Digital Foundry Article Technical Discussion [2020]

Status
Not open for further replies.
The other 3.5 GB has capable bandwidth of 336 GB/s, which exceeds the entire bandwidth of Xbox One and One S, and is 75% the bandwidth of PS5.
yes, do you think you can use as vram at the same time ram from pool of 336gb/s and from 560gb/s without complication ?
 
Last edited:
huh yea. You either access at 560 or you access at 336.
ok so no problem for xsx gpu to access at same time 10GB at 560 and for example 1GB more at 336 ? I thougt it would be problematic becouse of this asymetric access and in practice always only 10 will be for gpu
 
ok so no problem for xsx gpu to access at same time 10GB at 560 and for example 1GB more at 336 ? I thougt it would be problematic becouse of this asymetric access and in practice always only 10 will be for gpu
No. You can’t access both pools at same time because you’d need separate memory controllers to do that. So on that front you are correct.
Whether it’s a problem is not really.
 
No. You can’t access both pools at same time because you’d need separate memory controllers to do that. So on that front you are correct.
Whether it’s a problem is not really.
so as I thought, my comment was because some is worried that xsx has 13.5gb and 3080 10 which is not exatly truth as xsx has also 10 for it gpu (btw small advantage for ps5 here)
 
so as I thought, my comment was because some is worried that xsx has 13.5gb and 3080 10 which is not exatly truth as xsx has also 10 for it gpu (btw small advantage for ps5 here)
The impact is unlikely to be as large as people are making out to be. Some of that memory will need to be reserved for game code, audio, etc. And further more if you’re seldom accessing it then the bandwidth is more than generous for minor needs.
I would say the main issues is asking developers to play a bit of Tetris to get things to fit as they expect.

this problem though; much larger on series s
 
Last edited:
On consoles you can rely on fast storage from SSD to GPU so you may well need a bit less RAM for the GPU.
 
On consoles you can rely on fast storage from SSD to GPU so you may well need a bit less RAM for the GPU.
Yep. Cerny explained this very well, when he talked about the amount of 'active' RAM used in games, and how next gen will see a much larger portion of the RAM being used for things we actively need on a short term basis, instead of wasting GBs of RAM storing data that we may not even need in the next few seconds of gameplay.
 
hm... so VRS Tier 2 netted them 5-12% savings.

hm... not quite a perfect 60fps. odd. Seems to be a bug between streaming and low latency optimizations? There's a fix in the works.
Not bad, the interesting thing with VRS is how aggressive you want to get with it. I suppose overly aggressive would damage the image quality.
Perhaps an issue that Dirt 5 is running into, they may need to pull it back.
 
hm... so VRS Tier 2 netted them 5-12% savings.

hm... not quite a perfect 60fps. odd. Seems to be a bug between streaming and low latency optimizations? There's a fix in the works.
Is the Tier 2 VRS what's causing the pop-ins?

The RT scene at 4min is also broken or at least looks broken to me.
 
Yep. Cerny explained this very well, when he talked about the amount of 'active' RAM used in games, and how next gen will see a much larger portion of the RAM being used for things we actively need on a short term basis, instead of wasting GBs of RAM storing data that we may not even need in the next few seconds of gameplay.

I think there's a question to be asked here around how much speed is enough to enable this kind of paradigm? Obviously more than an 80MB/s HDD otherwise it would have been possible last gen. But do you need more than a 600MB/s SATA SSD for example? Or more than a 3GB/s PCI3 NVMe?

You won't be constantly refreshing your VRAM at maximum speed from the SSD or else you'll consume your entire game content in seconds. So the real streaming speed is going to be much, much slower for a typical game install - limited not by IO throughput, but by game size. Even at 600MB/s you will consume a typical 100GB game in about 3 minutes so it seems to me that where the players movement through the game world is linear and can be predicted (i.e. non fast travel like situations) then unless you can progress through all the game content in under 3 minutes (for example an open world map where you can fly from one extreme of the map to the other in less than 3 minutes) then even 600MB/s would be sufficient to entirely remove the streaming bottleneck.

Obviously none of the above would apply to initial loading screens or fast travel type scenario's (the unpredictable ability to instantly move to any point in your game content) where more speed is always beneficial.
 
I think there's a question to be asked here around how much speed is enough to enable this kind of paradigm? Obviously more than an 80MB/s HDD otherwise it would have been possible last gen. But do you need more than a 600MB/s SATA SSD for example? Or more than a 3GB/s PCI3 NVMe?

You won't be constantly refreshing your VRAM at maximum speed from the SSD or else you'll consume your entire game content in seconds. So the real streaming speed is going to be much, much slower for a typical game install - limited not by IO throughput, but by game size. Even at 600MB/s you will consume a typical 100GB game in about 3 minutes so it seems to me that where the players movement through the game world is linear and can be predicted (i.e. non fast travel like situations) then unless you can progress through all the game content in under 3 minutes (for example an open world map where you can fly from one extreme of the map to the other in less than 3 minutes) then even 600MB/s would be sufficient to entirely remove the streaming bottleneck.

Obviously none of the above would apply to initial loading screens or fast travel type scenario's (the unpredictable ability to instantly move to any point in your game content) where more speed is always beneficial.
Regardless of the ultimate results, which will be very much based on a huge number of variables, I think the idea is that while now we essentially need to account for the next, say, 30 seconds of gameplay and store the data required for that in RAM, with the new SSDs we only need to have the next second or two, or whatever the number will be. And lowering this number more and more (say, half a second?) would theoretically allow us to have more and more 'active RAM', resulting in more detail on screen.

That's the idea, anyway. As I said, there are so many variables!
 
Regardless of the ultimate results, which will be very much based on a huge number of variables, I think the idea is that while now we essentially need to account for the next, say, 30 seconds of gameplay and store the data required for that in RAM, with the new SSDs we only need to have the next second or two, or whatever the number will be. And lowering this number more and more (say, half a second?) would theoretically allow us to have more and more 'active RAM', resulting in more detail on screen.

That's the idea, anyway. As I said, there are so many variables!

Oh yes I absolutely agree with the principle of the theory and if games had unlimited install size then it may be the case that we could indeed fill the entire 16GB VRAM with assets that will be consumed in the next 2-3 seconds for some insane visuals. However I think this potential will be limited by the size of the game content. So even though you could just store 1-2 seconds worth of gameplay in VRAM, in reality you'll still likely be storing several minutes worth (or say 30 seconds in any given direction in a open world game) and just streaming in beyond that as the player moves through the game world. Of course the faster SSD's provide the option to allow the player to zip from one side of the map to the other really quickly if the game design demands it but I see that being more of a fast travel scenario than a normal streaming scenario that amplifies VRAM size.
 
Status
Not open for further replies.
Back
Top