Yeah, also devs consistently put much more effortns into PS ports, we are seeing it for almost every game.
ML SR can be a 4x boost to perf in GPU-bound games, CBR is only a 2X.
1 Like
It’s probably less than that for both because the reconstruction itself takes some processing.
There are probably some benchs regarding DLSS on Pc, but for CB we got that slide from Dice in ME Andromeda:
4 Likes
I think this has more to do with the maturity of the tools than anything else.
Not quite accurate, but in this scenario, why wouldn’t a developer just opt for a lower resolution instead of halving the frame rate? As long as the game isn’t a jaggie or shimmery mess on a screen, most people wouldn’t know nor care about a lower resolution but a half frame rate would definitely gather attention.
1 Like
I think the res you’d need is lower than you assume it’d be. Let’s say a game on XSX renders internally at 1080p, then uses ML to hit 4k at 60fps. The same game on PS5 would presumably need to have an internal res of 720p or so, then try CBR to get to 4k at 60fps. That would not look very good on screen. Could they turn other stuff way down to help instead of destroying IQ that badly? Maybe, sure. It probably depends on the type of game it is whether they go with fps at the expense of IQ or vice versa.
I am expecting there to just be different display modes on XSX tho. For instance, say a game on PS5 has a RT fidelity mode and a 60fps perf mode sans RT. On XSX it might just have a single RT perf mode that is best of both worlds and so you’d have the 60/30 fps gap there in a like for like comparison.
1 Like
That was still real early in CBR’s use though. I think it has seen some gains in recent yrs. That game was a ‘launch’ window game for CBR tech in the Pro iirc.
This is a real good way to show off VRS and its effect on IQ both locally and globally in the frame: Current Generation Games Analysis Technical Discussion [2020-2021] [XBSX|S, PS5, PC] | Page 100 | Beyond3D Forum
This is just a mock up image from the poster there I think, but gets the point across well I think:
1 Like
Here’s the first pass at a public write-up at the following B3D thread for those technically inclined. Constructive and open feedback is welcome in the thread.
2 Likes
Hopefully as Iroboto improves the tools, they will become more widely used with full disclosure over the entire run. I also hope it will put an end to the cherry picked scenarios and screenshots going on that are used to suit unscrupulous reviewers narratives.
1 Like
Are you a scientist?
It went all over my head😅
I’m not the poster in question, but do talk and game with them.
However, I worked at a real time machine vision inspection company back in the 90s for a few years, so some of the concepts I was familiar with but have long since forgotten. I need a refresher course before I could provide meaningful insight.
The main takeaway is objective image analysis instead of subjective. Or at least more analysis to assist the technical reviewers.
2 Likes
I figured any tool of that nature would need to rely on FT’s. I do wonder if there might be an issue though, namely that in his methodology he has to rely on greyscale images, which add another layer of processing to pixel values and can potentially mask deviations between neighboring pixels.
Because a low enough resolution can look like a jagged/blurry mess depending on the reconstruction tech and the art style of the game.
We have some games like Avengers where despite the higher resolution, SX performs better than Ps5. Imagine if SX was able to render 4 times less pixels with no noticeable drop in image quality due ML upscaling. And that’s a game that already have a messy IQ on Ps5.
With such a lower baseline resolution devs would be able to push SX visuals way further…
I think that would put developers at a very tough spot, because since Ps5 IQ is already bad, lowering it even more will make it only worse, and on the other hand, if they don’t there’s no way they will be able to match SX visuals.
But I’m not sure that’s a feasible scenario, because they likely opted for CB because it was already implemented on Pro. For a new game they would most likely use a far better reconstruction tech that would allow them to drop the resolution perhaps with even better IQ than what they got from CB.
Could be, but it’s interesting nonetheless because they were using all of Pro features already (the index buffer hardware, fp16)… Though there are worse implementations than that one, so perhaps you can let quality slip even more for higher savings?
There was some footage straight from Coalition themselves showing VRS on Gears 5. I don’t remember where they showed but the effect seemed way more granular, with smaller pixels being coarsed.
However, that’s still VRS being applied only to the final buffer. As Ms already noted in their researches, performance gains are much higher for intermediate buffers, and the effect is even harder to notice (for example, shadows, lighting buffers, they mentioned even for RT early tests show it’s a huge performance boost when coupled with sampler feedback because they can tell exactly how big a surface occupies on the screen and the direction it’s facing to adjust the amount of rays).
1 Like
Wrt VRS, the image isn’t showing what you think it is. It is showing tiles of some size (unrelated to VRS granularity) where the resolution in that tile is compared between the consoles. Green tiles show areas where XSX has higher res in that region and blue for PS5. The blue tiles have VRS somewhere within them, at least enough to shift the average resolution in that tile towards the blue result.
What is great about that image is it is a nice, clear cut an objective way to measure resolution as a coarse grained mapping of pixel density instead of just a single number that fluctuates both in time (DRS) an in screen space when VRS is used. FAR more informative to folks wanting to see how pixel density stacks up than just simple overall metrics.
1 Like
If I recall correctly, the analysis is broken up into pixel clusters of 16x9 to keep the display ratio.
They appear to be squares…?
That’s just how all the smaller clusters worked out assembled up.
I realize that much of the discussion has been technical, but what I’m most intrigued by is player perception of all these techniques.
We were all floored by a 1440p TSAA upscaled to 4K 30 fps UE5 demo on the PS5. People are happy with a 1080p up-rezed to 1440p and then CBRed to 4K with Returnal.
A native 4K Halo Infinite was panned because the fidelity wasn’t there.
I’m sure most could see the difference between native and upscaled 4K, and could even appreciate 6K if they saw it. But when it comes to making a good impression, resolution seems to matter less than fidelity and framerate.
I’d like to see actual polling research on this with data split up based on demographics (age being an interesting one, PC vs console being an obvious one).
I bring this up because I’m thinking Sony made a good call with hardware based CBR, which is essentially free, and that MS really needs ML upscaling to actually work with minimal impact to render times. Visual fidelity seems to be the thing to get people’s attention and that things like upscaling and optimizations like VRS are needed to get these running at “acceptable” resolutions and target framerates.
The other thing I want to bring up is that we are talking about a delta between both machines that will increase when games start to take advantage of DX12U. But what throws a wrench in all of this is Unreal Engine 5 because their tech is all done in software. All the great things accomplished by mesh shaders, sampler feedback, and raytracing hardware is largely accomplished through their software solutions. Even Project Acoustics, something that Microsoft has hardware for in this series consoles, has a similar software solution in Unreal Engine 5. From the sounds of it they will not be making use of hardware features like mesh shaders in new graphics cards and the Xbox consoles because they are so confident in their software solution. Microsoft can take a fork of the engine and optimize it themselves, if they can find optimizations. But this is something that throws a wrench in Xbox’s exclusive hardware feature set.
This last point is why I’m taking a vested interest in The Coalition’s presentation. If they have found some DX12U optimizations in the little time that they’ve spent with UE5 (which I’m guessing is a year or so at this point), it bodes well for these exclusive hardware features.
2 Likes
Even CryEngine has a software solution for RT but nothing beats the hardware accelerated ones.
1 Like
So he is using larger squares and then breaking those down into 16x9 pixel clusters…but then what? Why don’t we see any 16x9 variation in the mapping? o.0
Alright, sorry about earlier, I crossed up discussions. I had a review of the discord discussions and here’s the details:
Right now there are 144 blocks, so take 4K and segment appropriately. If I did that right, it would be a 12x12 grid of 320x180 pixel clusters. The ratio still works out to 16:9 – 1.77777. The blocks can be made smaller for more fine-grained analysis. The images are just for visualizations, and there are a lot of options for metrics like an overall one for the full image or per block etc.
2 Likes