Great stuff here. Hope see more use of this.
Wonder what the chances are of Starfield making use of this.
1 Like
Other than The Coalition and ID, what XGS studio do we think will use VRS tier 2 in their games? I myself think Ninja Theory, Turn 10 and Playground. The others will probably start using it later.
Is their Creation engine now support DX12U? I believe it hinges a lot on that.
Iāve never personally ported any code from being DX11 to DX12 compliant, or anything like that so Iām not sure about the work involved. There is a couple of layers to it that I know of one being that you are changing how you deal with the API and then secondly is taking advantage of the new features of the API.
But I would hope that we do see this with the Creation engine as well as many other engines. The Slipspace engine supporting DX12U could see a lot of improvements to Halo Infinite.
I expect Unreal Engine 5 to fully support DX12U and so all of these features would be available for any of those titles. Which is a massive boon for Xbox and PC versions.
The interesting thing about UE5 and Nanite is that It does a lot of new stuff in software, that I imagine developers like Turn 10 will do in hardware using Mesh Shaders. So I wonder if VRS 2.0 results will be wildly different or very similar.
Itās exciting to see constant improvement and I hope that they are also discovering improvements and optimizations to their ML upscaling tech which didnāt look too performant the first time they showed it. There simply isnāt enough raw horsepower comparable to tensor cores to have something comparable to DLSS but if performance gains could be around 10 to 25% even that is good to see.
2 Likes
VRS is basically a hardware implementation to identify which parts of the image where high definition is not needed and can selectively reduce the quality in those areas of the image instead of reducing resolution of the entire image which can degrade image quality too much.
But the same hardware can be used for any intermediate buffer the gpu has to render (lights, shadows, particles etc) and on those super expensive buffers the gains are much better.
Few games are supporting it yet, and still not fully using it, but the results are impressive. They said that not only it reduces the time it takes to render an expensive effect, but also significantly reduces gpu usage on multiple parts (processing, texture fetches, bandwidth etc) and that opens up opportunities to increase the base resolution or add more expensive effects to the scene.
In the case of Gears 5 and Doom Eternal, they were able to keep base resolution really close to native all the time, and drastically increase settings over 1x version of these games (and in the particular case of Doom, the resolution delta compared to ps5 is also one of the biggest ones thus far). In the case of Halo Infinite, it enabled them to do a 120fps mode even for the campaign.
They are still researching some bits that can lead to even further performance gains and less quality loss, for example camera velocity is still not used in their implementation and that can enable them to be more aggressive while in movement. Another one is using VRS as a āreconstructionā tech. They had good results initially. Around the same performance gains as checkerboard rendering, but the image was better (CB adds some artifacts specially at thin transparent objects that doesnāt happen with this method).
They also said that for one particular game thatās still in development (Iām assuming itās Hellblade 2 from the description) that VRS allowed them to ramp up the post processing immensely. Like having subsurface scattering across multiple characters and even at large scale snow landscapes, with long distance shadows, God rays, volumetric particles (and lots of them) that receive lights and cast shadows, high quality motion blur and depth of field all at the same time whereas before they needed to pick which effects to use in each scene because they were too expensive to use together.
Tldr; Itās still untapped, but early results show even fast implementations can make the game look better by either increasing the base resolution or improving settings quality āfor freeā due the gpu resources that are fred up
1 Like
They do mention that at this point it was expected lower usage because devs were still making games using last gen engines. Adopting should ramp up once engines for dx12 are created and it helps that Coalition ported it to UE4 and 5
BGS uses Creation Engine 2 for Starfield, wonder if this will be making use of it.
So Microsoft is also trying to implement Dlss tech into VRS? Could it be possible that SuperRes is going to be another tool set within VRS as well?
No. Thatās a misunderstanding of what VRS is. It has nothing to do with DLSS and you donāt implement any of this āintoā VRS. Itās just performance savings from using VRS as a tool in the graphics rendering pipeline allows for more resources spent elsewhere like increased base resolution or better post processing effects.
4 Likes
Canāt wait until VRS and SFS become standard for Xboxās first-party.
3 Likes
No no, itās not at all like dlss.
What they said is that they tried to use VRS with settings that could degrate image quality (reducing resolution on parts of the image where it normally wouldnāt) and compared that to reconstruction techniques.
They found out that compared to CB using VRS that way produces similar performance gains (around 30-40%) with a better image quality.
They also explained why. For using VRS like that and for CB itās necessary to project where the pixels will move on the next frame. If this prediction fails on VRS you get a portion of the image thatās still correct, but lower res. On CB you get an artifact.
Like this:
Given their findings, while itās definitely possible to use reconstruction on top of an image generated with VRS (and there are games doing so already), but it may need a balancing out, because VRS performance gains increase proportionally with resolution while reconstruction tech usually becomes more expensive the higher the base resolution is (because thereās more data to process).
Coalition did mentioned on other presentations that the UE5 reconstruction tech is really good and can make 4k like images from 1080p base resolution like DLSS (the reason for that is that with nanite they can pool the silhouette of objects in higher definition than 1080p, and the edges are more important for the perceived detail, so perhaps we will see better reconstruction tech on consoles even if they never tap into ML (though Iām sure weāll see some cases)
4 Likes
Thanks for the explanation, I got confused with the using VRS to reconstruct an image part, as it sounded like what dlss does.