VG24/7:Epic rewrite I/O subsystem with PS5 SSD in mind
- Thread starter chris 1515
- Start date
Is this the part where we start dreaming about an OutRun revival and end up inevitably disappointed?
I wonder how expensive it'd actually be to produce a game like this on UE5.
that would be cool but i have long dreamed of games doing the kind of intercutting of scenes movies do. especially nolan movies where he has different characters in different places doing different things and it all comes together beautifully as one scene.
Man, this is a really good point that I didn't even think about. The ability to do hard cuts in real time will be a game-changer for cinematic titles.
I am no expert, but I think machine learning will help. Matt may be able to help us all.
6 Ways Machine Learning will be used in Game Development | Logikk
Machine Learning in Game Development is still in its infancy but it is set to change how we play and experience games in the future. Heres how...
www.logikk.com
The question is probably too broad and complex.
But I guess it could be cheaper than usual if game asset creation and utilization is easier for developers, and if the I/O improvements allow them to spend less time optimizing around the asset streaming needed when a car is cruising at high speed on a highway surrounded by lush, detailed environments.
I think at the moment I'm most interested in fighting games. Nothing revolutionary but this technology promises to bring you back to the days of the arcade or SNES/Genesis. Instant loading, instant stage swaps for rematches, stage transitions ala MK3 without loading like Injustice, Shang Tsung morphs, etc.
I remember those early MK games, where they had those other fights happening in the background. Now, we could have an attack kick the opponent into the background and the fight continue into that plane.
We have internalized the ways modern games are limited so much that it can be a challenge to imagine what could even be possible under a new set of restraints.
I certainly can see a certain game developer making full use of that...
Oh no ...I certainly can see a certain game developer making full use of that...
I certainly can see a certain game developer making full use of that...
Oh lord.
We haven't even seen Kojima's final form.
I think if we look at everything as information it would help. You'll get higher quality visuals of course which will result in bigger open worlds with more details. But there are a ton of "invisible" details you will get as well. More realistic and sophisticated sound, more complex and intelligent AI in enemies, more realistic physics simulation, etc. All of this is data that can now be loaded much faster.
I think this is probably why so many people are a little down on SSD's, thinking that they will only bring loading increases. Once designers get their hands on PS5 and Series X, I am sure they are going to make some radical shifts in how they make games but I don't see it happening at launch. Maybe in the 2 or 3 year of the systems.
This is really solid to point out. Anything the game needs access to is available quickly.
Besides Kojima i'm very interested in what Miyazaki can come up with.
I'm willing to bet Sony has at least one showpiece game ready for launch. They've been quiet for years and they need a "drop the mic" game to silence any doubt about their hardware.
I am sure they have. They certainly have to show off the tech of course. The SSD is their main talking point for next-gen, so it would be silly to launch without something that really raises the bar. Spiderman PS5 is probably one of them, since they have shown tech off already for that. What else is to be seen but I doubt they would leave room for people to question their SSD capability.
Yeah, I've finally got a chance to watch Cherno's videos and I I think I finally understand the excitement. I have absolutely no experience with programming, etc but from what I gather the PS5 will be able to load up to at least 5.5gb of data per second. That's 5.5gb of gaming information (visuals, sound, ai, physics, whatever really) for every second from the disk alone. This could free up more RAM because you won't need to store that information onto the ram as much as you did before. This is 5.5gb without compression. 8-9gb a second typically with compression.
Of course, I can be pretty off base and most likely missing a lot of details with my lack of experience.
I agree for the most part, and I think it's going to be a long while before this really comes to full fruition.
I'm talking like, people that are kids today growing up in a world with SSD's will later on grow up to be designing games in fundamentally different ways than we can even wrap our heads around at the moment.
For now though we'll likely see a lot of indies cooking up wild new ideas along with first parties. I think 3rd parties will be the slowest due to their insistence of making things as multiplat as possible.
Agreed on the 3rd party side. The good thing is that both the PS5 and Series X have fast SSD's, so the sooner the last gen is dropped, 3rd parties should have a better foundation going forward.
I don't think it will take that long. The current crop of developers have spent years slamming against walls trying to make their games work. Once a couple of those walls are gone it's not going to take a long time for them to notice and take advantage. It will take a while for big orgs to alter their existing tools, but I don't even think that will take super long. This is a change everyone is ready for.
or sega making a rally game with deep tracks of effcts that are possible due to the level of streaming. They gutted revo cause memory and io speed just wasn't there. Would've been bad ass too. I hope codemaster really flexes now that cpus are strong.
I meant more along the lines of mind bending changes to what games are. I agree that we'll be seeing really cool never seen before the stuff within a few years easy.
User Banned (3 days): Mod Whining
So they're derailing the thread, and not Issen?
Everything wrong with moderation on this site in a single example.
Every RTX GPU has dedicated Ray Tracing units. Enabling RT on games reduced performance. It's just a fact. It takes power, bandwidth, cycles, causes heat. It's not going to be free.
Modern AI, Neural Nets, and Machine Learning all function in a similar intertwining manner conceptually. They are end results, means to, and the study of the creation of minimizing/maximizing functions respectively for a given test environment. Understanding how they function helps us understand how they can be used, what their limitations are, where and why they fall short of our expectations, and just how much time and work may go into using these concepts.
To give an example lets say in your game you want to have your car traveling down the highway at 70 MPH and all of a sudden luscious trees start growing everywhere, and you want to use AI to instantly and dynamically grow these trees +400 years in the next 10 seconds. Sounds great, right? But what does it actually take to do something like this? Is it even possible in realtime?
Let me take you through the hypothetical story of Outrun-dev, the individual trying to make this single task possible. Little do they know what they just signed up for.
Outrun-dev is studying machine learning so that he can create a neural net that makes a tree growing AI.
This individual starts by looking for some basic code off the internet, but no one in their right mind has every really tried to grow a 3D digital tree with AI before. This means he is going to have to program how the AI interacts with the game engine and environment. So after conceptualizing and reading through about 25 scientific papers written over the last 20 years in the course of a week, Outrun-dev decides they are going to have to create a dynamic topology based geometry mesh transform ("DTBGMT" for short). Oh, boy.
This piece is going to act as an efficient way for the game engine to 3D sculpt these trees in real time which already have their final meshes loaded behind the scenes. This method was chosen because keyframe sculpting would overstretch polygons and would be less capable of generalization to arbitrary shapes.
Of course, they need to prototype any of this before it goes anywhere, so they start up a new project file. They create a seed point and a tree they pulled from a data scanned library. The AI's outputs are intended to adjust the parameters of the DTBGMT. Of course it's never just that easy. it's now 3 months later. The work required to develop just this piece is comparable to some of the work done on UE5's Nanite tech. And at this point it still doesn't actually do anything.
It's now time to create and train the neural net.
Neural nets are made by taking a set of inputs multiplying them by random numbers, creating a new layer, adding random numbers, triggering with an activation function, and repeating with different sizes and numbers of layers, and then using multi-dimensional gradient decent through backpropogation to repeatedly run the sim and slowly optimize with each new attempt.
First attempt: system crashes almost instantly. Welp... memory overflow.
Turns out the level of detail the DTBGMT is trying to sculpt is far greater than the mesh it is aiming to replicate. Quickly tweaking the min and max level of geometric distance to match the mesh and dynamically applying tessellation overcame these issues. It's taken 2 weeks of troubleshooting.
First run that doesn't crash (#42): "Was that supposed to be a shrub?" questions a teammate. "And what are these bulges?"
Unfortunately, for all intents and purposes all first runs are like that. After 5000 runs it should look a bit better. The developers PC chugs away all night.
"You know, that kinda works, but just doesn't look right." responds the teammate to Outrun-dev's excited plea the next morning.
They were right about that. +9001 runs later.
It still hasn't improved. In this instance Outrun-dev's problem is the design of the neural net. And with each new neural net design all the progress from previous runs is scrapped. Outrun-dev's first realization of this was soul crushing.
Other work becomes priority and this prototype gets refined in the background. It's been 2 years now and the prototype actually looks pretty good now. The boss stops by one morning and sees the prototype running in the background cycling through different types of flora: leaves, trees, shrubs, vegetation, and flowers, each with their own optimized AI applied at different surface angles and in different environments. They chime in "That's pretty smooth, how are the texture sets coming?"
The last piece of a cream cheese bagel hits the floor.
It was using a dynamic geometry mesh. Any texture applied to it would break as it grew. An entirely different texture system would have to be developed.
The piece of bagel was quickly swept up. "We'll be starting that shortly." Outrun-dev replied.
Creating that dynamically resizing 3D warped nearest location texture mapping would be an additional 2 months.
The truth is Outrun-dev could have pre-rendered/cached the animations of growing plant life and just have them play at specific points in levels. So after all this work that took much longer comparably, what was actually gained?
A lot, actually.
a) the ability to generalize the process to more than just a single type of tree.
b) the ability to dynamically apply it in arbitrary environments.
c) a dramatic speed up in not only end user results, but also the creation process for these types of environments.
d) more accurate automated placement and root structures that look grounded.
e) a better means for realtime deformation in engine.
Sadly, even all of this potential doesn't mean we will see all of this in this coming gen. Machine learning will be great for improving the game creation pipeline for sure, yet it is far from the silver bullet some dream it to be right now. There will still need to be much work around the engine changes we will be going through as worlds become more dynamic. And furthermore even highly optimized scenes may still bring high end hardware to it's knees given enough complexity or a misdirected approach.
Still, here's to hoping we all eventually see that dream game.
PS5 is the same. The dedicated hardware intersection engine can be finding an intersection at the same time as shaders are executing. In both the PS5 and XSX the the intersection engine is memory intensive, so you would have a very hard time using it to effectively "double the performance/compute power" of the system as MS/DF claims. As Cerny puts it, it works best in parallel with logic heavy code that has a lower demand on system memory.
I think you're missing some pretty important information to reach this conclusion. I'm not the tech guy on this forum but consider this:
One of the reasons games are already so massive is because they are duplicating objects a lot to be on several areas of the HDD/disc so when the game is rendering a scene it doesn't have to go search around because it's physically nearby other objects from that scene. You know those same objects that you see constantly (same signs, benches, debris, etc.)? It's the same you see 25 other places in the game. This balloons file size a ton. With an SSD, you only need to have that same object once.
And the level of detail isn't going to be hundreds of times higher. There will be higher polygon counts, sure, but there will more more layers of effects and the models won't have to be as compressed or have other objects rendered at lower detail in certain situations. Think about it like this, let's say your average AAA game is 75GB right now. Now let's say the difference between removing duplicated objects equals an increase in higher resolution/more detailed models. There is so much more that is possible on top of that because of the SSD's like increasing LOD broadly, increasing draw distances, implementing procedurally generated/assisted textures, environments, shaders, etc. Having the computational power for limited raytracing and other effects that increase the visual fidelity of assets without being something that takes up more space on the hard drive.
Not to mention removing limitations such as movement speed, level design, "cinematic" squeezing through cracks and such.
So yeah, developers are absolutely going to lean on 5.5GBs and push 8-9GBs as they mature with the console. There's just nothing stopping them.
They create a library called TreeGrow and start licensing it out to other developers, who start using it everywhere in open world RPG projects, and suddenly have a new income stream that keeps money coming in during the quiet time between projects. Eventually a big engine developer offers them money and they spend the rest of their days doing Scrooge Mcduck gold angels.
Very true though, these things are rarely zero effort, and can be a lot more work over time than taking the simple path. It's just whether you can ensure that effort is worthwhile once you've hit the goal (if you can hit it, R&D can sometimes just go down a dead end).
Makes me hanker even more for a Motorstorm: Apocalypse remake.
Yeah yeah, I get it. I was just pointing out how Microsoft and Sony were emphasizing that particular point in opposite directions, but to seriously not read much into it. In the end the most likely scenario is that they're just using AMD's built-in RDNA 2 RT solution anyway.
Cerny said basically the same as MS, but with technical caveats. I'd expect both to be the same RDNA 2 implementation.
MS said the RT is done with dedicated units inside the CUs so can run in parallel with regular shading ops. They claim its equivalent of 13TF, so you could almost consider it a 25TF of available compute.
Sony/Cerny said the RT is done with dedicated units inside the CUs so can run in parallel with regular shading ops. However, Cerny added that RT is heavy on memory bandwidth so that will be likely to impact the effectiveness of the general shaders - you can perhaps run 'local' tasks that don't need a lot of memory access but don't expect to get fully parallelized performance
A PS5 will be worth it alone if the SSD somehow reduces SFV loading times.
Since the dream for next gen RT is pretty much hybrid rasterisation, what will help most for RT next gen is how denoising is handled. That's one of the key areas where RT performance can improve.
Could you clarify? Are you suggesting MS is using more of a dedicated block for ray tracing like Nvidia while Sony is using a GPU hardware accelerated approach? Because as far as I've seen both consoles have the same CU-accelerated RDNA2 ray tracing approach.
Exactly the opposite. I am suggesting that they were in fact talking about different aspects of the exact same, standard RDNA 2 technology.
I just thought it was curious how they worded it in such different ways. That's why I originally said to "not read into it", "it" being the fact that they emphasized different aspects of the technology.
Enabling RT reduces performance because the RT cores are a bottleneck, as RT cores get more powerful you will be able to turn on RayTracing with close to no performance impact
Can you elaborate? Clearly you know more than the rest of us.
Epic are independent 3rd party. UE5 is multiplatform. So explain the vested interest?
Add to that, hopefully next year we can see 3rd parties have a ssd listed as the minimum spec on PC.
I would hope that is the minimum we can hope for. I wonder if we will get a PS5 version of SFV announced this year.
This is probably the big thing. Until games specify SSD on PC, they will always have to program around a much slower HDD.
The XSX is going to be more capable with RT because of their higher CU count and memory bandwidth. Do I expect them to focus on it more then Sony.
This is so correct.
I see the lack of imagination around and it is probably because of what you said.
Question, with the tech in both next gen consoles, will there be improvements to MP/online play? I know your connection/provider is still the major part of it but with aside, what will it improve? or not improve anything?
Games like Destiny and Division would allow almost instant/few seconds when loading/fast traveling in those worlds and just imagine what it will allow in future games of this type.
Games like Destiny and Division would allow almost instant/few seconds when loading/fast traveling in those worlds and just imagine what it will allow in future games of this type.
A first-party developer will have a bigger impact on how we view the medium in the future. I don't expect Microsoft, considering their publicly documented strategy, to show how an SSD can change how we play. I do expect to see more innovation from an organisation that has publicly stated they want exclusive PS5 experiences.
Microsoft are waiting it out. That's fine. Sony are going all-in. That's why I think Sony's showcase will be more revealing than Microsoft's right now.
It seemed like you didn't know so I was just correcting you. Nothing more.
It's OK to be pessimistic, no matter who is delivering the information.
Yup, im expecting a PS5 PRO monster on 3nm. Say 2023, 2024?