Unreal Engine 5, [UE5 Developer Availability 2022-04-05]
- Thread starter mpg1
- Start date
Im about to upgrade from my current 2080Ti, what SSD should i choose? Seriously, i thought we where past the point of where storage disks equal rendering performance and visual fidelity/effects. What dictates graphics is the GPU for the most, otherwise, sony could have went with the PS4 Pro's GPU and called it a day.
If you believe all this, then the PS6 is going to offer no improvements in GPU raw power capabilities, just a faster SSD. If sony could, they obviously would packed that box with a 20TF gpu (all things being equal).
Fast ssd enables higher quality assets. Case in point ratchet&clank offloading behind camera things and streaming them in if/when needed. All memory used for visible things. If it's just about flops how would one explain halo and how it looked? I thought it's only about flops?
Obviously this kind of technology doesn't replace everything in game rendering, it's one method among many to render different kinds of game assets. RT being another for different things like lighting and reflections. The point is smart use of very high speed asset streaming for scene geometry, where to achieve the same result with more traditional techniques you would need a far more powerful GPU and large amounts of graphics memory.
I'm struggling to understand how a fast SDD somehow makes up for a more powerful GPU. I understand it making up for more VRAM, but how exactly does it make up for a more powerful GPU?
To be clear, the R&C:RA installation is <40GB. So with compression it's probably equal to around 80GB of game assets. Assuming the PS5 is only using about 10GB as actual VRAM that means you can refresh the VRAM at most 8x throughout the entire game without replicating assets. Obviously there will be lots of asset reuse throughout the game, but if people are seriously believing that R&C is achieving it's look by stuffing the RAM chock full of brand new content every 2 seconds then I'd love to understand how there's so much variation in the game environments when in theory, apparently, even the current viewport in any given scene is taking up 10-16GB, or around 15-20% of the entire game content.
Not everything in memory is loaded data from the SSD. Most of it might just contain the render-target and other data needed for RT, ...
It doesn't, not in the traditional sense of having (more, more, and more) compute/render power. The goal for PS5 SSD/IO isn't necessarily to replace GPU/CPU workloads, but more so to alleviate texture/streaming/decompression workloads as much as possible [away] from the GPU/CPU. In a sense, making it more efficient than trying to brute force everything and/or requiring more powerful CPU/GPU designs.
Of course it matters it’s just not necessarily the most important factor. The same developer will always be able to achieve more with more processing power though.
I would say it's main purpose is to alleviate the rising cost of VRAM.
To be able to have dense complex highly detailed assets, that requires space, unless you're generating those values in real time, in which you'd then need compute and bandwidth. Because space is so expensive and you require exponentially more memory as texture sizes increase to 4K and beyond, memory cost scaling is no longer a cost efficient solution. So improving I/O resolves that problem dramatically, especially if you've run out of memory, and this limits design potential.
In a perfect world, you'd solder the game directly onto your GPU
To serve more compute means you require more bandwidth; I/O or even having more VRAM doesn't increase bandwidth so we've still got challenges there on how far we can realistically push compute while being cost efficient.
All we're really discussing right now is which bottleneck is worse, and for this generation the chosen bottleneck to alleviate is the amount of VRAM, and vendors are doing so by improving I/O speeds dramatically since it's more cost efficient than taking on 32+ GB of VRAM. At least for now. I mean eventually something must give and we will need to move to 32 GB of VRAM next generation, as developers will naturally find a way to fill out all of this memory so its clear eventually developers will require more memory on tap to keep going further; but alleviating I/O should come first.
They made the right choice imo. The I/O change should be a dramatic change in asset quality. I think that's where the generational leap will be.
I think he talks about Unreal Engine 5 and the virtualisation of geometry. The virtualisation of geometry in Unreal Engine 5 transform the data on disk and I suppose this is cheaper to render than if it used a traditional method but it needs a fast storage to transfer data from the disk to memory.
Last edited:
Yes so the point here is twofold:
- game development is highly art constrained, ie getting good 3D models and textures into your games becomes ever more expensive as detail levels increase. So a lot of the innovations you see are as much focussed on improving the asset pipeline as anything else. In that light, being able to create your assets at the highest details and testing your levels with that saves a lot of time. Now you can postpone or even get rid of manual LOD management and just focus on the art, and have UE5 compress the assets automatically later based on possible player movement through the game. The same goes for lighting, something I think the DF video on Metro Exodus Enhanced shows well - no longer do you have to worry about placing lights all over the place to mimic realistic lighting, mess around with shadow maps, and bake lighting into your textures. All of these help immensely with the speed at which you can produce content in games, and I am willing to bet it will also make things like photogrammetry etc even easier to use.
- thanks to streaming geometry from SSD, you can keep much higher details of geometry in your games than you can fit in memory, simply because more detail can be streamed in when needed. And because so much can be streamed straight from SSD into the GPU, both memory use and bandwidth requirements go down as well.
UE5 will not be the only engine seeing these kinds of improvements, as I think this actually reduces complexity to the point where it is relatively easy for engines to move here, or even write new engines, especially for a situation like Metro where you choose to target the new consoles and RT capable GPUs. Development of completely new rendering engines will also get a boost, although they will always have to solve the art creation bottlenecks.
The bottleneck for now will be the commercial decision of how many systems will be able to run your game. We
- game development is highly art constrained, ie getting good 3D models and textures into your games becomes ever more expensive as detail levels increase. So a lot of the innovations you see are as much focussed on improving the asset pipeline as anything else. In that light, being able to create your assets at the highest details and testing your levels with that saves a lot of time. Now you can postpone or even get rid of manual LOD management and just focus on the art, and have UE5 compress the assets automatically later based on possible player movement through the game. The same goes for lighting, something I think the DF video on Metro Exodus Enhanced shows well - no longer do you have to worry about placing lights all over the place to mimic realistic lighting, mess around with shadow maps, and bake lighting into your textures. All of these help immensely with the speed at which you can produce content in games, and I am willing to bet it will also make things like photogrammetry etc even easier to use.
- thanks to streaming geometry from SSD, you can keep much higher details of geometry in your games than you can fit in memory, simply because more detail can be streamed in when needed. And because so much can be streamed straight from SSD into the GPU, both memory use and bandwidth requirements go down as well.
UE5 will not be the only engine seeing these kinds of improvements, as I think this actually reduces complexity to the point where it is relatively easy for engines to move here, or even write new engines, especially for a situation like Metro where you choose to target the new consoles and RT capable GPUs. Development of completely new rendering engines will also get a boost, although they will always have to solve the art creation bottlenecks.
The bottleneck for now will be the commercial decision of how many systems will be able to run your game. We
you act like new consoles only have a new SSD tech and nothing else, fact is if the PS4 add the same tech it would achieve better results than just doubling it's TFLOPS number. and yes this time the new data rates are a much bigger deal than the raw TFLOPS count in these new consoles. Just look at the PS4 pro, it's not doubling the details on screen over the PS4 even in games with a "detail mode" at the same res as the standard PS4.
GPUs have experimented also with plugging in an SSD straight into the GPU instead of having to go through RAM and the CPU, for I hope now obvious reasons. @PSman1700 you are mistaking the fact that many developers have provided input to console designers and EU5 developers alike, what their bottlenecks are, and the console and graphics engine designers have worked on how to alleviate these bottlenecks.
Everything particularly in a console is always a balancing act. The basic formula depends on where things will go and will always be tricky, but it is not just compute, or the CPU, or the amount of VRAM, or the RAM bandwidth, or the storage speed, or the speed of game (primarily art) development, it is everything combined. All of these can be bottlenecks as we have seen in the past, and many developments you see now show different approaches to solving these bottlenecks. Solve just one of these, like increasing the CU count, and the next will appear, like not being able to feed the CU s fast enough or render the output from the CUs fast enough, or not being able to provide enough mesh data to render in the first place, or not being able to render textures fast enough.
What you see in the consoles is the best possible balance they could find to raise the baseline performance as much as possible within the constraints of price (299 for series S, 399 for PS5 digital, 499 for disc based versions series X and PS5), and guesses on demands from current, past and future games and game engines.
This is not easy, to put it mildly, and there are definitely different routes that can be taken, depending on priorities. For instance if you think towards the future only, textures could become irrelevant and be replaced completely with geometry. If you design only for that, then you can get rid of any hardware that helps with that. Instead of a fast SSD, you could just use a lot of RAM - my PC is about 200 euro away from 64GB+6GB of VRAM, the much slower DDR4 could act as a cache for the harddrive. Etc.
But right now 64GB of RAM is about 400 euros, and that is the price of the digital PS5 in its entirety.
Everything particularly in a console is always a balancing act. The basic formula depends on where things will go and will always be tricky, but it is not just compute, or the CPU, or the amount of VRAM, or the RAM bandwidth, or the storage speed, or the speed of game (primarily art) development, it is everything combined. All of these can be bottlenecks as we have seen in the past, and many developments you see now show different approaches to solving these bottlenecks. Solve just one of these, like increasing the CU count, and the next will appear, like not being able to feed the CU s fast enough or render the output from the CUs fast enough, or not being able to provide enough mesh data to render in the first place, or not being able to render textures fast enough.
What you see in the consoles is the best possible balance they could find to raise the baseline performance as much as possible within the constraints of price (299 for series S, 399 for PS5 digital, 499 for disc based versions series X and PS5), and guesses on demands from current, past and future games and game engines.
This is not easy, to put it mildly, and there are definitely different routes that can be taken, depending on priorities. For instance if you think towards the future only, textures could become irrelevant and be replaced completely with geometry. If you design only for that, then you can get rid of any hardware that helps with that. Instead of a fast SSD, you could just use a lot of RAM - my PC is about 200 euro away from 64GB+6GB of VRAM, the much slower DDR4 could act as a cache for the harddrive. Etc.
But right now 64GB of RAM is about 400 euros, and that is the price of the digital PS5 in its entirety.
The PS4 Pro example doesn't really apply as devs have never tried to target it as a baseline. It's also constrained by more or less the same CPU and RAM.
I think most people will agree that the best example we have of real time graphics to date is the UE5 demo. And while we know the SDD's have enabled the Nanite technology (although the information we have puts the streaming requirements in the 100's MB/s rather than GB/s), we also know the rendering cost of that and Lumen would make it impossible on say a PS4 level GPU.
So both the increase in GPU performance and the improved IO systems are important to next gen and I don't think it's right to assume one is much more important than the other as it depends on what goals the developers have. Or put another way, if we ask which would create more impressive looking games; a PS4 with the PS5 IO or a PS5 with an HDD, I don't think the answer that immediately jumps out as correct is the PS4.
This is Unreal Engine we are talking about
1/ Import High detail stuff in Unreal Engine (with highend PC)
2/ Go on the round the world trip
3/ come back tanned from your round the world trip
4/ look at PC, -> 12% imported
I havent tried UE5 (which is just a rebrand of UE4.2X ) but lets hope the improved their asset management by several orders of magnitude
EDIT: - To clarify I am in totally agreement with Arwin, the less artists have to spend time optimizing/changing the original data the better. I just fear the thought of importing models 10 or 100 times greater data than what we deal with now
Last edited:
As above with the PS4P this is not a good example because those PC's have almost never been targeted as a baseline. Or where they have been targeted, the results have been well beyond what the last gen consoles have been capable of. Take Flight Simulator 2020 for example. It'll happily run off an HDD. Or Metro Enhanced. Both are clearly well in excess of that the last gen consoles are capable of but don't require an SDD.
The SDD's will bring new opportunities as well, and are certainly important, but you're being far too dismissive of the increased GPU capabilities.
i'm not dismissive i just think this time the new fast data reading / trasnfert speeds are a bigger deal than the increase in GPU, and it will leverage what's available on PCs too.
if anything, past years have been too dismissive about data speeds in favor of raw GPU power. It's time to balance things more.
if anything, past years have been too dismissive about data speeds in favor of raw GPU power. It's time to balance things more.
So do you believe that a PS4 with the PS5 IO would offer superior results to a PS5 with an HDD?
This explains. I'm not claiming anything about Most Important Factor. Consoles are sum of all of their parts. Would one rather have hdd on ps5 and 32GB ram for example? It feels like ssd is turning out to be key part of enabling "next gen" experiences. Extreme streaming in r&c, spiderman, unreal5, minimizing load times via hw and dev focus on optimization... There are many ways to create games. Same solution isn't optimal for everything.
My point about tflops was that tflops alone doesn't create great game. Someone has to put those tflops into good use. For example new resident evil doesn't benefit from tflops in pc side. It's full of artifacts no matter what pc hw one uses. Game quality and graphics quality comes much more from the team doing the game than having 10tflops or 20tflops available.
Last edited:
Similar threads
