-
Hey, guest user. Hope you're enjoying NeoGAF! Have you considered registering for an account? Come join us and add your take to the daily discourse.
Graphics detail and SSD bandwidth.
- Thread starter psorcerer
- Start date
jakinov
Member
The SSD is not going to give it compute power it doesn't have but it allows for compute units who need data to wait less before they can actually start processing the data resulting in less CPU/GPU idling. The reason FLOPS are considered PEAK THEORETICAL performance is because it's the highest you COULD get but probably will NOT ever get. You can have a lot of power but if your CPU has to wait on storage (one of the biggest bottle necks of modern computing) it's going to be harder to get full performance out of your compute units. Of course this depends on your game and what you are trying to achieve because if you don't need data to go from 0 to 100 (0-Storage 100-CPU) it won't matter but if your game relies on or you want to rely on being able to get data from 0 to 100 very fast then it makes a huge deal in what you can do and can help you use your CPU/GPU more efficiently. If your game is fine with using most of what's already in RAM, then the SSD would not be a huge deal. But again, if a game relies a lot on data going 0 to 100, the PS5 will feed the CPU faster despite the faster RAM because of the storage bottleneck and thus again saving them CPU idling and also giving developers the ability to rely on this so that can arguably create new/better experiences (how much better, how many devs take advantage are not clear). The concept/idea that storage can be a huge bottle neck and you want data as "close" to the CPU/GPU as possible to get better performance is something that's been discussed since the 70s and taught in schools for decades. The concept is usually applied to CPU and memory but the CPU relying on data that's ultimately from the secondary storage is just that but exasperated.
Virtual memory is also already a general computer science concept and has been around for a long time. You can very much use secondary storage like an SSD (or even a HDD) without GDDR6 level of speed. The only qualification be virtual memory is for it to be secondary storage.
B_Boss
Member
Sure thing but the SSD + Custom I/O Unit?:
....It is something else altogether. Not the end all be all of tech but engineering that is significant enough to analyze and discuss.
Last edited:
VFXVeteran
Banned
It shouldn't take long to master implementing a cache system like the HZD demonstration. I'll be watching out for the very first game exclusive to the PS5 only.
alabtrosMyster
Banned
With only the necessary data loaded in memory at any given time,
The consoles will load faster in games that it makes a difference and they will be able to handle streaming of higher details data in games that offer them...
If the GPU/CPU in the PC are faster, well complex scenes would be able to render at a faster frame rate and/or resolution. Most games are still designed around a mechanical drive, so it's only an annoyance, add to this that many old games are somehow sluggish on an SSD (saving in Zoo Tycoon 2 takes forever on an SSD, I have never tried it on a PC at the time, another old PC game I tried recently gave me similarly long loading times for no apparent reason, I can't recall the name of this one).
So with current practices and APIs the SSD is a certain benefit, but not nearly as much as it should be, the faster the drive the truer it is... I have a 3.2GB/s SSD and on day to day use I can barely feel any difference compared to a good SATA drive.
RaySoft
Member
Well.. If you halfway though building your house, you come up with some genious ways you could do it better, you don't just tare down your framework to start from scratch, you implement it on your next project.
Yes, for the frustum aspect of the engine, we should see imediate benefits with PS5.
Last edited:
KlickKlack
Banned
that means:I have seen a lot of confusion in different threads on how SSD bandwidth can impact graphics fidelity. So let's do it.
There are a lot of misconceptions like: SSD has nothing to do with better graphics, more bandwidth means more RAM per frame, it's useless, and so on.
TL;DR
Higher SSD bandwidth can give you much higher detail around the player while keeping RAM usage and RAM bandwidth exactly the same.
Now let's start.
I have a picture for you.
WTF is that?
It's a graphical approximation of two LoD systems: orange and blue.
The horizontal axis (x) is distance from the player (player being always at zero coordinates)
The vertical axis (y ) is asset quality (texture resolution and such) the closer we are to the player (at zero) the better the asset quality should become.
The blue LoD system has 2.66x worse assets close to the player but kind of a better ones farther from them.
The area below each graph represents the amount of memory that each system uses.
You can calculate it yourself if you want, but I will tell you that the area is exactly the same.
I.e. both systems will use exactly the same amount of RAM and RAM bandwidth,
On the bottom of the graph you can see two lines with dots at the end.
Each line represents the length of the curve for the same color graph.
As you can see it looks like the orange graph is 2x longer than blue one.
And it's indeed the case (1.9x longer) but what does it mean?
Let's think about it.
When player moves anywhere, what happens to the graph?
It moves right.
Why? Because when player moves anywhere (left, right, forward or backward) the LoD system needs to discard the closest data (to the player) and stream-in the newer one.
Player always needs to see the best LoD around.
In fact the closer the data is to the player the faster it becomes discarded. And long distance data stays almost the same for a pretty long time.
(when you pass a building in wilderness all its textures are not needed anymore, but the mountains on the horizon will stay the same)
So the length of the curve represents the amount of data that needs to be reloaded in each frame when the player moves somewhere.
In our blue vs orange systems we load 1.9x more data each time into the orange system and get 2.6x more details around the player.
Here we obviously sacrificed some "longer distance" details for that.
We can tweak the orange system and reduce details around the player to 2x, or even below 1.5x and then we will have enough RAM to fill the "farther" details whuch will be pretty close to the blue system.
Still overall picture will look much better than in the blue system.
Also, in real LoD systems the far details will change so slow that they will be probably never reloaded anyway. And the real curve length will be shorter for both systems thus increasing the difference even more.
Update: how things compare to no streaming
in best case scenario ps5 can process 10,2 TF of data
and xsx 12,1 TF of data
that's great.
Radical_3d
Member
TF are not a measure of data. GB are.
DrKeo
Member
Your graph is made out of two totally valid LOD but does not represent a 3rd party game running on two systems with different SSD speeds. You are showing two LOD systems, I'm assuming you are riffing off the Naughty Uncharted VS TLOU LOD graph (link), but what this graph really shows is a game that prefers very high assets close and very low assets far way VS high assets very close and low assets far aways. You are basically saying with that graph, things up close will look better on PS5, things far away will look better on XSX and that's how the memory requirement remains the same. It's a choice, not an SSD VS SSD speed based decision. As you (and Naughty Dog) have shown, both graphs contain the exact same amount of data which means both graphs will require the exact same GB/s SSD to sustain (after all, Naughty dog did do that between Uncharted 1 and TLOU using the exact same PS3 HDD). Your graph represents a single frame of a game, there is nothing "overtime" in it, each point in time will have a bit different graph, same same but different, and if we lived in a perfect world and the game was perfectly balanced, as all things should be, all graphs will be 100% the same because each and every frame of the game will have the exact same load.I have seen a lot of confusion in different threads on how SSD bandwidth can impact graphics fidelity. So let's do it.
There are a lot of misconceptions like: SSD has nothing to do with better graphics, more bandwidth means more RAM per frame, it's useless, and so on.
TL;DR
Higher SSD bandwidth can give you much higher detail around the player while keeping RAM usage and RAM bandwidth exactly the same.
Now let's start.
I have a picture for you.
WTF is that?
It's a graphical approximation of two LoD systems: orange and blue.
The horizontal axis (x) is distance from the player (player being always at zero coordinates)
The vertical axis (y ) is asset quality (texture resolution and such) the closer we are to the player (at zero) the better the asset quality should become.
The blue LoD system has 2.66x worse assets close to the player but kind of a better ones farther from them.
The area below each graph represents the amount of memory that each system uses.
You can calculate it yourself if you want, but I will tell you that the area is exactly the same.
I.e. both systems will use exactly the same amount of RAM and RAM bandwidth,
On the bottom of the graph you can see two lines with dots at the end.
Each line represents the length of the curve for the same color graph.
As you can see it looks like the orange graph is 2x longer than blue one.
And it's indeed the case (1.9x longer) but what does it mean?
Let's think about it.
When player moves anywhere, what happens to the graph?
It moves right.
Why? Because when player moves anywhere (left, right, forward or backward) the LoD system needs to discard the closest data (to the player) and stream-in the newer one.
Player always needs to see the best LoD around.
In fact the closer the data is to the player the faster it becomes discarded. And long distance data stays almost the same for a pretty long time.
(when you pass a building in wilderness all its textures are not needed anymore, but the mountains on the horizon will stay the same)
So the length of the curve represents the amount of data that needs to be reloaded in each frame when the player moves somewhere.
In our blue vs orange systems we load 1.9x more data each time into the orange system and get 2.6x more details around the player.
Here we obviously sacrificed some "longer distance" details for that.
We can tweak the orange system and reduce details around the player to 2x, or even below 1.5x and then we will have enough RAM to fill the "farther" details whuch will be pretty close to the blue system.
Still overall picture will look much better than in the blue system.
Also, in real LoD systems the far details will change so slow that they will be probably never reloaded anyway. And the real curve length will be shorter for both systems thus increasing the difference even more.
Update: how things compare to no streaming
If you wanted to do a PS5 VS XSX SSD showdown and make a case for the PS5 displaying higher-quality assets because of its' SSD, this is how your graph should have looked without bias (please excuse my poor 5-min-photoshop skills):
This is a single frame, the Y-axis is the LOD, the higher it is, the higher detail the models have. the X-axis is the distance from the camera. PS5 is the orange area, XSX is the grey area (which 100% of it is on top of the orange), and as you can see XSX is always contained inside the PS5 because every asset in every LOD on the XSX, will be a lower quality subset of the PS5's asset. The orange is the extra better-looking assets, and they are residing in memory (because they are all in the frame) and load the GPU, CPU, and memory bandwidth.
If you stream better-looking assets, you display better-looking assets on every frame which has a performance load penalty on different parts of the console (which part is depending on the type of asset).
Radical_3d
Member
Your graph is wrong because the PS5 would need to be more powerful than the SX to be like that. Uncharted an TLoU may run in the same hardware but the explanation between the two (which the developer mad when he posted those examples) is a way more advance tech in moving information across the system. You can’t have the two graphs in the same spec because the moment the player moves you have to unload more detail and then load more detail again in the orange than in the blue one, thus the orange configuration does not require more power (which is saved in long distance detail) but yes a better memory system.Your graph is made out of two totally valid LOD but does not represent a 3rd party game running on two systems with different SSD speeds. You are showing two LOD systems, I'm assuming you are riffing off the Naughty Uncharted VS TLOU LOD graph (link), but what this graph really shows is a game that prefers very high assets close and very low assets far way VS high assets very close and low assets far aways. You are basically saying with that graph, things up close will look better on PS5, things far away will look better on XSX and that's how the memory requirement remains the same. It's a choice, not an SSD VS SSD speed based decision. As you (and Naughty Dog) have shown, both graphs contain the exact same amount of data which means both graphs will require the exact same GB/s SSD to sustain (after all, Naughty dog did do that between Uncharted 1 and TLOU using the exact same PS3 HDD). Your graph represents a single frame of a game, there is nothing "overtime" in it, each point in time will have a bit different graph, same same but different, and if we lived in a perfect world and the game was perfectly balanced, as all things should be, all graphs will be 100% the same because each and every frame of the game will have the exact same load.
If you wanted to do a PS5 VS XSX SSD showdown and make a case for the PS5 displaying higher-quality assets because of its' SSD, this is how your graph should have looked without bias (please excuse my poor 5-min-photoshop skills):
This is a single frame, the Y-axis is the LOD, the higher it is, the higher detail the models have. the X-axis is the distance from the camera. PS5 is the orange area, XSX is the grey area (which 100% of it is on top of the orange), and as you can see XSX is always contained inside the PS5 because every asset in every LOD on the XSX, will be a lower quality subset of the PS5's asset. The orange is the extra better-looking assets, and they are residing in memory (because they are all in the frame) and load the GPU, CPU, and memory bandwidth.
If you stream better-looking assets, you display better-looking assets on every frame which has a performance load penalty on different parts of the console (which part is depending on the type of asset).
polybius80
Member
the problem is people dont understand graphics at all, you show a pretty scene and mention it need some specs and people try to equate that to a form of metric they understand(16-bit, 32-bit 64-bit) when there is a lot of tricks that can save performance here or there or effects that look impressive but dont actually require much in terms of "power", if a scene look pretty its difficult to explain any gains or their importance because people only see graphics and how look or bad they appreciate them, they dont know how things work under the hood or the real implications of certain specs
SSD doesnt make the system more powerful, it makes it able to utilize it more efficiently allowing lot of performance saves(performance that otherwise may be wasted) that can help other parts of the rendering pipeline which in turn allows the scenes to look more detailed or more "pretty", but this just gives capacity to do that developers have to actually use that in some way or figure out a use that can help them
the next videos are not related to SSD but its an interesting approach of looking at a way to improve things in a system just using it more efficiently
I recommend this youtube channel since it explain the performance gains and how recognized effects where achieved
I dont know of a sega fan that wont praise Sonic R in Saturn for these effects
a fast SSD allows to save space and bandwidth from memory that in turn can be used for something else, to put it in a crude and incorrect way is like "being more powerful" but the correct explanation is that the power is intact is just that it offers ways to do certain things more efficiently
Last edited:
DrKeo
Member
Yes, that's exactly my point. My graph is the graph implementing the SSD streaming higher-quality assets. And that's the point - streaming higher-quality assets means more GPU/CPU/memory/memory bandwidth load because in the end, you have to render higher quality assets.
Last edited:
Radical_3d
Member
Eh, ok. But that’s not the point of using the better visuals with the same performance. The idea is save performance details far from the player so you don’t need more power for the same scene. That’s it, Uncharted and TLoU use the same 400GF of the PS3. But memory management makes details closer to the player in the later more data-heavy than in Uncharted, while Uncharted draws more detailed assets away from the player because can’t be swapped fast enough.
New Yorks Finest
Banned
Don't worry, OPs got a major in Tech field and works at Sonys HQ PS5 dev kit department.
Basically, this whole thread will be bum rushed with tech Wizards.
Cool. At least now we're talking!
Yep. Exactly like that. But you miss the important detail: slower SSD cannot magically choose the "orange" approach.
Its SSD is too slow for that. It won't be able to keep up.
And if it will chose that approach but scale it down appropriately it's "far" details will be much worse.
DrKeo
Member
100% of what's on-screen needs to be streamed, not just the detailed close object. When you turn the camera left the highly detailed rock needs to be streamed but the 40 rocks 15 meters way also need to be streamed. In the end, when the camera turns, 100% of what the camera sees needs to be streamed in and needs to be rendered. In the end, if you want a more detailed world, you need to:
1) Stream all that data in.
2) render all that data.
PS5 might have an advantage in #1, but a disadvantage in #2. If you think GPUs are sitting idle with nothing to do just because the storage isn't fast enough, you probably never owned a PC.
Radical_3d
Member
The OP is one of the two advantages of fast storage. First, you can choose to have the orange model. Second, you can have less RAM occupied as cache.
I though this thread was ridiculous since all of this is pretty obvious but four pages later I’m aware it’s not.
This is how I imagine the OP making the post knowing how we, gaffers, are tho:
So it is in the distance where streaming can be helpful. If a rock 40 meters away can be loaded with less detail you free GPU resources for the closer details (something that, as you pointed, the PS5 is worst at doing). More over. You can choose not to change the amount of details you are using in the PS5 version and have less RAM used as cache. This is fantastic because, as I said before, the main bottleneck in this generation is not computational power but the scarce amount of RAM.
I though this thread was ridiculous since all of this is pretty obvious but four pages later I’m aware it’s not.
This is how I imagine the OP making the post knowing how we, gaffers, are tho:
When you turn the camera the data is already loaded. No HDD is fast enough to steam at the speed you turn your character. Cerny said something about “if you turn slow enough” but that’s not a realistic scenario in 99% of the games. Unless you make something ridiculous like… tank controllers from the 90s games very slow or something like that.
So it is in the distance where streaming can be helpful. If a rock 40 meters away can be loaded with less detail you free GPU resources for the closer details (something that, as you pointed, the PS5 is worst at doing). More over. You can choose not to change the amount of details you are using in the PS5 version and have less RAM used as cache. This is fantastic because, as I said before, the main bottleneck in this generation is not computational power but the scarce amount of RAM.
Last edited:
That's not entirely true.
For GPU it's approx the same amount of effort (there are minor differences in how cache will work) to render a higher or lower rez texture if the resulting amount of pixels on screen will be the same.
GymWolf
Member
i mean, during this gen open world games already have some of the best graphics.
horizon, forza horizon 4, rdr2, days gone, metro exodus, infamous 3, ac origins, some section of gow are pretty wide open, same for unchy4 or tlou2 or gears5 or shadow of tomb raider etc.
the only linear game with on par graphics are ryse, hellblade and the order (and maybe someone else) and i can argue that they all look kinda worse compared to rdr2 if you consider stuff on screen and scope.
fuck me star citizen has some of the best graphic ever and it is a huge open world game.
Last edited:
DrKeo
Member
Data streams in while you are turning even in Spider-man on a 50MB/s HDD, it's able to load the MIPS0 of the stores for instance. Will the PS5 and XSX be fast enough to hold a buffer just for the next frame? Time will tell, but MS sounds pretty confident about that so I'm sure Sony is in an even better position for that. The next frame needs just a few dozens of MB, even when you turn really fast you still need to replace just a very small percentage of the screen. And if you fail to get all data in time for the frame? Not a big deal, you will have the lowest LOD in memory as a failsafe. I'm sure gamers will survive 8% of the edge of their screen being at a lower LOD for 33ms, especially considering that area will be motion blurred anyway because the camera is turning.The OP is one of the two advantages of fast storage. First, you can choose to have the orange model. Second, you can have less RAM occupied as cache.
I though this thread was ridiculous since all of this is pretty obvious but four pages later I’m aware it’s not.
This is how I imagine the OP making the post knowing how we, gaffers, are tho:
When you turn the camera the data is already loaded. No HDD is fast enough to steam at the speed you turn your character. Cerny said something about “if you turn slow enough” but that’s not a realistic scenario in 99% of the games. Unless you make something ridiculous like… tank controllers from the 90s games very slow or something like that.
So it is in the distance where streaming can be helpful. If a rock 40 meters away can be loaded with less detail you free GPU resources for the closer details (something that, as you pointed, the PS5 is worst at doing). More over. You can choose not to change the amount of details you are using in the PS5 version and have less RAM used as cache. This is fantastic because, as I said before, the main bottleneck in this generation is not computational power but the scarce amount of RAM.
It doesn't matter, each part of the streaming assets will tax another part of the GPU/CPU/memory/memory bandwidth. Just because the higher res on screen textures won't tax the GPU much doesn't mean they won't take more space in memory. Just because loading more alpha into the screen won't affect the CPU, it will hurt the GPU Performance and so on. Every little part you upgrade will have an effect on something so in order to render it, PS5 will need more powerful hardware.
If you stream it, you need to render it. If you render it, it takes a toll. Using different LOD curve won't help you because everything needs streaming, having a close 2048x2048 texture takes the same time to load as a further away 4 1024x1024 textures. When you turn left, that little 10% of the left side of the screen you need to load has X amount of polygons, X amount of textures and so on and in the end, the system doesn't really care if that data is close or far away (especially with the new texture paging streaming method MS is using), in the end, you stream in the same amount of data, just like your graph you made yourself.
I mean, we already saw consoles with a much larger speed difference in storage than PS5 VS XSX and still, you didn't really see storage impact graphics unless the difference was a few orders of magnitude (like the PSX VS N64).
Last edited:
Shmunter
Member
Where were you when Sony designed their super fast SSD solution? Oh well too late now, it was all for nothing.
Last edited:
It's hand-waving. Numbers?
DrKeo
Member
What's hand-waving? That if you are just streaming in 10% of the screen you need (on average) to stream 10% of the data? Even less than that if your engine is smart enough because the eye doesn't focus on the edges of the screen and while the camera is moving you have motion blur? Be more specific.
Last edited:
No. Faster I/O gives you a choice on how to do your memory management. Even faster I/O - even more choice.
Slower I/O requires you to hold more needless assets because you will never have time to load them later.
All of that is described in detail in OP.
Yet you still handwave it with "more data = more GPU" mantra.
Last edited:
Sota4077
Member
Being patronizing doesn't really change the truth of the matter. I've not downplayed anything Sony have done. In fact I have deliberately made it a point to never mention Sony or Microsoft aside from one post because even the OP didn't go there. The SSD will make the GPU operate more efficiently on console. It will not make it operate higher than spec. You can selectively choose all the datapoints you like to try and mislead people all you like, but reality is not on your side. At the end of the day my very simple analogy still is both fitting and accurate. You cannot fill a 1 gallon bucket(GPU Frame) with more than one gallon of water. You can only be more efficient about keeping it filled or refilling it quickly with a better hose or higher water pressure (SSD). That is what these SSD's will do. They will not unburden the job of the GPU which is to put what you see on your screen.
n0razi
Member
The CPU/GPU also cant draw faster than the storage can provide data. It's all about reducing bottlenecks.
"Notto disu shitto again"...
Fillrate has nothing to do with mesh density or texture resolution.
Fillrate is related to targets not sources (assets).
If resolution of output, shadows, PBR buffers, deferred buffers remains the same. The load on GPU is the same.
Yes there would be problems if triangles on screen are too small (smaller than 1 pixel). And there would be problems if tex resolution is too high (again smaller than 1/4 of a pixel).
But if you control for that it's the same load.
DrKeo
Member
Your whole first post in this thread is one huge hand-wave and you are calling my posts hand-waving? If you are streaming in the next frame, you have very little advantage if it's the same game (all you can do is basically move faster or turn faster or you will increase the load on screen) and if you stream in a delay of a few frames, you need a buffer so you add a saving on memory of part of that buffer (let's assume it scales perfectly, so ~x2 smaller buffer).
If you change the LOD curve like in your graph, you are still keeping the same load on the hardware, just changing the places it focuses most of its' power. GPU load remains the same, memory usage remains the same but also streaming requirements remain the same, that's why it worked in Uncharted 1 VS TLOU while staying on the same HDD.
Yes, slower I/O requires you to hold extra assets unless your storage is fast enough to feed the next frame. If what you are doing with that extra I/O is stream in higher quality data, you better have the power to render it (GPU/CPU/memory/memory bandwidth). You can't just throw x2 more polygons, textures, alpha, etc. and expect everything to be fine as if no bottlenecks exist in consoles except for storage.
I have shown that streaming requirements increase 1.9x
In TLOU they used a different LOD curve.
Ah. Now I get it. You actually implying that if the system I/O speed has the max speed that this system GPU will ever need then there is no advantage of having a faster SSD on another system.
That's a bold claim. Essentially it's like claiming: "we don't need more RAM".
Unfortunately in reality we do need more RAM. And purely by numbers: 2TF last gen -> 12TF this gen means we will need more RAM than 16GB.
And actually as I've already said: because you can always trade RAM for CPU/GPU - it's never enough RAM.
ResilientBanana
Member
Is there a graph showing how the power of the GPU effects the amount of triangles and textures that can be displayed and at what resolution for both consoles?
DrKeo
Member
No, you really didn't, you just made false assumptions about moving without turning the camera. more Objects move faster in and out of the frame when they are further away. If you turn 2 degrees to the left, the tree 2 meters from you will still be in the frame, but 100 trees a kilometer away might have popped in or out of the frame. I mean, what you say is true if you are playing an on-rails game without free camera movement but you are just taking a specific case in order to support your claim.
Ah? No, it doesn't. What the hell, did I say RAM without a streaming pool is always enough? It's the other way around, you are the one claiming that you will put more textures on screen without them taking more RAM. I'm actually saying the exact opposite, it doesn't matter how fast your streaming is if you don't have the power to use it (and by power I mean CPU, GPU, memory and memory bandwidth). You still need more RAM because VRAM and system RAM is used for a lot more than just a VRAM streaming pool. Does it matter if you stream to the next frame or the 7th? Streaming pools will be insanely small in both cases.
What a weird point you are trying to make.
Last edited:
Shmunter
Member
You seemingly understand the concept of filling and refilling faster, but are unable to conceptualize how that materializes in available ram vs buffer ram ratio. Not much more can be done here, you're just going to have to go on your own journey -or- not.
megreotsugua
Banned
I agree. Idling in the sense that the system is not rendering higher geometry, textures and details because the data cannot be delivered in time. Or worst case, pop-ins. 10.3TF of RDNA2 with geometry engine (rendering only what visible on the screen) is going to be data-starved next-gen.
Trimming down of assets to fit in the 5.5GB of RAM (supported by slow HDD) is a problem today. Next-gen, it's the creation and production that will be a problem if dev studios become very ambitious with their game.
I hope MS and Sony are both doing something to solve that problem. Good software perhaps for off-line procedural creation of assets. Or, perhaps sharing of assets between studios. Or, perhaps a new studio dedicated for production and creation of arts and assets to support the World Wide Studios. Production will be the key next-gen.
Last edited:
Radical_3d
Member
I find this very unlikely but very interesting, where can I find a source?
Kazekage1981
Member
SSDs give the engine (GPU) the fuel it needed for so long...… or something.
DigitalScrap
Member
DrKeo
Member
Yes, it's from Insomniac GDC talk about the Spider-man technical postmortem. They talk about streaming the low-quality MIPS first and the MIPS0 load fast enough that they load before you turn and see the shops. I guess that's why Cerny chose to show Spider-man for the SSD demonstration, it's already built really well for streaming:
(I don't have the exact time-stamp but the whole talk is extremely interesting
)They have the initial 20MB of the tile streaming in before you reach it but when you enter the tile, they are still streaming 12MB per second of data around you. Obviously it's cruder than what the PS5 and XSX can do and also results in some pop-in, but the idea is similar, having the low-quality assets in memory while the MIPS0 load and if the fail to load in time, no big deal, you get a tiny pop in for a fraction of a second.
edit:
I've timestamped the video.
Last edited:
Nice graphs in the start page, but there is one small piece missing.
Texture tiling makes really sense this gen (PS5 & xbox)
e.g. Microsoft presented this in 2013 (I did not find the fligh-demo, just the build conference video)
The HDD made texture tiling almost impossible or at least not really practical. The SSD now with lower access times has the ability to fully utilize that feature. In the current gen almost all the time most texture had already been loaded fully into the main memory. Yes, there were also be streamed (slowly with a large buffer) in packets but if they were needed, even if only a tiny bit was needed, the whole texture had to be in memory.
The SSD changes that a lot. Now the buffer can be much smaller, you no longer need packets to get everything into memory and therefore can only load those parts of a texture that are really needed before rendering the whole scene.
Texture tiling makes really sense this gen (PS5 & xbox)
e.g. Microsoft presented this in 2013 (I did not find the fligh-demo, just the build conference video)
The HDD made texture tiling almost impossible or at least not really practical. The SSD now with lower access times has the ability to fully utilize that feature. In the current gen almost all the time most texture had already been loaded fully into the main memory. Yes, there were also be streamed (slowly with a large buffer) in packets but if they were needed, even if only a tiny bit was needed, the whole texture had to be in memory.
The SSD changes that a lot. Now the buffer can be much smaller, you no longer need packets to get everything into memory and therefore can only load those parts of a texture that are really needed before rendering the whole scene.
DrKeo
Member
And that's exactly one of the big features of Direct Storage. Textures are cut into pages and the system isn't streaming the required texture, it streams just a few pieces of that texture required by the camera view. So if you have a painting hanging on the wall in front of you but 3/4 of the painting is outside of the frame, only 1/4 of the painting will stream on the XSX. Once you start turning, another 1/4 of the painting will stream, and as you keep turning the whole painting will stream (obviously it's not really going to happen in quarters, each texture will be cut into thousands of pages). That's an API feature baked into Direct Storage. And we PC gamers will also enjoy that feature which is awesome.
Radical_3d
Member
Yes, it's from Insomniac GDC talk about the Spider-man technical postmortem. They talk about streaming the low-quality MIPS first and the MIPS0 load fast enough that they load before you turn and see the shops. I guess that's why Cerny chose to show Spider-man for the SSD demonstration, it's already built really well for streaming:
(I don't have the exact time-stamp but the whole talk is extremely interesting
They have the initial 20MB of the tile streaming in before you reach it but when you enter the tile, they are still streaming 12MB per second of data around you. Obviously it's cruder than what the PS5 and XSX can do and also results in some pop-in, but the idea is similar, having the low-quality assets in memory while the MIPS0 load and if the fail to load in time, no big deal, you get a tiny pop in for a fraction of a second.
edit:
I've timestamped the video.
Maaaaaan. If this is posible in a sh**ty ass PS4 hard drive I have a lot of hopes for the next generation. Maybe the RAM is not going to be so much of an issue this generation, after all.
DrKeo
Member
I couldn't find it, but the XSX optimization and engine architect talked in twitter about streaming to the next frame while holding the lowest-quality assets in memory at all times and if an asset fails to make it in time from the SSD, the GPU just falls back on the best quality asset it has in memory. So it seems MS is and obviously Sony is trying to reach an unbelievable level of streaming that will try to stream to the next frame. That's mind-blowing considering it will make the streaming pool the size of a single frame new data requirements which are just a few dozens MBs.
Last edited:
alabtrosMyster
Banned
With only the necessary data loaded in memory at any given time,
The consoles will load and they will be able to handle
At first when I saw the image I though: Oh that explains it better!
However after further consideration I figured what was wrong, it does not show is what the less friendly curve graph is showing is that objects that are closer to the camera are using exponentially higher level of details and higher texture resolutions than those that are further away, the fast SSD and support APIs allows the software to keep only the necessary LODs for every object on the screen, thus freeing up space for the highest levels as they don't need to be kept in memory for every object at all distances at all time because otherwise they would take too long to load.
alabtrosMyster
Banned
Teraflops are not datas... Peak teraflops are kind of like max RPM for your car's motor, it doesn't tell you much about how many pounds of trust the vehicle has, or even how fast it could really go.
Radical_3d
Member
It is amazing. Later on I’ll see the whole video, but... it doesn’t load on a twist of the cámara but based on the position of the player. Hence, a distance based LOD.
alabtrosMyster
Banned
The techniques aren't completely new, but the jump is in order of magnitudes better than what we had until now.
Yup, the only difference is probably how much of it will be possible in the end, and how much each architecture allows them to extract the full potential it has on hand--This holds true for both the SSD/CPU/GPU, etc.
DrKeo
Member
Has said something about he the MIPS0 loading in faster than the player can turn the camera at some point but you are right, it's not like the loading of data is based on where you are looking, it's just streaming in everything around you as you move through the world and hoping the HDD will keep up well enough so there isn't too much pop in.
DrKeo
Member
Yup, I'm really eager to see DF cooperation videos between the PS5 and XSX in various third party games. I'm sure we are going to see some more storage based pop-in on XSX for instance.
Radical_3d
Member
I just want to pay 600€ for a 24GB machine. Is that a crime?