Current Generation Games Analysis Technical Discussion [2023] [XBSX|S, PS5, PC]

Status
Not open for further replies.
Hardware Unboxed Looks at DLSS vs. FSR in recent games

tl/dv: Just a whitewash for DLSS unfortunately. I say unfortunate, as I really would not like to be tied to one architecture in order to get competent reconstruction, but it is what it is. FSR 2.1 Quality at 4K can be 'ok' in some spots, but once you go below that it falls apart quickly. Even at 4k/Quality that there can still be some egregious shimmering issues, aspects which really stand out to me and I hate when I see it in DLSS.

What it ultimately means is that effectively, for low to midrange cards, you really don't have an effective reconstruction technique on Radeon, which really has to be factored into their price/performance IMO. Better than before...I guess? Certainly not as much as I would have hoped. I mean if you're comparing cards and one is with DLSS at 1440P and another is FSR 2, are you really comparing the same thing if one image is flickering in and out of existence? Unreal 4's TAAU looks to be far superior, it may not be as crisp but it's far more stable - there's probably a reason you're not seeing FSR 2 take the console world by storm compared to other reconstruction techniques, albeit on a closed platform I reckon it will receive more attention from devs on tailoring it to games to avoid these poor edge cases.

I mean, yikes man:

View attachment 8705
Yeah, I wasn't expecting it to be such a mismatch. Sigh, I guess we'll have to wait a while for a hardware-agnostic solution that makes DLSS obsolete.
 
I'm interested in knowing why we haven't seen something trained by Microsoft for DirectML that could run across many gpu vendors. Seems like DirectML has been sitting around totally unused, which makes me think it's got some serious flaws. Or it could just be the cost of creating the models.
 
Seems like DirectML has been sitting around totally unused, which makes me think it's got some serious flaws.
Intel mentioned DirectML is too slow for them to use it to implement XeSS, even for XeSS version that run on hardware other than Intel. So yeah, it seems it's not fast enough.
 
I'm interested in knowing why we haven't seen something trained by Microsoft for DirectML that could run across many gpu vendors. Seems like DirectML has been sitting around totally unused, which makes me think it's got some serious flaws. Or it could just be the cost of creating the models.
There is no incentive for any one to devote the software effort, which I'd imagine is herculean.
 
I would think if MS could train a performant AI upscaler it would give them an advantage in the console space.
Xbox probably lacks the tensor math to make anything worthwhile. I also doubt Microsoft wants to deal with getting it performant on the breadth of PC GPUs out there. Almost all of their 1st party studios will be using UE5 which has respectable reconstruction tech.
 
I would think if MS could train a performant AI upscaler it would give them an advantage in the console space.
cost prohibitive I think. You could offer it as part of the DirectX toolset, but they're already using it anyway.
 
Hardware Unboxed Looks at DLSS vs. FSR in recent games

tl/dv: Just a whitewash for DLSS unfortunately. I say unfortunate, as I really would not like to be tied to one architecture in order to get competent reconstruction, but it is what it is. FSR 2.1 Quality at 4K can be 'ok' in some spots, but once you go below that it falls apart quickly. Even at 4k/Quality that there can still be some egregious shimmering issues, aspects which really stand out to me and I hate when I see it in DLSS.

What it ultimately means is that effectively, for low to midrange cards, you really don't have an effective reconstruction technique on Radeon, which really has to be factored into their price/performance IMO. Better than before...I guess? Certainly not as much as I would have hoped. I mean if you're comparing cards and one is with DLSS at 1440P and another is FSR 2, are you really comparing the same thing if one image is flickering in and out of existence? Unreal 4's TAAU looks to be far superior, it may not be as crisp but it's far more stable - there's probably a reason you're not seeing FSR 2 take the console world by storm compared to other reconstruction techniques, albeit on a closed platform I reckon it will receive more attention from devs on tailoring it to games to avoid these poor edge cases.

I mean, yikes man:

View attachment 8705

(Also this means if a PC publisher adds just FSR 2 because it's 'good enough' or (more likely) due to AMD sponsorship, fuck off. You deserve to be critiqued as harshly as not including any reconstruction at all.)

I wonder if those Ties at 4k would also become DLSS+ if they played on a bigger 4k screen like the LG TVs with more pixelization for FSR in distance that is not as obvious with higher PPIs for 27'' monitors.
 
Are
Xbox probably lacks the tensor math to make anything worthwhile. I also doubt Microsoft wants to deal with getting it performant on the breadth of PC GPUs out there. Almost all of their 1st party studios will be using UE5 which has respectable reconstruction tech.

Are ML models not hardware agnostic? I would have thought they’d be very portable, but is that not the case?
 
It's frustrating, cuz Microsoft themselves claimed they included some level of ML acceleration for upsampling purposes:

xsx ai supersampling.jpg

Given some of Microsoft's demonstrations of DirectML, I was pretty confident they were ultimately developing it for use in Xbox games. But that seems to have either never been a correct assumption, or just has evaporated for whatever reason(too slow, etc). I know the ML capabilities of Xbox aren't exactly substantial, but you'd think they might have had some real plan to make use of it...
 
I'm interested in knowing why we haven't seen something trained by Microsoft for DirectML that could run across many gpu vendors. Seems like DirectML has been sitting around totally unused, which makes me think it's got some serious flaws. Or it could just be the cost of creating the models.

MS is releasing a DirectML upscaler. However, it’s not for games but a video upscaler intergrated into Edge.
 
It's frustrating, cuz Microsoft themselves claimed they included some level of ML acceleration for upsampling purposes:

View attachment 8707

Given some of Microsoft's demonstrations of DirectML, I was pretty confident they were ultimately developing it for use in Xbox games. But that seems to have either never been a correct assumption, or just has evaporated for whatever reason(too slow, etc). I know the ML capabilities of Xbox aren't exactly substantial, but you'd think they might have had some real plan to make use of it...
MS so far this generation has really failed to deliver on anything they've demonstrated or promised this generation so far. Pretty much no games yet to demonstrate their DX12U technologies. It's pretty sad.
 
It's frustrating, cuz Microsoft themselves claimed they included some level of ML acceleration for upsampling purposes:

View attachment 8707

Given some of Microsoft's demonstrations of DirectML, I was pretty confident they were ultimately developing it for use in Xbox games. But that seems to have either never been a correct assumption, or just has evaporated for whatever reason(too slow, etc). I know the ML capabilities of Xbox aren't exactly substantial, but you'd think they might have had some real plan to make use of it...
Microsoft sold VRS as the second coming of Jesus and I seen the magnificent boost it gives in perfomance in game like Dead Space Remake. I stop to believe to whatever they claimed about such stuff time ago.
 
MS so far this generation has really failed to deliver on anything they've demonstrated or promised this generation so far. Pretty much no games yet to demonstrate their DX12U technologies. It's pretty sad.
Let's hope atleast their own games like FM do. If that also doesn't support it, that would be pathetic.

I'm not too hopeful though, these features were never mentioned once in their developer directs.
 
If you guys don't want to get disappointed by the lack of flashy public facing releases highlighting developer-focused features, stop scrutinizing news about developer-focused features.
 
If you guys don't want to get disappointed by the lack of flashy public facing releases highlighting developer-focused features, stop scrutinizing news about developer-focused features.
Tell them to stop marketing them directly to us to sell their hardware then...
 
If you guys don't want to get disappointed by the lack of flashy public facing releases highlighting developer-focused features, stop scrutinizing news about developer-focused features.

Scrutiny of marketing of said features vs. real world impact of their implementation, or in this case the complete lack of real-world implementation, does not seem out of pocket for a technology discussion forum.
 
MS so far this generation has really failed to deliver on anything they've demonstrated or promised this generation so far. Pretty much no games yet to demonstrate their DX12U technologies. It's pretty sad.

Well they've delivered the hardware where they can (Xbox), and the API and documentation for PC, and it's supported by all the three main GPU vendors. MS's ability to force developers to use certain features that aren't high enough priority for the developer is limited. There's a huge amount of inertia in complex, long running software development projects. The technologies MS have been talking about are real though, and they do what MS have told developers they do.

DXR has found a lot of traction, and mesh shaders are going to be incorporated into the worlds most popular engine where appropriate. But you maybe don't need to target mesh shaders on cross gen games ....

SM 6.4 allows for the acceleration of mixed precision packed integer stuff - the "AI acceleration" that MS talked about for Xbox which became a standard part of PC RDNA2. It's able to accelerate the general version of Intel's XeSS upscaling technology. It's just that no-one used this on Xbox yet (remains to be seen if that will change). These inference upscaling engines are many year, very expensive technologies to develop.

And VRS has been used, it's just it's none trivial to use it well, and the performance gains can be limited if you don't want to hurt visuals. Last gen systems, PS5, and a good many still popular PC cards don't support the hardware version DX12U specifies, and some games are better off using a software implementation. For the moment, in many cases, plain old DRS is probably good enough, easier to implement, and better suited to be used pretty much universally.

Sampler feedback streaming is another promising technology, but the motivation to fully leverage it probably won't come until DS is widely adopted, as you're talking about many many tiles per frame potentially being swapped in / evicted per frame and you want to be doing that with as close to zero frames of latency as possible. DS and rapid, accurate changing of tiles of individual mip map levels is another complicated thing to incorporate into your streaming game engine. I can only imagine how difficult incorporating all this into a game engine that has to run, properly and predicatably, on all kinds of configurations is.

These features are taking a long time to incorporate into PC games too. Games running on hardware none of it supplied by MS, with most games not developed or published or touched by MS. Nvidia and now AMD and even Intel have been banging on about these features, and trying to help developers get a handle on them.

It's always frustrating to see interesting new technology and approaches take time to be used, but its the reality of things. And some times, something better comes along in the mean time.... :unsure:
 
Well they've delivered the hardware where they can (Xbox), and the API and documentation for PC, and it's supported by all the three main GPU vendors. MS's ability to force developers to use certain features that aren't high enough priority for the developer is limited. There's a huge amount of inertia in complex, long running software development projects. The technologies MS have been talking about are real though, and they do what MS have told developers they do.

DXR has found a lot of traction, and mesh shaders are going to be incorporated into the worlds most popular engine where appropriate. But you maybe don't need to target mesh shaders on cross gen games ....

SM 6.4 allows for the acceleration of mixed precision packed integer stuff - the "AI acceleration" that MS talked about for Xbox which became a standard part of PC RDNA2. It's able to accelerate the general version of Intel's XeSS upscaling technology. It's just that no-one used this on Xbox yet (remains to be seen if that will change). These inference upscaling engines are many year, very expensive technologies to develop.

And VRS has been used, it's just it's none trivial to use it well, and the performance gains can be limited if you don't want to hurt visuals. Last gen systems, PS5, and a good many still popular PC cards don't support the hardware version DX12U specifies, and some games are better off using a software implementation. For the moment, in many cases, plain old DRS is probably good enough, easier to implement, and better suited to be used pretty much universally.

Sampler feedback streaming is another promising technology, but the motivation to fully leverage it probably won't come until DS is widely adopted, as you're talking about many many tiles per frame potentially being swapped in / evicted per frame and you want to be doing that with as close to zero frames of latency as possible. DS and rapid, accurate changing of tiles of individual mip map levels is another complicated thing to incorporate into your streaming game engine. I can only imagine how difficult incorporating all this into a game engine that has to run, properly and predicatably, on all kinds of configurations is.

These features are taking a long time to incorporate into PC games too. Games running on hardware none of it supplied by MS, with most games not developed or published or touched by MS. Nvidia and now AMD and even Intel have been banging on about these features, and trying to help developers get a handle on them.

It's always frustrating to see interesting new technology and approaches take time to be used, but its the reality of things. And some times, something better comes along in the mean time.... :unsure:
For sure. However, I'm more referring to their own studios. I'm not expecting tons of 3rd party developers to adopt all of these things. Many of them have their own technologies which work for them and that's great. But MS owns a lot of studios, and it's been years and years now and we still aren't really seeing anything pushing these technologies in any way that's obvious coming from them. I guess I'm speaking more for console owners here than anything.. but if I was a console Xbox fan.. I'd be somewhat disappointed that these features used to market the console being so advanced aren't being demonstrated very well, or at all. I look at Sony and they took a more reserved approach with how they talked up the PS5, and basically were all too happy to let the games speak for themselves.

I think some of it has to do with Xbox's studio management maybe not being in a great place. Their teams don't have the greatest output and often miss schedules. To be fair lot of them were going through transition periods by becoming part of Microsoft, and of course covid perhaps set them back more than other studios.. but still things are moving slow. I know MS is a big company and things can take a long time to happen. Hellblade 2 looks phenomenal though, and I know The Coalition's next games will look astonishing.. so I'm not saying their teams don't have awesome stuff coming.. but we're already a couple of years in and there's not much to show for it utilizing the "Velocity Architecture" for the Series consoles yet.. and all the various DX12U features. VRS has been used, but unfortunately some of the implementations have been less than flattering.

You're not wrong though. Perhaps getting the PC up to feature parity has been holding the Xbox side back and is part of the reason for the lack of products featuring all this stuff yet.. Seems like a very likely possibility. That's one of the things to keep in mind about all of this. Microsoft has a lot more to consider than Sony does when they're building their consoles and development tools, which makes everything less efficient for them than it could be.
 
Status
Not open for further replies.
Back
Top