Xbox Series S RAM options

[Seanspeed]

PS5 will no doubt be capable of a similar 'memory multiplier' trick. They didn't put such a big priority on their I/O just to boost load speeds a fraction better. They've got the exact same issue with lack of a RAM jump from the last generation that they need to figure out a way around.

How they do it may differ, but it wouldn't shock me to learn that they've managed an even more efficient way to push memory usage thanks to their incredibly fast storage solution.

But yea, it's gonna depend on the title, of course. Not every game is gonna push things as hard, especially earlier on, but I do expect it to become a problem as time goes on.

 

[Ronaldo8]

PS5 will no doubt be capable of a similar 'memory multiplier' trick. They didn't put such a big priority on their I/O just to boost load speeds a fraction better. They've got the exact same issue with lack of a RAM jump from the last generation that they need to figure out a way around.

How they do it may differ, but it wouldn't shock me to learn that they've managed an even more efficient way to push memory usage thanks to their incredibly fast storage solution.

But yea, it's gonna depend on the title, of course. Not every game is gonna push things as hard, especially earlier on, but I do expect it to become a problem as time goes on.

If you listened to the Road to PS5 presentation, you should know that the blazing fast SSD is itself a RAM multiplier. However SFS is also a bandwidth multiplier. All modern game engines have a virtual texture streaming capability which optimises memory usage; it's just that SFS is more capable than them (2.5x on average).
Does the PS5 has sampler feedback? I seriously doubt it.

 

[peceed]

And as for just having 2GB more, I agree that would have made a difference, but it also would have required a 192-bit bus rather than the 128-bit one. Which would have knocked out both birds with one stone - capacity and bandwidth. I'd have rather seen the system at like $320-330 with these upgrades.

The downside of 192 bit bus is that it limits future die size reductions, 128 bit gives similar effective bandwidth / tflops.
Anyway, the computing power of Series S should be also increased, it has 1.3 TFLOPs for resolution independent computations, and it is way too little in order to provide comparable fidelity/gameplay (GI/AI). 192 bit/24cu/1825mhz is a much more balanced design and BOM still fits $300/$500 ratio.
The minimal satisfying and easy improvement of Series S is 2GB more and GPU clock boosted to 1825 Mhz.

 

[Seanspeed]

If you listened to the Road to PS5 presentation, you should know that the blazing fast SSD is itself a RAM multiplier. However SFS is also a bandwidth multiplier. All modern game engines have a virtual texture streaming capability which optimises memory usage; it's just that SFS is more capable than them (2.5x on average).
Does the PS5 has sampler feedback? I seriously doubt it.

I would be surprised if Sony did not think things through with a way to further leverage their impressive I/O pipeline other than simply raw capabilities.

As for whether PS5 has sampler feedback, in terms of the actual hardware feature? I really wouldn't want to guess either way. There's reasons to think it would and reasons to think it wouldn't. Beyond that Road to PS5 event, they've really given us no further deeper tech dive with the hardware, so it's hard to say.

 

[Seanspeed]

The downside of 192 bit bus is that it limits future die size reductions, 128 bit gives similar effective bandwidth / tflops.
Anyway, the computing power of Series S should be also increased, it has 1.3 TFLOPs for resolution independent computations, and it is way too little in order to provide comparable fidelity/gameplay (GI/AI). 192 bit/24cu/1825mhz is a much more balanced design and BOM still fits $300/$500 ratio.
The minimal satisfying and easy improvement of Series S is 2GB more and GPU clock boosted to 1825 Mhz.

Die size reductions might not come as we usually expect them. The high jump in cost for new processes nowadays might well start to outweigh the benefits of a smaller die/higher yields like they have in the past. Or at least it could push this option out further than we're used to. Similarly, I'm not expecting price cuts to be as forthcoming as they usually are, either. Which might be one of the reasons why the Series S is here to begin with, as a more longer term solution to those who usually wait for price drops before jumping into a new generation. We'll see, I guess.

As for more compute power, obviously more would be better if they can keep it within budget(cost/power/heat), but it wouldn't be my first area to address if I couldn't have both. And most devs that I've seen talk about things seem to emphasize the memory concerns as well.

 

[HolySmoke]

There's one thing I haven't seen mentioned much in these discussions which is the texture streaming pool size. The lower target resolution of the XSS won't need as large pools as the XSX simply on account having fewer pixels to fill.

Doom Eternal has already given us a taste of what an SSD brings to the table where the smallest pool size basically looks identical to the largest and especially at lower resolutions. It adds significantly to the disk load (I saw ~2x the amount of reads) but it also saves ~3.5GB of video memory according to RenderDoc.

You can see the statistics here and here but note that these are both at the same rendering resolution. The XSS should see further savings on account of having smaller render targets etc.

 

[Deleted member 13524]

And as for just having 2GB more, I agree that would have made a difference, but it also would have required a 192-bit bus rather than the 128-bit one. Which would have knocked out both birds with one stone - capacity and bandwidth. I'd have rather seen the system at like $320-330 with these upgrades.

They could have 2 channels with 2GB + 2 channels with 4GB, instead of what they have now (3 channels with 2GB + 1 channel with 4GB).
It would also increase the bandwidth of the "slow memory pool" by twofold. I.e. it would have 8GB @ 224GB/s + 4GB @ 112GB/s.

If you listened to the Road to PS5 presentation, you should know that the blazing fast SSD is itself a RAM multiplier. However SFS is also a bandwidth multiplier. All modern game engines have a virtual texture streaming capability which optimises memory usage; it's just that SFS is more capable than them (2.5x on average).

You make it sound like sampler feedback is a magic sauce that multiplies effective bandwidth by 2.5x. It's a method to provide higher granularity in texture shading. AFAICS this would be revolutionary if engines+developers hadn't been using MipMaps for decades. There's probably performance gains from using sampler feedback, but definitely not in the order of the suggested "2.5x gains" as if games were always using the full texture size at all times.

Faster storage allows for caching less assets of the game in the RAM at any given time because the on-demand transfer of assets is faster.
If you needed 30 minutes to bring potatoes from your storage room to your kitchen then you'd need to keep a bunch of potatoes in your kitchen all the time or you'd spend too much time cooking.
If it takes 10 seconds then you don't need to keep any potatoes in the kitchen, meaning you have more kitchen space for appliances that make your cooking faster.
So the use of fast storage allows for a more efficient+effective use of the RAM, but it doesn't mean that devs can cache more assets per-GB in the PS5/Series than they could in e.g. in the PS4 Pro / OneX (e.g. color delta compression hasn't changed since Polaris IIRC).

 

[Ronaldo8]

They could have 2 channels with 2GB + 2 channels with 4GB, instead of what they have now (3 channels with 2GB + 1 channel with 4GB).
It would also increase the bandwidth of the "slow memory pool" by twofold. I.e. it would have 8GB @ 224GB/s + 4GB @ 112GB/s.




You make it sound like sampler feedback is a magic sauce that multiplies effective bandwidth by 2.5x. It's a method to provide higher granularity in texture shading. AFAICS this would be revolutionary if engines+developers hadn't been using MipMaps for decades. There's probably performance gains from using sampler feedback, but definitely not in the order of the suggested "2.5x gains" as if games were always using the full texture size at all times.

Faster storage allows for caching less assets of the game in the RAM at any given time because the on-demand transfer of assets is faster.
If you needed 30 minutes to bring potatoes from your storage room to your kitchen then you'd need to keep a bunch of potatoes in your kitchen all the time or you'd spend too much time cooking.
If it takes 10 seconds then you don't need to keep any potatoes in the kitchen, meaning you have more kitchen space for appliances that make your cooking faster.
So the use of fast storage allows for a more efficient+effective use of the RAM, but it doesn't mean that devs can cache more assets per-GB in the PS5/Series than they could in e.g. in the PS4 Pro / OneX (e.g. color delta compression hasn't changed since Polaris IIRC).

All modern texture (Gen 9) streaming engines does exactly that, i.e Granite and UE5 where offline calculations are done to determine visibility and determine the required mip in the mip chain. Let's now go back to said presentation:

At the 8:20 time mark, the guy doing the presentation clearly says: "We can estimate what an optimised generation 9 traditional texture streaming engine will need to have in memory". He then goes on to show said 2.5x advantage over those texture streaming engines due to absolute precision in determining visibility and the large gain in granularity. No one should be surprised that granular certainty is going to outperform coarser guesswork.

 

[Deleted member 13524]

At the 8:20 time mark, the guy doing the presentation clearly says: "We can estimate what an optimised generation 9 traditional texture streaming engine will need to have in memory". He then goes on to show said 2.5x advantage over those texture streaming engines due to absolute precision in determining visibility and the large gain in granularity.

What is an optimized Gen9 traditional texture streaming engine? I find the use of the word "traditional" to be very odd here.
The Gen9 consoles released 6 months ago and IIRC so far there's only one game (The Medium) for the Series consoles that isn't cross-platform with Gen8 consoles. If anything there are some early Gen9 texture streaming engines for the Series consoles that don't make use of SFS, but at best it's on games that aren't out.
There's not enough of a repetition of events that would warrant the use of the word "traditional" on the PS5, let alone the Series consoles.

It seems they're claiming a 2.5x RAM occupancy advantage over what is actually a Gen8 texture streaming engine adapted to the I/O speeds of the Series consoles. That sounds like a hypothetical comparison point, as we don't really know how devs will be adapting their streaming engines to the new I/O speeds in the long-term.

No one should be surprised that granular certainty is going to outperform coarser guesswork.

No one is. I'm only advising caution on the expectations created by buzzwords being thrown like "2.5x better".
Don't expect the Series X to behave like a 40GB GDDR6 console, or the Series S like a 25GB one, because of sampler feedback.

 

[iroboto]

What is an optimized Gen9 traditional texture streaming engine? I find the use of the word "traditional" to be very odd here.

A non SFS combined with a non Tiled Resources based type VT system. Essentially, a software derived VT system is traditional texture streaming. And that's largely what we've known up to now and it's worked out very well for us.

I think for SFS to work, you must use Tiled Resources (maybe? not necessarily however), and typically developers haven't done that period (only Gears 4/5 comes to mind). So it's going to be a while to see if developers adapt this SFS streaming. It will be up to developers to determine if the trade off is worth it. In the video posted above, it (I think) should be the bar graph in the middle if you are using a software VT system. With SFS it's the lower bar graph. Without any sort of VT, it's the upper bar graph.

 

[Ronaldo8]

A non SFS combined with a non Tiled Resources based type VT system. Essentially, a software derived VT system is traditional texture streaming. And that's largely what we've known up to now and it's worked out very well for us.

I think for SFS to work, you must use Tiled Resources (maybe? not necessarily however), and typically developers haven't done that period (only Gears 4/5 comes to mind). So it's going to be a while to see if developers adapt this SFS streaming. It will be up to developers to determine if the trade off is worth it. In the video posted above, it (I think) should be the bar graph in the middle if you are using a software VT system. With SFS it's the lower bar graph. Without any sort of VT, it's the upper bar graph.

Both Unity (Granite) and UE4 have tiled-based VT systems implemented ages ago. It is definitely not something particular to Gen 9. And what do you mean by "software derived VT system".

 

[iroboto]

Both Unity (Granite) and UE4 have tiled-based VT systems implemented ages ago. It is definitely not something particular to Gen 9. And what do you mean by "software derived VT system".

those are software derived VT systems. They don't use hardware acceleration for their VT. Those solutions can customize their tile sizes, everything is done through compute shaders, they have fixed tile sizes and figure out their ordering and recall they way they want to. SFS and Tiled Resources are hardware VT support systems, as a result the parts of the pipeline they accelerate is faster. But it may come as a detriment if the pipeline does not provide everything the developers want.

We have a very long thread on it's advantages and disadvantages here:
https://forum.beyond3d.com/posts/1757463/

at the 9:30 mark for "Hardware Tiling" you get "Anistropic Filtering" for free! (this is in relation to the "Software Tiling" approach where AF is not free)

even with these advantages, developers ignored hardware tiling.

 

[Jay]

How the fuck do you even take that from this post? Your comment is so convoluted I had to search for what the hell you were talking about.
And how is that even a remotely positive contribution to the discussion?

How did you come to your figures then?
No one thinks going to have equivalent of 40 & 25GB respectively.

 

[Ronaldo8]

What is an optimized Gen9 traditional texture streaming engine? I find the use of the word "traditional" to be very odd here.
The Gen9 consoles released 6 months ago and IIRC so far there's only one game (The Medium) for the Series consoles that isn't cross-platform with Gen8 consoles. If anything there are some early Gen9 texture streaming engines for the Series consoles that don't make use of SFS, but at best it's on games that aren't out.
There's not enough of a repetition of events that would warrant the use of the word "traditional" on the PS5, let alone the Series consoles.

It seems they're claiming a 2.5x RAM occupancy advantage over what is actually a Gen8 texture streaming engine adapted to the I/O speeds of the Series consoles. That sounds like a hypothetical comparison point, as we don't really know how devs will be adapting their streaming engines to the new I/O speeds in the long-term.



No one is. I'm only advising caution on the expectations created by buzzwords being thrown like "2.5x better".
Don't expect the Series X to behave like a 40GB GDDR6 console, or the Series S like a 25GB one, because of sampler feedback.

those are software derived VT systems. They don't use hardware acceleration for their VT. Those solutions can customize their tile sizes, everything is done through compute shaders, they have fixed tile sizes and figure out their ordering and recall they way they want to. SFS and Tiled Resources are hardware VT support systems, as a result the parts of the pipeline they accelerate is faster. But it may come as a detriment if the pipeline does not provide everything the developers want.

We have a very long thread on it's advantages and disadvantages here:
https://forum.beyond3d.com/posts/1757463/


even with these advantages, developers ignored hardware tiling.

Tiled resources is simply a way to provide an indirection layer, has been supported since 2013-14 in GPUs and is the crux around which Granite and the UE4 Streaming Virtual Texturing (both are which are tiles-based) are built. The "acceleration" is a hardware page table that facilitates tile mapping in lieu of an indirection texture.

 

[mr magoo]

They could have 2 channels with 2GB + 2 channels with 4GB, instead of what they have now (3 channels with 2GB + 1 channel with 4GB).
It would also increase the bandwidth of the "slow memory pool" by twofold. I.e. it would have 8GB @ 224GB/s + 4GB @ 112GB/s.




You make it sound like sampler feedback is a magic sauce that multiplies effective bandwidth by 2.5x. It's a method to provide higher granularity in texture shading. AFAICS this would be revolutionary if engines+developers hadn't been using MipMaps for decades. There's probably performance gains from using sampler feedback, but definitely not in the order of the suggested "2.5x gains" as if games were always using the full texture size at all times.

Faster storage allows for caching less assets of the game in the RAM at any given time because the on-demand transfer of assets is faster.
If you needed 30 minutes to bring potatoes from your storage room to your kitchen then you'd need to keep a bunch of potatoes in your kitchen all the time or you'd spend too much time cooking.
If it takes 10 seconds then you don't need to keep any potatoes in the kitchen, meaning you have more kitchen space for appliances that make your cooking faster.
So the use of fast storage allows for a more efficient+effective use of the RAM, but it doesn't mean that devs can cache more assets per-GB in the PS5/Series than they could in e.g. in the PS4 Pro / OneX (e.g. color delta compression hasn't changed since Polaris IIRC).

This was a comment from someone from msft during the presentation

"SFS results in approx 2.5x the effective IO throughput (SSD performance) and memory usage above and beyond hardware capabilities on average.
In other words, it gives a speed boost till 6GB/s (raw) / 12GB/s (compressed) on Series X."

 

[iroboto]

Tiled resources is simply a way to provide an indirection layer, has been supported since 2013-14 in GPUs and is the crux around which Granite and the UE4 Streaming Virtual Texturing (both are which are tiles-based) are built. The "acceleration" is a hardware page table that facilitates tile mapping in lieu of an indirection texture.

Its possible that outside of Gears 4/5 that tiled resources are being used for SVT for UE4. But there's no guarantee unless explicitly stated, many game engines have done the same thing without hardware TR/PRT. See the original Rage for instance. There are other titles that also employ SVT without hardware PRT.

 

[Deleted member 13524]

A non SFS combined with a non Tiled Resources based type VT system. Essentially, a software derived VT system is traditional texture streaming. And that's largely what we've known up to now and it's worked out very well for us.

I think for SFS to work, you must use Tiled Resources, and typically developers haven't done that period (only Gears 4/5 comes to mind).

Tiled Resources were available for the XBox One though. I don't know why developers didn't use that tech, but wouldn't that mean a virtual texturing system using tiled resources is in fact a Gen8 texture streaming engine?

How did you come to your figures then?
No one thinks going to have equivalent of 40 & 25GB respectively.

I directly quoted a post that stated SFS is "a RAM and bandwidth multiplier".
40GB = 2.5x 16GB on SeriesX; 25GB = 2.5x 10GB on SeriesS.

 

[iroboto]

Tiled Resources were available for the XBox One though. I don't know why developers didn't use that tech, but wouldn't that mean a virtual texturing system using tiled resources is in fact a Gen8 texture streaming engine?

Yup.
Only SFS would make it gen 9. Which has been labelled as PRT +

https://microsoft.github.io/DirectX-Specs/d3d/SamplerFeedback.html

Terminology
Use of sampler feedback with streaming is sometimes abbreviated as SFS. It is also sometimes called sparse feedback textures, or SFT, or PRT+, which stands for “partially resident textures”.

They didn't use it, because they didn't want to get locked into 64KB tiles.
https://docs.microsoft.com/en-us/windows/win32/direct3d11/how-a-tiled-resource-s-area-is-tiled
Buffer tiling
A Buffer resource is divided into 64KB tiles, with some empty space in the last tile if the size is not a multiple of 64KB.

@sebbbi found this to be extremely limiting on his SVT system.

 

[Jay]

I directly quoted a post that stated SFS is "a RAM and bandwidth multiplier".
40GB = 2.5x 16GB on SeriesX; 25GB = 2.5x 10GB on SeriesS

Yea, that's not taking into account even the OS allocation much less everything else.

Just saying no one (here) would think it equates to those sorts of figures.

They didn't use it, because they didn't want to get locked into 64KB tiles.

Don't believe everything that's exposed in DX12U was exposed before either. Whether that be due to drivers or just api I don't know.

 

[iroboto]

Yea, that's not taking into account even the OS allocation much less everything else.

Just saying no one (here) would think it equates to those sorts of figures.
Don't believe everything that's exposed in DX12U was exposed before either. Whether that be due to drivers or just api I don't know.

Forza Horizon 3 and 4 also required Tiled Resources Tier 1.

So the only games I've seen that support the feature have only been MS first party titles.

Looking at some newer UE4 titles, like Outriders and such, there is no option to turn on/off tiled resources. And if AF is free with tiled resources on, then we would expect Xbox to have the maximum AF allowable, which it doesn't. It's quite low.

So the technology is there and has always been available to use, but on PC it was locked to certain versions of windows (Win 8) and at the time made no sense for anyone to pursue. Not to mention other limitations around the technology itself. Moving into DX12U generation (and most people now on Win10), yes, I believe you are more likely to see TR/PRT supported because the baseline of consoles has all shifted upwards, perhaps even take use of SFS. But modifying VT systems do take time.

Just need to wait and see.