PC system impacts from tech like UE5? [Storage, RAM] *spawn*

[Shifty Geezer]

We don't know yet. It may be that devs adapt to resources available and implement the solutions that best fit the hardware profiles. The main reason to think IO is important is Sweeney said it was. PS5 being God-tier doesn't mean XBSX and future PC won't be though, as its an undefined term.

 

[iroboto]

We don't know yet. It may be that devs adapt to resources available and implement the solutions that best fit the hardware profiles. The main reason to think IO is important is Sweeney said it was. PS5 being God-tier doesn't mean XBSX and future PC won't be though, as its an undefined term.

If the topic is respective of UE5; they mentioned the ability to handle textures is done through their VT system. The mesh data seems to be unknown. Shouldn’t that be where the emphasis of our research be instead of just guessing the amount of bandwidth required ?

 

[Shifty Geezer]

We've tried a little speculation on what they could be doing, but it's complete guesswork. That part of the discussion really needs to wait on Epic to explain Nanite. TBH most of all this discussion needs to wait on a lot of info.

 

[zupallinere]

Your looking at $400 for a 2TB ssd with 5gb read speeds. I'm a believer that prices will come down but I'm not sure how fast for the higher capacity higher speed drives.

So PCs are looking at a deficit regarding I/O for a decent price ?? So much for Directstorage API fixing this

 

[eastmen]

So PCs are looking at a deficit regarding I/O for a decent price ?? So much for Directstorage API fixing this

Its okay ps5 is limited to 16 gigs of total ram. You can get 64 gigs of ram for about $70 more than a 1TB 5gb/s nvme drive.

 

[zupallinere]

Its okay ps5 is limited to 16 gigs of total ram. You can get 64 gigs of ram for about $70 more than a 1TB 5gb/s nvme drive.

Well you were using fast 2TB so howsabout 140 added to the 180 bucks to get from 16 to 64.
added: sorry I wasn't reading as well as I should have. I shall leave my response to kill future hero-worship
So to make up for lack of I/0 you could indeed get 64gb if the game scales that way.

 

[eastmen]

Well you were using fast 2TB so howsabout 140 added to the 180 bucks to get from 16 to 64.

here is a 64gig kit of ddr 4 3200 https://www.amazon.com/Corsair-Veng...=1&keywords=64+gig+ddr4&qid=1592233217&sr=8-1

Here is a 1TB nvme drive that this 5gb/s https://www.amazon.com/GIGABYTE-PCI...ld=1&keywords=1tb+nvme&qid=1592234766&sr=8-35

Its a $100 price difference and it doesn't meet the greater than ps5 ssd that another poster said would be required to use in the ps5.

I wonder if its better to have a faster nvme drive vs a slower nvme drive and more system ram ? What about a lot of system ram , a lot more graphics ram and a slightly slower ssd ?

We will know when we see unreal engine 5 running on more hardware

 

[zupallinere]

Thankful

I wonder if its better to have a faster nvme drive vs a slower nvme drive and more system ram ? What about a lot of system ram , a lot more graphics ram and a slightly slower ssd ?

We will know when we see unreal engine 5 running on more hardware

Thankfully we have the ps5 (and Series X) driving interest in making a faster I/O important in PC gaming.

 

[pjbliverpool]

So PCs are looking at a deficit regarding I/O for a decent price ?? So much for Directstorage API fixing this

Depends what speed drive you purchase. I doubt having 6 priority levels vs 2 or 3 would allow a 5.5GB/s perform like a 7GB/s drive in all circumstances.

I was thinking about this earlier and I'll post more in the PC NVMe thread about it shortly but I would expect on PC, IO to GPU memory to be from system memory rather than SSD anyway, which is much faster and lower latency than the PS5 SSD. Streaming will still be required into system memory of course but with a much lower throughput requirement given the amount of data you're already caching in system and video memory. So the slower SSD speed in a PC may be a red herring anyway. Even if you had a faster IO from SSD to GPU than the PS5, you'd probably still be better off streaming from system memory.

 

[eastmen]

Thankful

Thankfully we have the ps5 (and Series X) driving interest in making a faster I/O important in PC gaming.

and thankfully we had the pc drive making faster I/O cheaper for consoles to finally adopt.

 

[manux]

Depends what speed drive you purchase. I doubt having 6 priority levels vs 2 or 3 would allow a 5.5GB/s perform like a 7GB/s drive in all circumstances.

I was thinking about this earlier and I'll post more in the PC NVMe thread about it shortly but I would expect on PC, IO to GPU memory to be from system memory rather than SSD anyway, which is much faster and lower latency than the PS5 SSD. Streaming will still be required into system memory of course but with a much lower throughput requirement given the amount of data you're already caching in system and video memory. So the slower SSD speed in a PC may be a red herring anyway. Even if you had a faster IO from SSD to GPU than the PS5, you'd probably still be better off streaming from system memory.

Using ram only works if data is in ram. Load times/popin inbound with slow drive and lot of ram approach. Game sizes are much bigger than ram sizes being used by gamers. Consider something like grand theft auto next and what kind of interiors cars/buildings could have if you can swap in something like 8GB of data per second(raw rate+decompression). It becomes much more of install size and how much effort can be put into content. You could go as far as the imaginary ants being rendered in full detail in an openworld game. Fast streaming opens up all kinds of possibilities thanks to being able to pull in huge amount of asets in short time.

 

[pjbliverpool]

Using ram only works if data is in ram.

Well yes that goes without saying. So you pre-load into RAM at the start of the game in an initial loading screen and then stream into RAM from there on in. The streaming rate into RAM where say you have 28GB allocated to a game cache along with a 16GB GPU would be far less than required on a console because your would be able to keep much more re-usable data in RAM rather than having to keep swapping it out and re-retrieving it again from the SSD.

Load times/popin inbound with slow drive and lot of ram approach.

This would require a longer initial loading period but would result in less pop in because you are now streaming data into GPU memory from system RAM which is a faster solution that streaming it in from the PS5's SSD.

Game sizes are much bigger than ram sizes being used by gamers.

Yes but game sizes are still only going to be in the 100-200GB (uncompressed) range across the entire game/all environments. That means at any given time you're not going to need instant access to anywhere near that much data. The assets used in the current game environment for example may amount to more than the consoles 16GB RAM but would fit into 32GB of system memory so can be entirely streamed from there. Or if you need access to 64GB of assets but have half of that stored in system memory you only need to stream half of what you would have had to if you had non cached in system memory, thus bandwidth requirements to SSD are much lower.

 

[manux]

Well yes that goes without saying. So you pre-load into RAM at the start of the game in an initial loading screen and then stream into RAM from there on in. The streaming rate into RAM where say you have 28GB allocated to a game cache along with a 16GB GPU would be far less than required on a console because your would be able to keep much more re-usable data in RAM rather than having to keep swapping it out and re-retrieving it again from the SSD.



This would require a longer initial loading period but would result in less pop in because you are now streaming data into GPU memory from system RAM which is a faster solution that streaming it in from the PS5's SSD.



Yes but game sizes are still only going to be in the 100-200GB (uncompressed) range across the entire game/all environments. That means at any given time you're not going to need instant access to anywhere near that much data. The assets used in the current game environment for example may amount to more than the consoles 16GB RAM but would fit into 32GB of system memory so can be entirely streamed from there. Or if you need access to 64GB of assets but have half of that stored in system memory you only need to stream half of what you would have had to if you had non cached in system memory, thus bandwidth requirements to SSD are much lower.

I think this approach can work and is more of the traditional pc way. Hack around limitations through brute force. Most amazing trick being installing games into ramdisk.

But if you consider following. Does it make more sense to allow fetching content roughly on demand(1s ahead caching) versus creating system that requires load time and something like 30s ahead caching? And then the not so quick travel in open world games. To me the really fast ssd and enough ram is better approach than slower ssd/hdd and more ram. Fast storage solution can really be the thing that allows open world games look like piperun games does. Piperuns can easily cache and load and fill ram, open world not so much. I believe both microsoft and sony came to this conclusion. They could easily have gone no ssd/only small cache of ssd and more ram but they decided less ram and faster storage.

For consoles the game is going to be compressed on disk and decompressed on the fly. If pc has to go with less compressed assets that will only make things less optimal(larger installs, even more to load from disk).

 

[zupallinere]

and thankfully we had the pc drive making faster I/O cheaper for consoles to finally adopt.

Technically I would credit the server market for drive towards faster i/o but we have what we have now to play with in any case.

 

[eastmen]

I think this approach can work and is more of the traditional pc way. Hack around limitations through brute force. Most amazing trick being installing games into ramdisk.

But if you consider following. Does it make more sense to allow fetching content roughly on demand(1s ahead caching) versus creating system that requires load time and something like 30s ahead caching? And then the not so quick travel in open world games. To me the really fast ssd and enough ram is better approach than slower ssd/hdd and more ram. Fast storage solution can really be the thing that allows open world games look like piperun games does. Piperuns can easily cache and load and fill ram, open world not so much. I believe both microsoft and sony came to this conclusion. They could easily have gone no ssd/only small cache of ssd and more ram but they decided less ram and faster storage.

For consoles the game is going to be compressed on disk and decompressed on the fly. If pc has to go with less compressed assets that will only make things less optimal(larger installs, even more to load from disk).

depends on what you want to do. We aren't talking about traditional hardrives vs sonys 6gb/s drive. We are talking about nvme drives that are 2gb/s and faster with loads of ram

 

[pjbliverpool]

I think this approach can work and is more of the traditional pc way. Hack around limitations through brute force. Most amazing trick being installing games into ramdisk.

That's one way of describing it but I'd say that making full use of what is essentially (from the perspective of the PS5) a large, very high speed cache inbetween the SSD and VRAM is more than just a hack because it has the advantage of potentially providing to higher streaming throughput to the GPU.

But if you consider following. Does it make more sense to allow fetching content roughly on demand(1s ahead caching) versus creating system that requires load time and something like 30s ahead caching? And then the not so quick travel in open world games.

Why would the RAM approach result in slower world traversal or any other gameplay compromises? You're still ensuring the GPU is fed with data as fast as it needs (potentially faster than the PS5 is capable of). You're simply reducing the streaming speed required from SSD by initially caching more data in RAM and not needing to swap as much moving forwards. The amount of RAM required would be a product of how fast your SSD is and how much data the game would need to access within a given time frame.

The compromise is an initial loading screen but it shouldn't take too long to fill 44GB of memory from a 5GB/s drive.

To me the really fast ssd and enough ram is better approach than slower ssd/hdd and more ram. Fast storage solution can really be the thing that allows open world games look like piperun games does. Piperuns can easily cache and load and fill ram, open world not so much. I believe both microsoft and sony game to this same conclusion. They could easily have gone no ssd/only small cache of ssd and more ram but they decided less ram and faster storage.

From a console perspective it's likely the very high SSD option was cheaper than adding enough additional GDDR6 (assuming UMA is a must have) to make up for the slower SSD.

For consoles the game is going to be compressed on disk and decompressed on the fly. If pc has to go with less compressed assets that will only make things less optimal(larger installs, even more to load from disk).

I agree that one consequence of not having a hardware decompression unit could be a larger disk (but not memory) footprint for PC games. Better usage of bandwidth from SSD to memory is obviously another advantage of hardware decompression but that's already included as an assumption in these kind of discussions. i.e. we're assuming the effective streaming throughout of the PS5 is 8-9GB/s rather than 5.5GB/s. That's still much slower than system memory to VRAM.

 

[Shifty Geezer]

Why would the RAM approach result in slower world traversal or any other gameplay compromises?

It depends on the game, but take for example something like R&C. Let's say there are 12 worlds each consisting of 10 GBs, for a 120GB game. As you play each level, the PC version would only need to load 10 GBs which won't take long. But when you start dimension hopping arbitrarily, on PS5 you access that data at any point on the SSD whereas on PC, half of it won't be in RAM (if you can't account for prefetching) and you'll have to wait until its copied into RAM before you can hop there.

So long as a game cannot traverse between 40 GBs of data at a time, the 64 GB PC would be able to prefetch effectively. If a game can hop anywhere around the 120 GB data, the 64 GB caching system won't work.

For another example, let's grab an Elder Scrolls. Let's say each vicinity has its own art style and vases and pots unique to an area. When you move to a new area, the PC can load all its assets to RAM. If you travel there over land, it can stream them in seamlessly, and if you teleport, a short teleportation animation could cover it up. But what if you grab a vase from one area and place it in another? Now you need to fetch data from all across the storage. Start moving and mixing assets and suddenly your cache might not work.

Worst case though, you may suffer some pop-in. It might be workable overall.

 

[manux]

Why would the RAM approach result in slower world traversal or any other gameplay compromises? You're still ensuring the GPU is fed with data as fast as it needs (potentially faster than the PS5 is capable of). You're simply reducing the streaming speed required from SSD by initially caching more data in RAM and not needing to swap as much moving forwards. The amount of RAM required would be a product of how fast your SSD is and how much data the game would need to access within a given time frame.

You are assuming all the data can be loaded into ram. What if this is not true? Consider gta next. Maybe each car interior is big enough that you cannot fit all of them into ram. Would you limit the amount of unique car interiors or endure random load time/pop up effect when entering car?

Consider entering buildings. Is every building cached in ram or would there be random hitch if building which wasn't in cache is accessed? Would you limit asset sizes or perhaps recycle same textures to overcome?

Consider fast travel. You might jump into completely different area of the game world which is not yet cached in ram.

Of course you can try to be super smart and cache and preload like crazy but is that best use of developer/artist time when you could just rely on streaming ~roughly in time.

I believe end game is fast storage. While it's not yet baseline it will be. How fast it needs to be who knows. Probably any decent nvme ssd with proper software support is enough to work but then higher end solution could be better.

 

[Shifty Geezer]

You are assuming all the data can be loaded into ram. What if this is not true? Consider gta next. Maybe each car interior is big enough that you cannot fit all of them into ram. Would you limit the amount of unique car interiors or endure random load time/pop up effect when entering car?

Consider entering buildings. Is every building cached in ram or would there be random hitch if building which wasn't in cache is accessed?

You don't have to cache every single thing. A game isn't going to be using more than, say, 16 GBs of assets at any moment, and there's only so much change that you can create in typical play. In something like GTA, each city block could be 5 GBs of unique assets, and you'd load the city blocks around the player and a few in advance - this is the basis of streaming game content we already use. With RAM of the order of half your game data size, you can cache the vast majority of data and be pretty confident whatever you need will be available from RAM rather than disk for most titles that enforce a linear progression, whether that's driving through the streets or working through a story. It only become an issue when the entire dataset can referenced by the player at any given moment. At the point, the player will be referencing data present in the RAM cache and not present, and then you are stuck with cache transfer speed. A game designed around this probably won't suffer any, but a game designed for direct data access on consoles will need work to fit a small RAM cache.

 

[manux]

I wonder how the pc statistics for nvme ssd's are. i.e. once directstorage is here how big the customer base supporting it will be?

It would also be interesting to know if someone has idea how big part of pc users really have 32GB or more memory? Is this use ram more academical thing and (sata or nvme) ssd is reality or opposite? If people have to go upgrade anyway it wouldn't be that expensive to have some basic 1TB nvme ssd at around 2GB/s read speed. That should be plenty fast to work well and avoid any shortcomings of use more ram and somehow preload/cache more smartly approach.