailed in the table below. It introduced the new Ultra Performance preset that does a 3.0x resolution scale, something that didn't exist in the previous versions of XeSS. On the other end of the spectrum is Native AA, a mode that has zero upscaling, but just the full application of the upscaler as a varnish—this is essentially Intel's take on DLAA.
21 Comments on Intel Releases XeSS 1.3, Improves FPS Across Presets with New Resolution Scaling, Improved Upscalers
Looking forward to comparisons.
1.2 Ultra Quality got 42 FPS, but 1.3 Ultra Quality Plus got 41 FPS
1.2 Quality got 50 FPS, but 1.3 Ultra Quality got 49 FPS
1.2 Balanced got 58 FPS, but 1.3 Quality got 57 FPS
1.2 Performance got 66 FPS, but 1.3 Balanced got 65 FPS.
On hardware that can't run DLSS, it's a race to see what gets more adoption, FSR3.1 or this. Either would fix the distracting artifacts of their implementations.
Doing some testing for fun, running Cyberpunk on an i5 1240p laptop, no dGPU. You have to go down to 1280x720 + Performance/Ultra Performance to get 30 FPS, that's 240p gaming. Both FSR 3.0 (mod) and XeSS 1.2 look terrible that low :D
"1.2 Ultra Quality got 42 FPS, but 1.3 Ultra Quality Plus got 41 FPS" you should compare Ultra Quality 1.2 to Ultra Quality 1.3 which is 42fps vs 49fps but you went with plus. same goes for every other comparison you made.
That screenshot seems to hint it does even better. But it's a sample size of 1 so...
No upscaling.
50% internal resolution, XeSS 1.2.
50% internal resolution, XeSS 1.3.
Overall, feels like a dozen minor improvements here and there with virtually zero performance penalty compared to older SDK. You WILL NOT be able to catch these differences on my screenshots, recordings, or whatnot, they're too subtle for it. Yet, no denying, it became better. Better enough for 50% internal resolution (1.3) to be at least on par with 58% internal resolution (1.2) in the Cyberpunk 2077 but still feel less jittery and shimmery. Not by much though.
Cool, I guess.
I'm curious about it's "AI" claims that perhaps it would use NPU or AVX VNNI instruction assistance to aid with speed, but nothing in the documentation points to that. It seems similar to motion estimation tech that was GPU accelerated in video encoding tech years ago, PRE AI. Who knows though I can't seem to find much.
It would be nice to test it on CPUs fps that support VNNI with and without. Also Arc GPUs vs non using it. I have a lowly 3950x with a rx 6800 XT so I can't test AVX VNNI.
Ryzen 7000 and Alder Lake and newer vs older ones.
Edit: It would be disappointing if they simply labeled it AI for the tone mapping and motion estimation feature. That's pre AI tech. They were the ones who pushed for performance optimized ai if they don't include it, just change the scales this is doesn't look great.