Nvidia to use GDDR5X memory with next generation GPUs

i dont know how much more ram they will throw on cards as ram gets faster.
there has to be a saturation point where more ram does not yield results proportional to the costs.
The slower the ram then the more of it you need. but a card with 4gb of vram working at 1/2 the speed of a different card with 2gb of vram (just ram speed gpu is the same) would be the same speed.
obviously this is not 100% correct but its close enough to be stated as such.

So there has to be a point where 4gb of ultra fast vram will be better than 8-10gb of slower ram. lets say the 4gb ram is 3x faster it will be able to off load and re populate 12gb of data in the same amount of cycles as the 10gb ram card was able to process 10gb.
So it is going to get to a point "pretty soon imho" where the amount of ram is not going to be as important as the speed of the ram. now that is not to say that more ram wouldn't be better. but there has to be a point where the extra ram is doing nothing other than holding data it does not need to hold at all, and it is giving no performance gains at all compared to having 2-4gb less ram on the card.

I honestly expect 8gb to be a saturation point for GPU's with 4gb cards performing just as well after memory speeds start getting faster.

edit
just thought id add this does not take in to account 4k. as i dont really expect that to be as big of a thing as some do. But i guess i could be wrong, but even then making the ram 2x faster is the same as adding twice as much memory. and if used properly dx 12 with sli/x-fire will take care of that any way. and id imagine people going for 4k gaming will be running 2 or more gpu's and because dx12 does not just ask both cards to que up the same things then you do have twice the ram to populate.
 
Last edited:
Yeah, I can't imagine we'd need more than 8GB of fast VRAM for many years. I'm not sure how much VR will demand.

I'm not sure what DX12 will bring with VRAM usage, but I heard it was down to game developers how the game orders the VRAM.
 
Yeah, I can't imagine we'd need more than 8GB of fast VRAM for many years. I'm not sure how much VR will demand.

I'm not sure what DX12 will bring with VRAM usage, but I heard it was down to game developers how the game orders the VRAM.

DX12 will influence ram requirements somewhat but mainly due to the fact that as more draw calls are being queued devs will then tend to store a little more in memory so instead of having to reprocess data it could already be there. Obviously it comes down to Devs and there ability to optimize and find the best balance but I don't think DX12 will drastically alter the amount of memory needed over previous DX APIs. I could be wrong and I wouldn't be surprised if Sub or SPS corrected me:p However the way I explained it seems logical enough to me^_^
 
Is there an expected timeframe on HBM 2?

Most games I've seen so far use about 3GB VRAM so there's still plenty of headroom left yet.
 
DX12 will influence ram requirements somewhat but mainly due to the fact that as more draw calls are being queued devs will then tend to store a little more in memory so instead of having to reprocess data it could already be there. Obviously it comes down to Devs and there ability to optimize and find the best balance but I don't think DX12 will drastically alter the amount of memory needed over previous DX APIs. I could be wrong and I wouldn't be surprised if Sub or SPS corrected me:p However the way I explained it seems logical enough to me^_^
I agree with you. It makes sense. What might change is how game developers and the software team for the GPU companies optimise for HBM. You probably saw the article about how a game developer recently decided to indefinitely delay the DX12 update for their game (I think it was ARK) as the drivers from both nVidia and AMD were not mature enough. That might be partially due to HBM in the Fury cards, though I doubt it plays that big of a part as there are many other GPU's without HBM.
 
DX12 will influence ram requirements somewhat but mainly due to the fact that as more draw calls are being queued devs will then tend to store a little more in memory so instead of having to reprocess data it could already be there. Obviously it comes down to Devs and there ability to optimize and find the best balance but I don't think DX12 will drastically alter the amount of memory needed over previous DX APIs. I could be wrong and I wouldn't be surprised if Sub or SPS corrected me:p However the way I explained it seems logical enough to me^_^


the only thing memory wize that could be better used with dx 12 IMO (that is well worth doing) is xfire/sli
right now both cards memory will be populated with exactly the same data. so 2x 4gb cards really only have 1x4gb of ram but theoretically twice as fast (theory never much pans out in practice mind you). with dx 12 both cards will be able to be populated with different data giving access to 8gb of ram at a single set speed.
how much better that will be for sli is not something im sure about but would probably benefit 4k.
 
I agree with you. It makes sense. What might change is how game developers and the software team for the GPU companies optimise for HBM. You probably saw the article about how a game developer recently decided to indefinitely delay the DX12 update for their game (I think it was ARK) as the drivers from both nVidia and AMD were not mature enough. That might be partially due to HBM in the Fury cards, though I doubt it plays that big of a part as there are many other GPU's without HBM.

Pretty sure that's Arks coverup. They are blaming AMD and Nvidia when even in DX11 that game is so horribly optimized:p
 
Pretty sure that's Arks coverup. They are blaming AMD and Nvidia when even in DX11 that game is so horribly optimized:p
That could be true actually. Some of the benchmarks I've seen have been really, really demanding. Like, you need to use medium settings with a 980ti at 1080p to surpass 60 FPS. I can't remember where I saw that, or whether it was for ARK or Ashes of the Singularity.
 
Back
Top