Question Video Memory effective speed vs system module size?!

[FreeBee101]

If you get 192 bit bus with gddr7 and 28gb/s both 16/32/64 and 24/48 gb modules divide evenly into the thoughput. But 256 and 512 bit busses only 16/32/64gb modules match. Does this matter to the systems ram size?

Are 24 and 48 gb modules somehow fit to having 192 bus size. Or does this only matter for Vram on the GPU? Is there any reason to line up your system ram sizes to the gpu capability? Is there any weird gain for any applications?

I'm assuming not, but I thought I'd check. And what is the point if any to 24/48gb vs 16/32gb modules?

And when/if the zen6 with 12/24 core/threads comes out will there be a reason to get modules in 24/48gb sizes to match?

6000x32\8 = 24000mhz. Does that mean 24gb modules are good as the bandwidth equals the module capacity? Or does that not have any real world effect?

 

[Lutfij]

System ram and VRAM are two different platforms/languages/architectures. The memory on your GPU is managed by the GPU's architecture(die). The memory in your motherboard's ram slots are managed by the integrated memory controller that's located on the CPU's die, formerly known as the NorthBridge(if you're familiar with tech from 2 decades or more ago).

You buy ram according to your workloads/needs. If you're on a DDR5 platform, you look at tight latencied, dual channel DDR5-6000MHz(or slightly higher frequencied) ram kits. If you're on Ryzen, look for ram modules with AMD's E.X.P.O advertised on it. If you're on Intel, look for Intel's X.M.P advertised on it.

 

[FreeBee101]

I thought I had heard of gpu combined with system ram and it's for AI computers. I was thinking of shared/combine/coherent memory that is in the M2Ultra computer.

I think I was hoping that would be in desktops at some point and got the idea in my head. In fact is there any chance that will be coming to normal computers soon? I haven't heard anything yet.

https://en.wikipedia.org/wiki/Memory_coherence

https://stackoverflow.com/questions/36241009/what-is-coherent-memory-on-gpu

Apple introduces M2 Ultra

Apple today announced M2 Ultra, a new system on a chip (SoC) that delivers huge performance increases to the Mac and completes the M2 family.

www.apple.com

Either way I was hoping to do it to theme the system for fun. Or that this tech would be good for large scale video games also one day. I swear I saw an article saying this was coming to desktops by like 2028.

 

[SkyNetRising]

If you get 192 bit bus with gddr7 and 28gb/s both 16/32/64 and 24/48 gb modules divide evenly into the thoughput. But 256 and 512 bit busses only 16/32/64gb modules match. Does this matter to the systems ram size?
Are 24 and 48 gb modules somehow fit to having 192 bus size. Or does this only matter for Vram on the GPU? Is there any reason to line up your system ram sizes to the gpu capability? Is there any weird gain for any applications?
I'm assuming not, but I thought I'd check. And what is the point if any to 24/48gb vs 16/32gb modules?
And when/if the zen6 with 12/24 core/threads comes out will there be a reason to get modules in 24/48gb sizes to match?
6000x32\8 = 24000mhz. Does that mean 24gb modules are good as the bandwidth equals the module capacity? Or does that not have any real world effect?

Each and every sentence of that post is nonsense.
What have you been smoking?

I swear I saw an article saying this was coming to desktops by like 2028.

Integrated graphics use system ram for vram. That is nothing new.
Integrated graphics have been available for decades.

 

[FreeBee101]

They are working on a new version for AI stuff. I'm not sure if it's coming to normal desktop PC's in the next few year or not. It's more advanced than normal shared memory for integrated graphics.

https://www.reddit.com/r/LocalLLaMA/comments/16v7ais/question_will_there_ever_be_unified_memory_on_a/

 

[Eximo]

They are working on a new version for AI stuff. I'm not sure if it's coming to normal desktop PC's in the next few year or not. It's more advanced than normal shared memory for integrated graphics.

https://www.reddit.com/r/LocalLLaMA/comments/16v7ais/question_will_there_ever_be_unified_memory_on_a/

You might be thinking of DRAM cache, which has also already happened. Basically putting fast memory directly together with the CPU, but isn't as costly as building more SRAM. I wouldn't be surprised if NPUs would start needing their own memory/cache to operate more effectively, or if they would just further integrate to give those chips DMA.

Integrated DRAM or HBM onto the physical CPU package is also not new. AMD and Intel have both done this. (And once in the same chip)

We are rapidly approaching a time where socketed memory will become another memory layer removed from the CPU cores. CPU, L1, L2, L3/X3D/l4, DRAM cache/HBM, System Memory, Storage.