The new memory usage method on 30xx

[Unseul]

So watched a lot of the stream, but the bit that really grabbed me was actually how they're going to bypass the CPU.

Got a few questions on that.

First, surely AMD are aware of this? It seems like the sort of thing that if possible, could make an enormous difference to users, and if AMD aren't doing this with their next lot of cards, they're going to struggle I imagine?

Next up, does that mean older CPU's are suddenly going to be able to cope a lot better with these new cards? A huge amount of the load that the card would put on them has just been removed, meaning more left over to just get on running the game?

Third, would this have an effect on the amount of Vram actually needed? If you no longer need to go through a slow (relatively) process to get the information in the graphics memory, then you don't need to be able to store as much of it at once I would have thought? Swapping in and out quicker is generally going to be as good as simply having more space (unless literally everything you are looking at on your screen is maxing out the vram immediately).

Or am I just being daft and don't particularly understand any of this?

 

[Chuk_Chuk]

First, surely AMD are aware of this? It seems like the sort of thing that if possible, could make an enormous difference to users, and if AMD aren't doing this with their next lot of cards, they're going to struggle I imagine?

Considering that they worked on the development of the consoles, I would be shocked if they didn't have something similar in RDNA 2. Also they did have HBCC cache (is that what it was called?) on the fury cards so they are no strangers to it.

Next up, does that mean older CPU's are suddenly going to be able to cope a lot better with these new cards? A huge amount of the load that the card would put on them has just been removed, meaning more left over to just get on running the game?

It does remove some workload from the CPU but it is too early to say how much performance will be gained from removing this workload.

Third, would this have an effect on the amount of Vram actually needed? If you no longer need to go through a slow (relatively) process to get the information in the graphics memory, then you don't need to be able to store as much of it at once I would have thought? Swapping in and out quicker is generally going to be as good as simply having more space (unless literally everything you are looking at on your screen is maxing out the vram immediately).

From my understanding, after watching the cerny presentation.
VRAM holds the data for the things your looking at, as well as anything you may look at within the next "X" seconds. Where "X" is the amount of time it takes to load data from storage to VRAM. Reducing "X" does mean you can cache less data, for a fixed character movement speed.

They will also have the option of increasing the characters movement speed (think Ratchet and Clank when they hopping through the portals between different sides of the same map, or increasing the density of assets on the screen.

 

[LePhuronn]

First, surely AMD are aware of this?

Yes, AMD developed High Bandwidth Cache Controller 3 years ago and was on the Vega cards. This is also how the PS5's IO system works.

It seems like the sort of thing that if possible, could make an enormous difference to users...

If devs implement it and the drivers are stable.

...and if AMD aren't doing this with their next lot of cards, they're going to struggle I imagine?

I'd wager there is zero chance AMD will disregard the work they've done on consoles for the dGPU space. Clearly Nvidia have seen the potential of the console's IO system and developed RTX IO accordingly, there is no reason to think AMD won't counter by implementing something they already made 3 years ago.