• Ever wanted an RSS feed of all your favorite gaming news sites? Go check out our new Gaming Headlines feed! Read more about it here.
  • We have made minor adjustments to how the search bar works on ResetEra. You can read about the changes here.

Irrotational

Prophet of Truth
Member
Oct 25, 2017
7,243
I don't really think people realize what he meant from all of that. If you go back to road to PS5 there is a chart where they were Targeting reading 2GB in .27 Seconds. Yeah. This basically can allow for some crazy stuff loading game data.
Hopefully some people on this board do, but I agree I think a lot of people don't.

The concept of a memory limit is baked into basically every game we've all ever played. The storage speed has always been too slow to load new areas or new items in real time... So every game has to load them ahead of time.

Hence every game has levels, from pong loading one set of bricks at a time, to tomb raider hiding the level load behind a zig zagged corridor.

It is genuinely a very different way of building the very concept of a game. How to do you factor in the frenetic bits and quiet bits if you don't have level loads? Where do you put score screens or even show scores? How do you give the player natural rest points or save points?
 

BreakAtmo

Member
Nov 12, 2017
12,967
Australia
This is so convincing argument I bow in awe.

Look, this is exactly like with cashiers in supermarket. Once you have more cashiers than customers, adding more doesn't improve processing speed.

No, it really isn't. There are multiple situations where you would want to swap out assets that quickly (one of the most common being when you simply want to seamlessly cut from one location to another in gameplay or a real-time cutscene, like movies and TV shows do all the damn time), but even ignoring that, one of the best things about the PS5 SSD is that it gives developers more headroom and cushions them from having to deal with sudden, rare spikes in I/O need. The less time they need to spend on optimising that sort of thing, the more they can spend on other stuff.

Additionally, if I'm understanding the tech right, stuff like Nanite in UE5 will also offer a visual advantage on PS5, as the detail level of assets presented on-screen will scale with I/O speed.
 

stryke

Member
Oct 28, 2017
2,347
it's going to be exciting what devs can do next gen
Insomniac said this seemless cutscene hiding exterior to interior transition took them one month to complete

 

panda-zebra

▲ Legend ▲
Member
Oct 28, 2017
5,746
Latency is not in relationship with bandwidth, latency is a constant you need to add on top of size/transfer speed, to get realistic real life timings.

Anyway, let me go with another example. You remember those corridors in God of War they had to put for loading? When Kratos squeezed through one for 10 seconds they were able to load around 800 megs of data. XSX will do it in 0.33 second(not including data compression and not having to optimize for serial reads, so even faster in practice), PS5 in 0,15 second. You really think those 183 ms make much of a difference, when over 9 whole seconds have been shaved?

PS5 would have to render much faster and on bigger data to saturate that bandwidth and its not going to happen. I would be willing to bet a small amount it will be a prime example of over engineering in a near future.
It's not over-engineering, it's under appreciation of how games will be made next gen. You're presenting a problem of running last gen game design on next gen hardware that's not aplicable to next gen only games.

A next gen console playing last gen games constantly streaming chunked assets to be used in the next say 30s in a fraction of a second vs 10 seconds+ is irrelevant and not looking to the true advantages of incredibly fast I/O. "183ms" or any such interval that might be 2x+ greater latency between platforms could make all the difference in the level of detail you're able to attain in your assets presented on screen.

These next gen machines don't really have the huge leap in RAM other gens have had, they're going to lean heavily on the speed textures etc. can be discarded and replaced with exactly those required, games will be designed around this rather than the current packaged lumps of maybes and possibles. Of course PC will be able to brute force past such things as fast I/O by having masses of VRAM to cache far more assets in and/or by having on-GPU texture storage with high I/O of its own, but at a cost.
 

bi0g3n3sis

Banned
Aug 10, 2020
211
On the other hand explain to me how >6GB/s means the theoretical max is 6GB/s. All we know is the typical throughput is above 4.8 and above 6GB/s is possible.

Same as DF said for XSX SSD theoretical max is 6 during interview with Goossen ( even if Goossen stated over 6, DF rounded it. If it is closer to 7, he would say closer to 7 surely ).

mmcGwen.jpg

No it's not agree to disagree dude. You are objectively wrong. There is absolutely Nothing in the XSX that closes that gap. The system is objectively faster in I/O as the XSX is more powerful. These systems have different ways of being balanced and Sony put more bias in the speed of the console and MS put more in the Power.

It's not going to go further than that. that's what we are saying. There is not enough software in the world. When the hardware constraint is 2.4 raw. and even the. All of the higher numbers is all based on the mix of compression it's never just all textures. just stop saying it's higher than what MS stated. it's not correct.

There is no in further discussing and explanation. Checked the post history, he tried the same stuff around 5 months ago in this thread ( from page 14 till 22 ). Well, it's worth reading if you like. LOL

www.resetera.com

Tom Warren: Developers are saying CPU and SSD are key to next gen, not teraflops, Lockhart won't hold back games

500GB wouldn't be bad if they can start reducing game sizes.
 

d3ckard

Member
Dec 7, 2017
212
No, it really isn't. There are multiple situations where you would want to swap out assets that quickly (one of the most common being when you simply want to seamlessly cut from one location to another in gameplay or a real-time cutscene, like movies and TV shows do all the damn time), but even ignoring that, one of the best things about the PS5 SSD is that it gives developers more headroom and cushions them from having to deal with sudden, rare spikes in I/O need. The less time they need to spend on optimising that sort of thing, the more they can spend on other stuff.

Additionally, if I'm understanding the tech right, stuff like Nanite in UE5 will also offer a visual advantage on PS5, as the detail level of assets presented on-screen will scale with I/O speed.

My argument is that once you get to the size of asset that PS5 will manage and XSX not, that asset will be too big for PS5 to render in time. Constraints on storage were pushed much, much more than everything else on the system(in both cases).

Yes, PS5 has ~70 times the last generation bandwidth, but only with twice the RAM and less than 6 times computing power. Simple napkin math shows they won't be able to utilize it in full.

The 60Hz scenario another user posted is much more interesting, but the argument stays the same - you run out of compute power before you can utilize additional data.

Data streaming was a problem last gen, but it is not anymore. The games were not "starved" for assets, because 60 FPS would be much more prevalent then.

I don't see other benefits from those crazy speeds than initial loadings, since they are pretty much the only situation you need to fill whole RAM fast. But it will end with waiting a couple of seconds longer. Nice, not particularly game changing.
 

BreakAtmo

Member
Nov 12, 2017
12,967
Australia
My argument is that once you get to the size of asset that PS5 will manage and XSX not, that asset will be too big for PS5 to render in time. Constraints on storage were pushed much, much more than everything else on the system(in both cases).

Yes, PS5 has ~70 times the last generation bandwidth, but only with twice the RAM and less than 6 times computing power. Simple napkin math shows they won't be able to utilize it in full.

The 60Hz scenario another user posted is much more interesting, but the argument stays the same - you run out of compute power before you can utilize additional data.

Data streaming was a problem last gen, but it is not anymore. The games were not "starved" for assets, because 60 FPS would be much more prevalent then.

I don't see other benefits from those crazy speeds than initial loadings, since they are pretty much the only situation you need to fill whole RAM fast. But it will end with waiting a couple of seconds longer. Nice, not particularly game changing.

Pretty sure PS5 has more than twice the RAM and more than 6x the GPU power. PS4 games RAM was 5GB, PS5 is unknown but absolutely more than 10GB. For the GPU it has less than 6x the TF, but if I remember right RDNA2 is a good 50% more efficient than the second-gen GCN used in the PS4 chip.

But never mind that. My opinion is that we need to wait and see what ends up being done with the machine, but keep in mind that if this whole situation is just as simple as "napkin math", don't you think the designers of the PS5 would have considered all this?
 

gofreak

Member
Oct 26, 2017
7,819
My argument is that once you get to the size of asset that PS5 will manage and XSX not, that asset will be too big for PS5 to render in time. Constraints on storage were pushed much, much more than everything else on the system(in both cases).

Yes, PS5 has ~70 times the last generation bandwidth, but only with twice the RAM and less than 6 times computing power. Simple napkin math shows they won't be able to utilize it in full.

That isn't how that works. Doubling RAM in a console - data access - doesn't mean you have to double your processors to see benefit. If this were the case we wouldn't even need multi-GB/s SSDs in the next gen consoles either, a ~100 or ~200MB/s SSD would do. There isn't some universal linear relationship between data access and processing capacity. Having more data access doesn't necessarily mean processing more data - it can mean getting to more accurate data or more variety of data in a given window of time. For example if I'm sampling between multiple views of an imposter to generate something suitable for my current view, being able to get to more data won't increase the amount of data I have to process, it will 'just' improve the relevance of the imposter views I'm sampling, and the quality of the final result vs a from-scratch rendering. Indeed there are scenarios where the ability to get to data can save processing capacity, classic time:memory trade offs.

Re. SSDs specifically, and storage usage, we will have to wait and see how hard devs go in on this stuff. But I find it really hard to believe Sony, or anyone, could really 'overspec' on data access at the moment.
 

Fisty

Member
Oct 25, 2017
20,366
My argument is that once you get to the size of asset that PS5 will manage and XSX not, that asset will be too big for PS5 to render in time. Constraints on storage were pushed much, much more than everything else on the system(in both cases).

Yes, PS5 has ~70 times the last generation bandwidth, but only with twice the RAM and less than 6 times computing power. Simple napkin math shows they won't be able to utilize it in full.

The 60Hz scenario another user posted is much more interesting, but the argument stays the same - you run out of compute power before you can utilize additional data.

Data streaming was a problem last gen, but it is not anymore. The games were not "starved" for assets, because 60 FPS would be much more prevalent then.

I don't see other benefits from those crazy speeds than initial loadings, since they are pretty much the only situation you need to fill whole RAM fast. But it will end with waiting a couple of seconds longer. Nice, not particularly game changing.

With those "crazy speeds" you don't have to keep the RAM full of old data or potentially unneeded data like right now. The reason people weren't fainting in shock & dismay when 16GB of RAM was announced is because the insane I/O speed means that a much larger proportion of the RAM can be used for much more "immediate" data and far less future and past data.

The reason PS5's solution is seen as a paradigm shift compared to XSX is because when you have the ability to fill roughly 8GB of RAM one second, just as Cerny said you can start rendering literally only what the player is looking at. That means the asset quality can increase exponentially, which also frees up the GPU from spending unnecessary time rendering what you arent looking at... all without the need for extreme memory management like in the past. Being 2x as fast as XSX might not seem like a whole lot when we are talking about sub-second speeds, but think about it in terms of size: if it takes 1 second to move 8GBs into RAM on PS5, you could only get 4GB on XSX in the same amount of time. Thats a huge difference.
 

BreakAtmo

Member
Nov 12, 2017
12,967
Australia
That isn't how that works. Doubling RAM in a console - data access - doesn't mean you have to double your processors to see benefit. There isn't some universal linear relationship between data access and processing capacity. Having more data access doesn't necessarily mean processing more data - it can mean getting to more accurate data or more variety of data in a given window of time. For example if I'm sampling between multiple views of an imposter to generate something suitable for my current view, being able to get to more data won't increase the amount of data I have to process, it will 'just' improve the relevance of the imposter views I'm sampling, and the quality of the final result vs a from-scratch rendering. Indeed there are scenarios where the ability to get to data can save processing capacity, classic time:memory trade offs.

Thanks for this, I'm always trying to learn more about this stuff.
 

d3ckard

Member
Dec 7, 2017
212
With those "crazy speeds" you don't have to keep the RAM full of old data or potentially unneeded data like right now. The reason people weren't fainting in shock & dismay when 16GB of RAM was announced is because the insane I/O speed means that a much larger proportion of the RAM can be used for much more "immediate" data and far less future and past data.

The reason PS5's solution is seen as a paradigm shift compared to XSX is because when you have the ability to fill roughly 8GB of RAM one second, just as Cerny said you can start rendering literally only what the player is looking at. That means the asset quality can increase exponentially, which also frees up the GPU from spending unnecessary time rendering what you arent looking at... all without the need for extreme memory management like in the past. Being 2x as fast as XSX might not seem like a whole lot when we are talking about sub-second speeds, but think about it in terms of size: if it takes 1 second to move 8GBs into RAM on PS5, you could only get 4GB on XSX in the same amount of time. Thats a huge difference.

Yeah, I'd like to see it. Rendering what you look at(which is by the way already the case, you mean loading only what you look at) is not feasible at that speeds, at least not how you describe it. You would have to restrict movement set and make character sluggish, since it will not be fast enough to follow the camera.

With data you have two basic approaches - just in time or preloading. For graphics rendering just in time would mean loading all necessary data each frame and even the PS5 is not even close fast enough for that. If you can't have that(and you can't), you have to stream to update your memory state as player traverses the world. And yes, PS5 will stream faster. Except it won't matter, because game usable traversing speed has limits and those limits correspond to the maximum requirements on bandwidth. You don't have to stream any faster than player moves to new areas in the world. Now, with current image fidelity, I claim that those requirements are vastly lower than available storage bandwidth. Better assets DO require better compute, since that quality has to be rendered, in high enough resolution and lighting, so you can actually see the difference. It's not like graphical chips were on vacation whole of last generation, they were plenty busy rendering assets we had. Even if we can improve on this front, I doubt the improvement will be more than 30x asset size, which is when XSX can't compete anymore. Alternatively, even if you use those assets, I doubt your rendering pipeline will use them properly, as you just won't have enough GPU juice.
 

score01

Member
Oct 27, 2017
3,709
Additionally, if I'm understanding the tech right, stuff like Nanite in UE5 will also offer a visual advantage on PS5, as the detail level of assets presented on-screen will scale with I/O speed.

It would be silly to argue that the additional 2TF the XSX has over the PS5 is over engineering and not needed because with graphics we can see or imagine what type of differences extra TF can bring. Issue with faster I/O is people don't really have an understanding of the type of things it can impact.

With VA coming to PC as well, I expect future game engines to start being able scale what they do based on I/O in a similar manner that they can scale what they do based on GPU TF.
 

west

Member
Oct 28, 2017
394
My argument is that once you get to the size of asset that PS5 will manage and XSX not, that asset will be too big for PS5 to render in time. Constraints on storage were pushed much, much more than everything else on the system(in both cases).

Yes, PS5 has ~70 times the last generation bandwidth, but only with twice the RAM and less than 6 times computing power. Simple napkin math shows they won't be able to utilize it in full.

The 60Hz scenario another user posted is much more interesting, but the argument stays the same - you run out of compute power before you can utilize additional data.

Data streaming was a problem last gen, but it is not anymore. The games were not "starved" for assets, because 60 FPS would be much more prevalent then.

I don't see other benefits from those crazy speeds than initial loadings, since they are pretty much the only situation you need to fill whole RAM fast. But it will end with waiting a couple of seconds longer. Nice, not particularly game changing.

Memory pool size actually matters less the faster your IO is. Imagine a theoretical scenario where disk IO is as fast as your ram, now you effectively have a terrabyte or more of RAM. We naturally are not there yet, since we are not near the speeds needed for rendering work; But for game assets not being rendered, it is definetly are fast enough to have it act like we had a RAM disk. In current gen games the IO speeds are so slow you need to store huge chunks of the levels you are not even seeing
wasting precious memory. Combined with better gpu decompression the actual memory difference between ps4 and ps5 is much much larger than it looks on paper.

Faster level loads? Thinking about loading levels is an outdated concept. The only reason the whole concept exists is to work around slow IO. Assets where packaged in zip files to make it as fast as possible, but then you got these large chunks that where indivisble and you had to design your levels around using them. Quake pak files anyone?

Assets where also not the issue with framerates last gen, weak cpus where. What IO gives you are things like persistent world states and non repeating assets.
No more repeating cars and npcs. Environmental damage can still be there a 100h later when you return, just write it back to disk.

Image almost infinite LOD when looking at objects up close in vr. You can allways maximize your memory to what you are looking at since replacing it is instant. This is not limited by compute, it is about having the right thing in memory TO compute.

We are already reaching levels of per pixel geometry and per pixel lightning. I truly believe that IO will scale better in these scenarios than raw compute and will have a more profound impact on gameplay than pushing resolution. I think the X Box series S is a testament to that. IO and Cpus are pushing next gen, more pixels is just gravy on top.
 

d3ckard

Member
Dec 7, 2017
212
It would be silly to argue that the additional 2TF the XSX has over the PS5 is over engineering and not needed because with graphics we can see or imagine what type of differences extra TF can bring. Issue with faster I/O is people don't really have an understanding of the type of things it can impact.

Do you actually have that understanding? GPUs are trivial to scale, a lot of your computing has natural granularity of one pixel, which for 4K is around 8 million similarly sized jobs. When all you do is matrix operations on sequences and sequences of data, it's not hard to scale.

What you see is what engine rendered. As long as your engine is fed with data as fast as you can process it, you're all set. There is nothing to improve as it won't change anything. And how much you can eat is bottlenecked on your mouth(GPU), not your food supply(storage). GPU sets a limit on how much data can effectively be used and once you reach that limit you're just pumping air.
 

behOemoth

Member
Oct 27, 2017
5,723
It would be silly to argue that the additional 2TF the XSX has over the PS5 is over engineering and not needed because with graphics we can see or imagine what type of differences extra TF can bring. Issue with faster I/O is people don't really have an understanding of the type of things it can impact.

With VA coming to PC as well, I expect future game engines to start being able scale what they do based on I/O in a similar manner that they can scale what they do based on GPU TF.
I think Epic already stated that the demo was scaling very hard with the GPU performance which means the Series X could probably push the resolution to ~ 1800 p
 

2Blackcats

Member
Oct 26, 2017
16,161
Yeah, I'd like to see it. Rendering what you look at(which is by the way already the case, you mean loading only what you look at) is not feasible at that speeds, at least not how you describe it. You would have to restrict movement set and make character sluggish, since it will not be fast enough to follow the camera.

Yeah, that's the perfect example really. You would have to restrict movement\camera speed. You'd have double your restriction on Xsx.
 

d3ckard

Member
Dec 7, 2017
212
Memory pool size actually matters less the faster your IO is. Imagine a theoretical scenario where disk IO is as fast as your ram, now you effectively have a terrabyte or more of RAM.

Mate, XSX game RAM has 560 GB/s bandwidth, 100 times what PS5 SSD has. That scenario is so far away it's not even worth discussing.

We naturally are not there yet, since we are not near the speeds needed for rendering work; But for game assets not being rendered, it is definetly are fast enough to have it act like we had a RAM disk. In current gen games the IO speeds are so slow you need to store huge chunks of the levels you are not even seeing
wasting precious memory. Combined with better gpu decompression the actual memory difference between ps4 and ps5 is much much larger than it looks on paper.

The issue is not with memory difference, but with compute difference. You will get better assets, just not 70x size better, because PS5 GPU won't have enough juice to process them.

Faster level loads? Thinking about loading levels is an outdated concept. The only reason the whole concept exists is to work around slow IO. Assets where packaged in zip files to make it as fast as possible, but then you got these large chunks that where indivisble and you had to design your levels around using them. Quake pak files anyone?

Levels can be artistic choice. The problem of better streaming doesn't require that level of bandwidth.

Assets where also not the issue with framerates last gen, weak cpus where. What IO gives you are things like persistent world states and non repeating assets.
No more repeating cars and npcs. Environmental damage can still be there a 100h later when you return, just write it back to disk.

How much state exactly you want to store? Because object id and position vector do not actually require killer bandwidths.

Image almost infinite LOD when looking at objects up close in vr. You can allways maximize your memory to what you are looking at since replacing it is instant. This is not limited by compute, it is about having the right thing in memory TO compute.
Its not instant. LOD doesnt require killer bandwidths. Even having everything in memory, 16 ms seems to be something hard to reach with current compute and even 33 ms is sometimes causing trouble, with current assets. How exactly faster IO is going to alleviate this?

We are already reaching levels of per pixel geometry and per pixel lightning. I truly believe that IO will scale better in these scenarios than raw compute and will have a more profound impact on gameplay than pushing resolution. I think the X Box series S is a testament to that. IO and Cpus are pushing next gen, more pixels is just gravy on top.

You're basically saying that gasoline will scale better than car engine. Sorry, no. And while CPUs can provide for a deeper gameplay, IO can only make assets load faster.
 

d3ckard

Member
Dec 7, 2017
212
Yeah, that's the perfect example really. You would have to restrict movement\camera speed. You'd have double your restriction on Xsx.

For PS5, assuming 8GB of assets per view and 60 degree camera, it would take six seconds for a character to do a 360 turn. Once you reduce the assets to get into one second, they will all fit in memory in the first place. I love how people spend hours speculating about tech, but are to lazy for a 5 second calculation reality check.
 

2Blackcats

Member
Oct 26, 2017
16,161
For PS5, assuming 8GB of assets per view and 60 degree camera, it would take six seconds for a character to do a 360 turn. Once you reduce the assets to get into one second, they will all fit in memory in the first place. I love how people spend hours speculating about tech, but are to lazy for a 5 second calculation reality check.

Hmmn

What about a track based racing game like Wipeout? Would being able to load the upcoming track at twice the speed have benefits? Could you have an extra speed class to take advantage of it?
 

rntongo

Banned
Jan 6, 2020
2,712
With those "crazy speeds" you don't have to keep the RAM full of old data or potentially unneeded data like right now. The reason people weren't fainting in shock & dismay when 16GB of RAM was announced is because the insane I/O speed means that a much larger proportion of the RAM can be used for much more "immediate" data and far less future and past data.

The reason PS5's solution is seen as a paradigm shift compared to XSX is because when you have the ability to fill roughly 8GB of RAM one second, just as Cerny said you can start rendering literally only what the player is looking at. That means the asset quality can increase exponentially, which also frees up the GPU from spending unnecessary time rendering what you arent looking at... all without the need for extreme memory management like in the past. Being 2x as fast as XSX might not seem like a whole lot when we are talking about sub-second speeds, but think about it in terms of size: if it takes 1 second to move 8GBs into RAM on PS5, you could only get 4GB on XSX in the same amount of time. Thats a huge difference.

If the PS5 had significantly more RAM then this argument would hold. At the throughput of say 4.8 for XSX and 9 for PS5, you'r filling 14GB of RAM in < 3 seconds. But even then Of that 14GB of RAM for games most likely you're going to have 2-3GB of data that isn't refreshed as often due to locality etc. Thats why the XSX has game optimal RAM as 10GB. The "working set of RAM" even on the PS5 won't be the whole 14GB available for games it will be less. So texture filtering comes into play as a memory multiplier. We already know about SFS on the XSX, and I'm sure Sony has something for that as well.

But basically most likely it will be less than 10GB of RAM that is refreshed constantly.
 

d3ckard

Member
Dec 7, 2017
212
Hmmn

What about a track based racing game like Wipeout? Would being able to load the upcoming track at twice the speed have benefits? Could you have an extra speed class to take advantage of it?

In theory, yes.

There is a max size of assets that will saturate your compute(optimum graphical performance). As long as the assets are ready, your compute will proceed normally and keep the frame rate. Now, in normal gameplay, your streaming needs is how fast your field of view changes, I.e. how fast you move. If you move faster, you need faster streaming. Question is whether the speed at which this would matter will make for a fun gameplay and whether you will be even able to spit a difference with image moving so fast.
 

rntongo

Banned
Jan 6, 2020
2,712
Same as DF said for XSX SSD theoretical max is 6 during interview with Goossen ( even if Goossen stated over 6, DF rounded it. If it is closer to 7, he would say closer to 7 surely ).

mmcGwen.jpg





There is no in further discussing and explanation. Checked the post history, he tried the same stuff around 5 months ago in this thread ( from page 14 till 22 ). Well, it's worth reading if you like. LOL

www.resetera.com

Tom Warren: Developers are saying CPU and SSD are key to next gen, not teraflops, Lockhart won't hold back games

500GB wouldn't be bad if they can start reducing game sizes.

I trust DF and what they posted in the video is not in line with what they posted in the interview on their website and furthermore at HotChips, the people that designed the console put it at above 6GB/s; here it is:

0WEGp7L.png


There's no point in making a false equivalence where it doesn't exist. We can only infer what the typical throughput for the XSX is not the theoretical max because they didn't mention it. On the other hand, the Oodle developers gave a solid estimate what the average throughput will be for the PS5, 2:1. It's impressive.
 

amoy

Member
Oct 28, 2017
1,230
Anyone else already over the SSD? Yeah it's hella fast, but I still can't wrap my head around the I/O complex, the way Cerny explained removing bottlenecks and how it will handle some things automatically and seamlessly. Things will feel like black magic soon enough.
 
Jan 3, 2019
3,219
For PS5, assuming 8GB of assets per view and 60 degree camera, it would take six seconds for a character to do a 360 turn. Once you reduce the assets to get into one second, they will all fit in memory in the first place. I love how people spend hours speculating about tech, but are to lazy for a 5 second calculation reality check.
What kind of fucked up scenario is that? Assuming If turning 180 degrees took half a second, that's 4-5GB of data being pooled from the SSD. It's less than half on a series S/X.
 

2Blackcats

Member
Oct 26, 2017
16,161
In theory, yes.

There is a max size of assets that will saturate your compute(optimum graphical performance). As long as the assets are ready, your compute will proceed normally and keep the frame rate. Now, in normal gameplay, your streaming needs is how fast your field of view changes, I.e. how fast you move. If you move faster, you need faster streaming. Question is whether the speed at which this would matter will make for a fun gameplay and whether you will be even able to spit a difference with image moving so fast.

I still think we're only looking at this from the way current engines have been designed which historically have had to take into account 40mb\s (worst case scenario) HDD transfer rates.

Once devs get a handle on the new minimum there will be shift in how this stuff works. If we can't properly conceptualise it yet, than it's all the more exciting.

Looking like PS5 is better equipped to deal with that paradigm shift though we'll probably be looking at mid gen refresh consoles by the time they software side of the tech catches up properly.
 

bi0g3n3sis

Banned
Aug 10, 2020
211
I trust DF and what they posted in the video is not in line with what they posted in the interview on their website and furthermore at HotChips, the people that designed the console put it at above 6GB/s; here it is:

0WEGp7L.png


There's no point in making a false equivalence where it doesn't exist. We can only infer what the typical throughput for the XSX is not the theoretical max because they didn't mention it. On the other hand, the Oodle developers gave a solid estimate what the average throughput will be for the PS5, 2:1. It's impressive.

Figures in screenshot is what Goossen stated in Eurogamer interview. He said in the interview "over 6", DF rounded it as 6. 4.8 GB/s is typical for XSX SSD, 6 is theoretical max ( yes, i said 6, you can round it something like 6.4 if you want ). There is no need for further discussion in that. It is what it is. Really.
 

d3ckard

Member
Dec 7, 2017
212
What kind of fucked up scenario is that? Assuming If turning 180 degrees took half a second, that's 4-5GB of data being pooled from the SSD. It's less than half on a series S/X.

Mate, the example shows what happens when you load data following the camera. If you're at any moment rendering 8 GBs of assets at 60 degree wide camera, you have to load 6 times that for a full 360 degree view. Even if you optimistically, consistently pull 8 GB per second, it will put a hard limit of six seconds to rotate fully. That's slow. Once you get to one second(by reducing size of assets), the whole 360 degree world fits in memory and you do not need to stream at all.

That simple and puts the claims of loading following camera to rest. Can't be done on that hardware. Once you talk normal assets streaming, no reason to believe it will be a bottleneck for XSX.
 

SDR-UK

Member
Oct 27, 2017
2,401
Sucker Punch's next game is gonna have reverse load times and gonna create a time paradox. 👀
 

rntongo

Banned
Jan 6, 2020
2,712
Memory pool size actually matters less the faster your IO is. Imagine a theoretical scenario where disk IO is as fast as your ram, now you effectively have a terrabyte or more of RAM. We naturally are not there yet, since we are not near the speeds needed for rendering work; But for game assets not being rendered, it is definetly are fast enough to have it act like we had a RAM disk. In current gen games the IO speeds are so slow you need to store huge chunks of the levels you are not even seeing
wasting precious memory. Combined with better gpu decompression the actual memory difference between ps4 and ps5 is much much larger than it looks on paper.

Faster level loads? Thinking about loading levels is an outdated concept. The only reason the whole concept exists is to work around slow IO. Assets where packaged in zip files to make it as fast as possible, but then you got these large chunks that where indivisble and you had to design your levels around using them. Quake pak files anyone?

Assets where also not the issue with framerates last gen, weak cpus where. What IO gives you are things like persistent world states and non repeating assets.
No more repeating cars and npcs. Environmental damage can still be there a 100h later when you return, just write it back to disk.

Image almost infinite LOD when looking at objects up close in vr. You can allways maximize your memory to what you are looking at since replacing it is instant. This is not limited by compute, it is about having the right thing in memory TO compute.

We are already reaching levels of per pixel geometry and per pixel lightning. I truly believe that IO will scale better in these scenarios than raw compute and will have a more profound impact on gameplay than pushing resolution. I think the X Box series S is a testament to that. IO and Cpus are pushing next gen, more pixels is just gravy on top.

Great post, I get your argument and it makes sense. Storage I/O has more room for growth in terms of speed and this will provide one of the biggest benefits or changes to gaming next gen. Open worlds that are less restrictive, even linear single player games could become more interactive. But, still advancements in Accelerator such as RT and architecture improvements in GPUs for example the VRS hardware and AI upscaling hw can really change the way games look and how to interact with them. For example with RT, objects in the game could be more destructible and interactive since the lighting isn't fixed on them.
 

west

Member
Oct 28, 2017
394
Mate, XSX game RAM has 560 GB/s bandwidth, 100 times what PS5 SSD has. That scenario is so far away it's not even worth discussing.

Yup, as I said i next sentence. It was just to illustrate to we are one pathway towards it, even if it only approaching it. We can see the benefits before we hit 560 GB/s. Surely you can agree?

The issue is not with memory difference, but with compute difference. You will get better assets, just not 70x size better, because PS5 GPU won't have enough juice to process them.

I don't see it as such. You don't need to render everything at once. Naturally each frame will not have 70x the data. But if you want to change the scene 4 times in a second, it will take 4 times the assets but same compute. Same number of frames, same compute, vastly different memory requirement. Completely theoretical example ofc.

Levels can be artistic choice. The problem of better streaming doesn't require that level of bandwidth.

I agree totally. But more is more.

How much state exactly you want to store? Because object id and position vector do not actually require killer bandwidths.

Id like to think big ;) Like a Jet Set radio or splatoon type game , where you can spray paint any surface of the game world and it will stay there forever. Something like that would require the capacity to stream in and out every texture in game. No tiling or decals.

Its not instant. LOD doesnt require killer bandwidths. Even having everything in memory, 16 ms seems to be something hard to reach with current compute and even 33 ms is sometimes causing trouble, with current assets. How exactly faster IO is going to alleviate this?

Are you sure compute is causing the issue there? Not the LOD algorithm? There are more than one challenge to be solved there. Virtual geometry seems like one way froward.

You're basically saying that gasoline will scale better than car engine. Sorry, no. And while CPUs can provide for a deeper gameplay, IO can only make assets load faster.

I don't think there is a proper car analogy for this.
 
Jan 3, 2019
3,219
Mate, the example shows what happens when you load data following the camera. If you're at any moment rendering 8 GBs of assets at 60 degree wide camera, you have to load 6 times that for a full 360 degree view. Even if you optimistically, consistently pull 8 GB per second, it will put a hard limit of six seconds to rotate fully. That's slow. Once you get to one second(by reducing size of assets), the whole 360 degree world fits in memory and you do not need to stream at all.

That simple and puts the claims of loading following camera to rest. Can't be done on that hardware. Once you talk normal assets streaming, no reason to believe it will be a bottleneck for XSX.
Why do you think you'd need up to 48GB of dynamic data for one 360 degree view?
 
OP
OP
platocplx

platocplx

2020 Member Elect
Member
Oct 30, 2017
36,076
Same as DF said for XSX SSD theoretical max is 6 during interview with Goossen ( even if Goossen stated over 6, DF rounded it. If it is closer to 7, he would say closer to 7 surely ).

mmcGwen.jpg





There is no in further discussing and explanation. Checked the post history, he tried the same stuff around 5 months ago in this thread ( from page 14 till 22 ). Well, it's worth reading if you like. LOL

www.resetera.com

Tom Warren: Developers are saying CPU and SSD are key to next gen, not teraflops, Lockhart won't hold back games

500GB wouldn't be bad if they can start reducing game sizes.
Well that is a level of insanity I've never thought possible. They reallly. Really wasn't to be right even after months of being told no. Lmao. I really hope they don't have an aneurysm when they see the PS5 UI and how much faster it is swapping games and loading games.
 

gofreak

Member
Oct 26, 2017
7,819
Yeah, I'd like to see it. Rendering what you look at(which is by the way already the case, you mean loading only what you look at) is not feasible at that speeds, at least not how you describe it. You would have to restrict movement set and make character sluggish, since it will not be fast enough to follow the camera.

With data you have two basic approaches - just in time or preloading. For graphics rendering just in time would mean loading all necessary data each frame and even the PS5 is not even close fast enough for that. If you can't have that(and you can't), you have to stream to update your memory state as player traverses the world. And yes, PS5 will stream faster. Except it won't matter, because game usable traversing speed has limits and those limits correspond to the maximum requirements on bandwidth. You don't have to stream any faster than player moves to new areas in the world. Now, with current image fidelity, I claim that those requirements are vastly lower than available storage bandwidth. Better assets DO require better compute, since that quality has to be rendered, in high enough resolution and lighting, so you can actually see the difference. It's not like graphical chips were on vacation whole of last generation, they were plenty busy rendering assets we had. Even if we can improve on this front, I doubt the improvement will be more than 30x asset size, which is when XSX can't compete anymore. Alternatively, even if you use those assets, I doubt your rendering pipeline will use them properly, as you just won't have enough GPU juice.


I think the relationship you're drawing between GPU compute and asset data size is too simplistic. If it were simplistic and linear you'd expect GPU's to 'only' need data throughput/streaming rates 5 or 6 times better than the last gen. That being the case, everyone is vastly over-speccing, no? Why would they?

When people talk about view dependent loading of rendering data... this is already happening. The extent to which it is applied is just relatively very limited. There's huge headroom to explore there with enough bandwidth. That headroom is basically the whole hypothesis underpinning the UE5 tech.

What will be needed for what quality of results? That's unknowable at the moment. But I say all this to say, the relationship between compute resources and data access is not some simple linear thing, and the efficiency with which renderers deal with data isn't staying still.
 

Tora

The Enlightened Wise Ones
Member
Jun 17, 2018
8,650
Supercharged is such a great marketing buzzword

Just makes whatever product feel that much better lol
 

Mubrik_

Member
Dec 7, 2017
2,733
I trust DF and what they posted in the video is not in line with what they posted in the interview on their website and furthermore at HotChips, the people that designed the console put it at above 6GB/s; here it is:

0WEGp7L.png


There's no point in making a false equivalence where it doesn't exist. We can only infer what the typical throughput for the XSX is not the theoretical max because they didn't mention it. On the other hand, the Oodle developers gave a solid estimate what the average throughput will be for the PS5, 2:1. It's impressive.
We can only infer what the typical throughput for the XSX is not the theoretical max because they didn't mention it.
Literally in the image you quoted.
~6GB/s
If it were lower it would be 5, higher it would be 7

I don't see other benefits from those crazy speeds than initial loadings, since they are pretty much the only situation you need to fill whole RAM fast. But it will end with waiting a couple of seconds longer. Nice, not particularly game changing.
Nanite?
 

Dizastah

Member
Oct 25, 2017
6,126
Memory pool size actually matters less the faster your IO is. Imagine a theoretical scenario where disk IO is as fast as your ram, now you effectively have a terrabyte or more of RAM. We naturally are not there yet, since we are not near the speeds needed for rendering work; But for game assets not being rendered, it is definetly are fast enough to have it act like we had a RAM disk. In current gen games the IO speeds are so slow you need to store huge chunks of the levels you are not even seeing
wasting precious memory. Combined with better gpu decompression the actual memory difference between ps4 and ps5 is much much larger than it looks on paper.

Faster level loads? Thinking about loading levels is an outdated concept. The only reason the whole concept exists is to work around slow IO. Assets where packaged in zip files to make it as fast as possible, but then you got these large chunks that where indivisble and you had to design your levels around using them. Quake pak files anyone?

Assets where also not the issue with framerates last gen, weak cpus where. What IO gives you are things like persistent world states and non repeating assets.
No more repeating cars and npcs. Environmental damage can still be there a 100h later when you return, just write it back to disk.

Image almost infinite LOD when looking at objects up close in vr. You can allways maximize your memory to what you are looking at since replacing it is instant. This is not limited by compute, it is about having the right thing in memory TO compute.

We are already reaching levels of per pixel geometry and per pixel lightning. I truly believe that IO will scale better in these scenarios than raw compute and will have a more profound impact on gameplay than pushing resolution. I think the X Box series S is a testament to that. IO and Cpus are pushing next gen, more pixels is just gravy on top.
Great post.
 

McFly

Member
Nov 26, 2017
2,742
I trust DF and what they posted in the video is not in line with what they posted in the interview on their website and furthermore at HotChips, the people that designed the console put it at above 6GB/s; here it is:

0WEGp7L.png


There's no point in making a false equivalence where it doesn't exist. We can only infer what the typical throughput for the XSX is not the theoretical max because they didn't mention it. On the other hand, the Oodle developers gave a solid estimate what the average throughput will be for the PS5, 2:1. It's impressive.
They did. It is > 6 but < 7. if it was greater than 7 that slide would state;

2 general + texture decompression engines, total > 7GB/s output.

it is for the same reason we can also accept when Mark Cerny says average of all compressed data yield 8 - 9GB/s and data that compresses really well can go up to 22GB/s. RAD Game Tools intimates it is closer to 11GB/s average but with data that compresses really well like BC1-7 texture data with their RDO reached compression ratio of 3.16:1 putting it at 17GB/s. Closer to the max capability Mark Cerny spoke about.

This is why i hate engaging in these disingenuous discussions. You would say direct storage is better but you don't even know why other than Microsoft said they rewrote their entire IO stack. While in the face of evidence from the people who made Kraken who have been doing benchmarks all summer you claim they are wrong about their own compression libraries achieving 3.16:1 compression ratio.
 
Last edited:
Oct 26, 2017
6,151
United Kingdom
They did. It is > 6 but < 7. if it was greater than 7 that slide would state;

2 general + texture decompression engines, total > 7GB/s output.

it is for the same reason we can also accept when Mark Cerny says average of all compressed data yield 8 - 9GB/s and data that compresses really well can go up to 22GB/s. RAD Game Tools intimates it is closer to 11GB/s average but with data that compresses really well like BC1-7 texture data with their RDO reached compression ratio of 3.16:1 putting it at 17GB/s. Closer to the max capability Mark Cerny spoke about.

This is why i hate engaging in these disingenuous discussions. You would say direct storage is better but you don't even know why other than Microsoft said they rewrote their entire IO stack. While in the face of evidence from the people who made Kraken who have been doing benchmarks all summer you claim they are wrong about their own compression libraries achieving 3.16:1 compression ratio.

Great post.

It's pretty telling what is driving a lot of these disingenuous arguments.
 

MrFox

VFX Rendering Pipeline Developer
Verified
Jun 8, 2020
1,435
They did. It is > 6 but < 7. if it was greater than 7 that slide would state;

2 general + texture decompression engines, total > 7GB/s output.

it is for the same reason we can also accept when Mark Cerny says average of all compressed data yield 8 - 9GB/s and data that compresses really well can go up to 22GB/s. RAD Game Tools intimates it is closer to 11GB/s average but with data that compresses really well like BC1-7 texture data with their RDO reached compression ratio of 3.16:1 putting it at 17GB/s. Closer to the max capability Mark Cerny spoke about.

This is why i hate engaging in these disingenuous discussions. You would say direct storage is better but you don't even know why other than Microsoft said they rewrote their entire IO stack. While in the face of evidence from the people who made Kraken who have been doing benchmarks all summer you claim they are wrong about their own compression libraries achieving 3.16:1 compression ratio.
Also, I'm not sure why anyone is surprised at the 6GB peak achieved with the equivalent computational power of 2 zen cores for decomp and decrypt, while Sony gets 22GB peak with a 9 zen cores equivalent for decomp alone.

Another interesting part of MS figures is that the >6 is the sum of the zlib and the texture decomp, while Sony have a unified I/O pipeline.
 

rntongo

Banned
Jan 6, 2020
2,712
They did. It is > 6 but < 7. if it was greater than 7 that slide would state;

2 general + texture decompression engines, total > 7GB/s output.

it is for the same reason we can also accept when Mark Cerny says average of all compressed data yield 8 - 9GB/s and data that compresses really well can go up to 22GB/s. RAD Game Tools intimates it is closer to 11GB/s average but with data that compresses really well like BC1-7 texture data with their RDO reached compression ratio of 3.16:1 putting it at 17GB/s. Closer to the max capability Mark Cerny spoke about.

This is why i hate engaging in these disingenuous discussions. You would say direct storage is better but you don't even know why other than Microsoft said they rewrote their entire IO stack. While in the face of evidence from the people who made Kraken who have been doing benchmarks all summer you claim they are wrong about their own compression libraries achieving 3.16:1 compression ratio.

So by that logic when Mark Cerny said 22GB/s was theoretical max he meant >22?? Cmon man.

And I never said the Kraken developers were wrong, all I said is they stated the average ratio will be closer to 2:1. Will there be data that has a 3.16:1 ratio, yes.
 
Last edited:

rntongo

Banned
Jan 6, 2020
2,712
I'm not seeing any such logic in the post you quoted.

I guess its just an argument over semantics. I saw there was a 3.16:1 ratio when Oodle and Kraken are combined in the blog post by the Oodle engineer. Its being used to signify typifying performance, yet in the same post the engineer clearly states the average ratio will be closer to 2:1.

My point is that although 6GB/s is likely the max typical throughput, for certain data, the XSX will have higher ratios hence higher theoretical numbers.

Correct me if I'm wrong.

Also you mentioned something that I didn't quite understand what do you mean by the PS5 has a unified I/O pipeline compared to MSFT. And further, since the PS5 has more data to decompress it's not surprising to see that its the work equivalent to 9 cores. But to again, Microsoft has stated it's decomp block does the equivalent of 5 zen 2 cores. In the picture I attached from hotchips, they stated >2 Zen 2 cores. So I think it's an issue with MSFT not being very clear about their numbers in some case.
 
Last edited: