General Next Generation Rumors and Discussions [Post GDC 2020]

[Shifty Geezer]

Umm, no they really can't ? Custom D3D on Xbox One exposes raw GCN2 ISA much like GNM does on PS4. Aside from maybe a couple of registers, they're largely the same GPU architecture. On consoles they can bypass the entire shader compiler.

That's the first I've heard of that. So MS will face similar problems in future BC for some titles.

As for why games aren't able to take automatic advantage of boost clocks, that's likely down to their reliance on timing specific behaviour on previous hardware.

Which is really bad practice these days! I hope devs drop that. the efficiency gains aren't worth it in the long run. A game should never break because a workload can be completed faster.

But that said, that's not a problem for a BC architecture that can clock lower for those titles. MS's abstraction allows them to run XB1 games on a GPU with a completely different CU count, where Sony's apparently does not.

 

[Jay]

Logically: if games were to be designed around PS5's hard drive speed, it wouldn't make a lot of sense for PSNow. You couldn't stream such high fidelity. But you're going to lock all of your servers to building out these custom SSDs that the performance will never be seen.

not sure if I'm understanding you correctly here.
Are you saying psnow won't be able to match the performance of ps5?
That could literally break games.

 

[Shifty Geezer]

Sure, but there are upper limits to that. Because your end state is still to interact with RAM. Whatever is being rendered on screen has to be resident in RAM. Unless you're suggesting we can stream off SSDs now and by-pass memory altogether.

Yes there's an upper limit, but potentially, depending on seek times, a scene can be cached early enough in advance for a game to be fully streamed.

If a game requires 10 GBs of data to be resident for the GPU to draw a frame, it won't be possible. But if a game needs 2 GBs to draw a frame, and that'll change no more than a few MBs per frame as the player moves, that data could be streamed with several GBs of cached data in RAM. It really depends on the engines and how much they cram into a frame. If they go full-one virtual texturing, the SSDs should be completely adequate for fully streamed textures.

 

[iroboto]

not sure if I'm understanding you correctly here.
Are you saying psnow won't be able to match the performance of ps5?
That could literally break games.

if you design a game for PS5 around streaming assets at 5.5GB/s, that technology doesn't exist server side. So you need to build that in the servers too.

 

[iroboto]

Yes there's an upper limit, but potentially, depending on seek times, a scene can be cached early enough in advance for a game to be fully streamed.

If a game requires 10 GBs of data to be resident for the GPU to draw a frame, it won't be possible. But if a game needs 2 GBs to draw a frame, and that'll change no more than a few MBs per frame as the player moves, that data could be streamed with several GBs of cached data in RAM. It really depends on the engines and how much they cram into a frame. If they go full-one virtual texturing, the SSDs should be completely adequate for fully streamed textures.

Right, so take advantage of PS5s speed over our current NVME; you'd have to find an extreme situation in which they would need to have this scenario where they are moving so quickly that they can't stream textures. But with such high fidelity that you can see the difference while in motion all the while, being able to render them in crisp full clarity that someone could appreciate it.

Or if we ignore what I wrote:
It seems like the basis for our discussion is how much can be streamed via virtual texturing while completely bypassing main memory. In which having 2x the amount of bandwidth would be more useful than having 1/2 the amount if we are looking strictly at hardware capabilities.

Because in all other scenarios we are moving stuff into memory anyway, and the only difference between their performance will be the amount of memory available.

 

[MrFox]

Right, so take advantage of PS5s speed over our current NVME; you'd have to find an extreme situation in which they would need to have this scenario where they are moving so quickly that they can't stream textures. But with such high fidelity that you can see the difference while in motion all the while, being able to render them in crisp full clarity that someone could appreciate it.

It will just be more fuzzy.

"turning" in a game is not an extreme situation.

 

[iroboto]

It will just be more fuzzy.

"turning" in a game is not an extreme situation.

You need high frame rate (and display refresh rate) to clear up motion blurriness from turning in combination.

 

[AbsoluteBeginner]

I have heard some fascinating things about the PS5’s operating system like this - one of the pitches they’ve been making to developers is “playing a PS5 game should be as easy as Netflix.” They want to make players feel like they can load up the game immediately and know exactly how much time a given activity is going to take them. They want people to feel more inclined to play in short bursts rather than only wanting to turn on the console when they have a few hours to spare.

From Jason Schreir.

Is this fascinating enough? Instantly loading the game? I am wondering if anyone will care between loading in 1s or 2s. This is why I am itching for Sony to show what is only possible on 5GB/s SSD gameplay wise, because 99% of the talk has been related to load times.

And I know devs will be biggest proponents of fast SSD because it makes their life easier, but will gamers see so much difference on their end?

 

[Jay]

if you design a game for PS5 around streaming assets at 5.5GB/s, that technology doesn't exist server side. So you need to build that in the servers too.

Then what would devs be told?
Develop your games to 2GB/s?

psnow will be upgraded to support the required spec.

 

[iroboto]

Then what would devs be told?
Develop your games to 2GB/s?

psnow will be upgraded to support the required spec.

I don't know. I don't have the answer for that.

 

[manux]

Then what would devs be told?
Develop your games to 2GB/s?

psnow will be upgraded to support the required spec.

Develop games to 100MB/s due to pc peasants with mechanical drives :/ or 500MB/s due to sata ssd's. It's probably platform exclusives that stretch nvme muscles. Other games it's probably some additional things(better textures, less popup, wider variety of models) but not baked fundamentally into engine as requirement to run the game.

 

[Shifty Geezer]

Right, so take advantage of PS5s speed over our current NVME; you'd have to find an extreme situation in which they would need to have this scenario where they are moving so quickly that they can't stream textures. But with such high fidelity that you can see the difference while in motion all the while, being able to render them in crisp full clarity that someone could appreciate it.

I'm not arguing a PS5 advantage. You said SSD was designed for speeding up loading and you can't stream as SSD isn't fast enough, RAM still needs to be populated. I'm saying both consoles are designed for streaming and streaming should be very effective. I'm not arguing that double the streaming speed will amount to any game advantage (I've already debated otherwise!).

 

[iroboto]

Oh right.
I agree, but I actually agree with you that this is what we should be talking about with respect to hard drive performance.
I think for the sake of clarity; I'm going to drop the 5.5GB/s and 2.4 GB/s talk. Its too high in the sky. I get lost in thinking about it to be honest.

The raw throughput numbers for PS5 are
88 MB @ 60 fps
176 MB @ 30 fps

XSX:
38.4 MB @ 60 fps
76.8 MB @ 30 fps

vs current generation (120MB/s)
1.92 MB @ 60fps
3.84 MB @ 30fps

Its about what they can do with those numbers. I did not account for compression. I did not account for any technologies that reduce the amount of data that needs to be streamed.

If the virtual texturing exceeds both values, buffering is required. If it's below those values, you may be able to go direct from SSD to GPU in theory.

 

[AbsoluteBeginner]

Its funny that no one is actually talking much about biggest difference for gamers next gen : 8 core / 16 thread Zen 2's.

I think SSDs we are getting will be amazing jump forward, and I am sure devs will appreciate it greatly, but It think this gen was mostly held back CPUs, that kept game design on 05' level.

More then high quality texture streaming, I think we need actual advance in interactivity, ai and physics of games.

 

[BRiT]

And 4K display is what, 24 MB with 24bit or 32 MB with 32bit pixels [alpha channel]?

3840 * 2160 = 8,294,400 pixels
8,294,400 pixels * 3 bytes = 24,883,200 ~ 24 Meg
8,294,400 pixels * 4 bytes = 33,177,600 ~ 32 Meg

 

[Shifty Geezer]

Its about what they can do with those numbers. I did not account for compression. I did not account for any technologies that reduce the amount of data that needs to be streamed.

If the virtual texturing exceeds both values, buffering is required. If it's below those values, you may be able to go direct from SSD to GPU in theory.

Did you see my back-of-a-napkin maths from Sebbbi's numbers? That was 7 MB a frame needed for perfect texturing at 720p. 4K is 9x the res, so you'd need 63 MB/frame. At half that drive speed, twice the caching should be enough(?). But you'd never go SSD > GPU because you have to cache tiles for access latency and performance; those tiles are read multiple times over multiple frames before being swapped out. As you said, SSDs are far, far, far too slow to replace RAM as working memory.

 

[iroboto]

Did you see my back-of-a-napkin maths from Sebbbi's numbers? That was 7 MB a frame needed for perfect texturing at 720p. 4K is 9x the res, so you'd need 63 MB/frame. At half that drive speed, twice the caching should be enough(?). But you'd never go SSD > GPU because you have to cache tiles for access latency and performance; those tiles are read multiple times over multiple frames before being swapped out. As you said, SSDs are far, far, far too slow to replace RAM as working memory.

lol okay thanks. I thought I was going crazy, thanks for adding some additional numbers there.

 

[iroboto]

And 4K display is what, 24 MB with 24bit or 32 MB with 32bit pixels [alpha channel]?

3840 * 2160 = 8,294,400 pixels
8,294,400 pixels * 3 bytes = 24,883,200 ~ 24 Meg
8,294,400 pixels * 4 bytes = 33,177,600 ~ 32 Meg

Doom has up to 50 render targets. But not all of them at 4K resolution necessarily.

 

[BRiT]

Doom has up to 50 render targets. But not all of them at 4K resolution necessarily.

Right, but in the end you're combining down to display just 8,294,400 pixels.

 

[iroboto]

Right, but in the end you're combining down to display just 8,294,400 pixels.

Yes, for a final output yea. The residency in memory may be more than 50x the size of that though for a single frame