Current Generation Hardware Speculation with a Technical Spin [post GDC 2020] [XBSX, PS5]

Status
Not open for further replies.
Can we assume that the cost/quality is the visible transitioning of texture mips that Microsoft's hardware texture filters seek to fix? Maybe the performance cost of doing the filtering through shaders was too high so Microsoft chose to engineer a hardware solution?
The bulk of the benefits come from Sampler Feedback which all DX12U cards support, and PS5 should also (outside chance it doesn't).

Sampler Feedback Streaming is MS looking at it and thinking with some minor (silicon) changes and extension to api can make the feature bit better.
How much it improves things is unknown in terms of slight additional performance and image quality with new filtering.
SFS might even make usage of SF easier for devs, but they would still need to implement it on DX12U anyway.
 
The bulk of the benefits come from Sampler Feedback which all DX12U cards support, and PS5 should also (outside chance it doesn't).

How much it improves things is unknown in terms of slight additional performance and image quality with new filtering.
SFS might even make usage of SF easier for devs, but they would still need to implement it on DX12U anyway.

You're saying SFS is only a bit better and a developer is saying it's a game changer. I'm betting it's the latter and why MS are shouting to the moon about the Velocity Architecture.

The truth is that the PS5 has better brute force IO and the XSX has more efficient IO for textures, which comprise the majority of data. There will be situations where each outperform the other, but neither has absolutely better IO than the other IMO, but we'll find out in a couple years.
 
You're saying SFS is only a bit better and a developer is saying it's a game changer. I'm betting it's the latter and why MS are shouting to the moon about the Velocity Architecture.
SFS is superset, it includes SF, so yes it's a big difference compared to not having SF like current gen.
He didn't say SFS is game changer compared to SF.
I also never downplayed SFS benefits, just highlighting the different parts of the technology as people seem to not understand the different aspects.
The truth is that the PS5 has better brute force IO and the XSX has more efficient IO for textures, which comprise the majority of data. There will be situations where each outperform the other, but neither has absolutely better IO than the other IMO, but we'll find out in a couple years.
We don't know how true that is at all yet.
Not if PS5 has SF, which I expect it to.
Although as I said there's an outside chance it doesn't.
 
The truth is that the PS5 has better brute force IO and the XSX has more efficient IO for textures, which comprise the majority of data. There will be situations where each outperform the other, but neither has absolutely better IO than the other IMO, but we'll find out in a couple years.

There is a lot of unknowns regarding the capability between the two systems, I think it is a stretch to say that the XSX has more efficient texture IO to any significant degree without knowing the entire story.
 
There is a lot of unknowns regarding the capability between the two systems, I think it is a stretch to say that the XSX has more efficient texture IO to any significant degree without knowing the entire story.
Indeed.

We will know more in few years when proper mutiplatform tech talks emerge.
 
Well, as far as I understood, SFS is a patented solution like VRS. So it might be on board of the chip, but sony can't use ist (... I hate patents if you want to exclude someone of using it :( ).
No, SFS hardware changes and software was developed by Microsoft. If Sony has something similar than they haven't talked about it.
 
There is a lot of unknowns regarding the capability between the two systems, I think it is a stretch to say that the XSX has more efficient texture IO to any significant degree without knowing the entire story.
SFS is also predicated on developer implementation, and given that it's platform specific and that devs haven't really made much use of tiled resources over the years, it's not really a sure-fire win across the board there. o_O
 
SFS is also predicated on developer implementation, and given that it's platform specific and that devs haven't really made much use of tiled resources over the years, it's not really a sure-fire win across the board there. o_O
Well, tiled resources and HDD didn't scale that well ;)
With SSDs this changes. And SF is more an evolution of Tiled resources (more or less done right). With SFS they also implemented filtering between the mips (as far as I understood) in hardware. So it needs less steps and less roundtrips.

Btw, some kind of tiled resources should also be used for PS5 (don't know the name for this feature on sonys side). So for everything that is next-gen this should be the way to use. But well, I guess sonys SSD is more the brute force of IO bandwidth does not really rely on that feature. But overall I would guess, even the 2.4 gb/s from xbox should be more than sufficient for the good old texture streaming. In the end, the GPU must somehow process it. It is just a question how big the buffer in memory is.

After all, something like SF/SFS not only saves IO bandwidth, but also memory and memory bandwidth. So in the end it is a welcome feature.

About loading times ... couldn't care less if it is 2s or 4s as long as we are under 30s it is fine for me.
 
tiled resources should work with SFS.. lol.

haha. It's taken a long time to get to this stage for all those small parts to add up to something useful.

In many ways, they need to drop off the PC market that isn't DX12U to get good progression, but I don't think devs are willing to.
 
You're saying SFS is only a bit better and a developer is saying it's a game changer. I'm betting it's the latter and why MS are shouting to the moon about the Velocity Architecture.

The truth is that the PS5 has better brute force IO and the XSX has more efficient IO for textures, which comprise the majority of data. There will be situations where each outperform the other, but neither has absolutely better IO than the other IMO, but we'll find out in a couple years.

My question about what you say is. Can SFS fetch compressed textures? Or textures need to be uncompressed for it to work? I ask this because SFS is a GPU feature, and not a part of the I/O system.
 
My question about what you say is. Can SFS fetch compressed textures? Or textures need to be uncompressed for it to work? I ask this because SFS is a GPU feature, and not a part of the I/O system.
The problem with compression is, that you most times have to decompress a good part of the data to get just a bit of it. So at best you compress each tile individual.
Just think of a file system (well don't think we have something that can be called like this) just to get the idea:
Instead of texture1.zip you have texture1.1.zip to texture1.112.zip on the SSD and some kind of map-information (just basic metadata) to know which texture part is where, and you only load those you need.
The loaded tiles, than get decompressed (just the data-compression) and written into the memory.
Texture-compression is still active.

This is a bit oversimplified but I think you get the idea.
 
tiled resources should work with SFS.. lol.

haha. It's taken a long time to get to this stage for all those small parts to add up to something useful.

In many ways, they need to drop off the PC market that isn't DX12U to get good progression, but I don't think devs are willing to.
Yea, can't remember the details but the tier of tiled resources on consoles especially had limitations that meant for a lot of studies was worth rolling their own streaming implementations.

Hopefully everything is good this time around. But it does go to show that you never no if a feature is a home run, especially for version 1.
 
Well, tiled resources and HDD didn't scale that well ;)
With SSDs this changes. And SF is more an evolution of Tiled resources (more or less done right). With SFS they also implemented filtering between the mips (as far as I understood) in hardware. So it needs less steps and less roundtrips.

Btw, some kind of tiled resources should also be used for PS5 (don't know the name for this feature on sonys side). So for everything that is next-gen this should be the way to use. But well, I guess sonys SSD is more the brute force of IO bandwidth does not really rely on that feature. But overall I would guess, even the 2.4 gb/s from xbox should be more than sufficient for the good old texture streaming. In the end, the GPU must somehow process it. It is just a question how big the buffer in memory is.

After all, something like SF/SFS not only saves IO bandwidth, but also memory and memory bandwidth. So in the end it is a welcome feature.

About loading times ... couldn't care less if it is 2s or 4s as long as we are under 30s it is fine for me.

But the PS5 SSD is still faster right?
 
Yea, can't remember the details but the tier of tiled resources on consoles especially had limitations that meant for a lot of studies was worth rolling their own streaming implementations.

Hopefully everything is good this time around. But it does go to show that you never no if a feature is a home run, especially for version 1.
IIRC the tile sizes were 64K for tiled resources. And for some engines this worked out well, and for others not so much. Most opted to run their own version because the speed up wasn't worth the bandwidth trade off if I recall that part correctly. (most wanted smaller, 64K was too large)

Tiled Resources is now supported up to tier 4 as of writing IIRC.
Tier 1 was 2d textures.
Tier 2 was special clamped on 2D
Tier 3 was 3D textures or volume tiled resources
Tier 4 ??? no clue wtf this is.

For Sampler Feedback:
Specifically D3D12_TILED_RESOURCES_TIER_2 is recommended as a dependency for a workable streaming implementation; hardware tier 2 covers the ability to control tile residency, also the ability to straightforwardly clamp to resident mip levels when sampling.

^^^
So I guess it depends once again on whether the developers want to support it. For the Gears of War team, it's already supported since Gears 4 I think? So SF and SFS is the next step for them as TR is already covered.

Best documentation on SFS is here:
https://microsoft.github.io/DirectX-Specs/d3d/SamplerFeedback.html
 
Last edited:
But the PS5 SSD is still faster right?
Should be. But we just don't know how effective the tiled resources algorithms from sony are. But RAW bandwidth is normally always better than optimized, because the raw-bandwidth is always available.
But as I stated, I don't believe the SSD will make much of a difference as both are really fast. There just may be some edge cases where the higher raw bandwidth might be better. But on the other hand there may be edge-cases where SF is much better (because it generally reduces memory "pressure"). But as sony might just have the same implementation (with the final S ^^ which only seems to save some roundtrips and not IO bandwidth) more is always better for edge cases.


One thing is for sure, the XSX/S can be repaired quite easy if the SSD is broken, as it is just a module to replace. I don't think the SSD of the PS5 will be on a replaceable module.
 
The problem with compression is, that you most times have to decompress a good part of the data to get just a bit of it. So at best you compress each tile individual.
Just think of a file system (well don't think we have something that can be called like this) just to get the idea:
Instead of texture1.zip you have texture1.1.zip to texture1.112.zip on the SSD and some kind of map-information (just basic metadata) to know which texture part is where, and you only load those you need.
The loaded tiles, than get decompressed (just the data-compression) and written into the memory.
Texture-compression is still active.

This is a bit oversimplified but I think you get the idea.

Yes... Thought about that. But it seems to me that can be done with sample feedback or even PRT. Where is the inovation then?
 
IIRC the tile sizes were 64K for tiled resources. And for some engines this worked out well, and for others not so much. Most opted to run their own version because the speed up wasn't worth the bandwidth trade off if I recall that part correctly.
Thanks, I couldn't remember the details, hopefully the spec allows for more sizes that will cover more use cases.
One thing is for sure, the XSX/S can be repaired quite easy if the SSD is broken, as it is just a module to replace. I don't think the SSD of the PS5 will be on a replaceable module.
have to admit i assumed when you upgrade the ssd on PS5 that you swap out the current one, not have an additional connector. But that does make sense.

Anyway we still haven't seen teardown to see if the PS5 is a simple module to upgrade.
 
Status
Not open for further replies.
Back
Top