Xbox Wire: Xbox Series X|S are the only next gen consoles to have full RDNA2

That tweet is wrong. They are misinterpreting an article in The Verge, which is misinterpreting the AMD event today. DLSS is Deep Learning Super Sampling. It requires machine learning inference capabilities (low level integer calculations) in order to be hardware accelerated and actually work. PS5 lacks any ML hardware, as confirmed by Rosario Leonardi (Sony’s Principle Graphics Engineer working on PS5) already.

6 Likes

SFS is also a feature.

They mentioned it in there presentation today and it also requires a dedicated hardware block.

1 Like

im pretty sure sampler feedback and sampler feedback streaming are two different things

1 Like

It’s a feature of DX12U. Not RDNA2.

3 Likes

Are they? Really?

1 Like

Correct, but SF itself is not an RDNA2 feature either. There is a difference between RDNA2 feature set (RT, VRS, perf/watt) and DX12U feature set, which has other stuff on top of those.

2 Likes

Seriously. I didn’t know. Thought both are the same things. So its a custom feature like PS5 i/o system.

Good to know.

yeah i think sfs is more on ram optimization. so less ram usage to render a scene. I don’t have the correct background to give you an accurate answer as two the differences between them. but xsx has both and Sampler feedback streaming is engineered by xbox

SF is hardware access that lets devs get feedback about what textures are rendered to pixels on screen, which mip levels are those from and where on the mip map itself are those samples from. This is not very useful on its own though, since utilizing that info to help predict what parts of textures/mips need to be there for the next frame results in lots of faults where the guess doesn’t work out.

SFS is SF+texture filtering hardware, which allows devs to avoid these faults being visually obtrusive. It lets you seamlessly transition these textures across mip levels without the texture popping effect all over the screen. When you make SF useful with this kinda hardware, it lets you confidently stream in only the parts of a texture ya need, knowing any faults get smoothed over and corrected quickly on the fly by the SFS hardware. That’s where the 2.5x boost to I/O and RAM utilization stems from.

2 Likes

PS5 IO system = XBOX Velocity Architecture. Not the same. Other feature sets part of RDNA2 are extra on XBOX part which PS5 doesn’t have.

Wow. That’s pretty cool. I can’t wait to see what games take full advantage of these features in the future.

Didn’t said equal!

But custom thing which is not AMD’s RDNA2 part.

Yes.

Sampler feedback is hardware for the code to get information about texture samples (basically which textures were used, at which level od detail and which part of them).

Sampler Feedback Streaming is also new hardware, thus far only available in SX and SS, that implements a end to end texture streaming solution using the information from sampler feedback to direct what to stream.

You can create a SF based streaming implementation in software only, but as Ms notes it increases the code complexity and also there’s some performance issues. (because you need to load, filter and store both the texture and the usage map data at once, and with a software based approach you have to coordinate that yourself and filter them separately paying twice the cost. Also with the hardware implementation Ms was able to resolve intermediate addresses without performance hit, and finally they had hardware based “page faults” that immediately trigger the correct lod when it misses)

2 Likes
4 Likes

So the raytracing could be better on XSX too? Ubi did mention hardware accelerated raytracing for WDL on XSX and didn’t mention it for PS5.

4 Likes

Good post. Just wanted to add to the very last thing ya said, the process of correcting the page faults (which normally is an abrupt texture pop that many of us are familiar seeing in games) is only able to happen smoothly and seamlessly over a few frames if ya have the texture filtering SFS provides. :slight_smile:

2 Likes

XSX has more CU units for better RT anyhow, but yes even beyond having more brute power to do the RT, having ML (not an RDNA2 feature, but is DX12U and XSX/XSS feature) allows for further denoising of RT, which lets you cast fewer rays per pixel without losing quality.

2 Likes

If you know Anime - you know this is Epic.

Hit like for Rock Lee Series X

6 Likes

Indeed, there is hardware for that too, for smoothing out the transition between the incorrect and correct mip levels

(and it’s indeed super smooth, seeing that SFS demo on SS at 25% speed makes it easier to spot when the transitions happen)

3 Likes

Interesting Xbox waited so long to get the full feature set. That explains why they have less games ready at launch and why John Linneman heard good stuff from developers about the performance of the ps5. Xbox was probably late and thats why early titles might perform the same or better on ps5 early on. I expect this to change dramatically over time in favour of Xbox. In the short term Xbox has less games ready for launch but over the course of a 7 year generation this should come out as a great decision!

3 Likes