Is 4GB enough for a high-end GPU in 2015?

silent_guy said:
I don't think we'll ever see nvlink in a product that matters for gaming, which means: an Intel or AMD CPU. But even if it happened, as you point out, it'd still be a small fraction of the local DRAM.
Click to expand...
Is nvlink is only for workstations or something? PCIe 4.0 is supposed to be double the bandwidth of 3 coming in @~30GB/s, dual channel DDR4 @ 3000 would saturate that at about 50% load. That's not bad, 30GB/s is about 25% of a single HBM stack bandwidth right. I don't know is 512MB/frame enough?
 
Good review to reference: https://www.techpowerup.com/reviews/AMD/R9_Fury_X/

Far Cry 4 and Assassin's Creed Unity seem to like the Fury X plenty, with no particular bottleneck for amount of ram. Other games, such as DA Inquisition, like the Nvidia Titan/980ti better, but don't appear to be bottlenecked by RAM either (390x v 980 non TI). Shadow of Mordor, no difference between 4gb Fury and 6gb 980ti even at 5k: http://www.guru3d.com/articles_pages/amd_radeon_r9_fury_x_review,15.html . Grand Theft Auto 5, no difference between 980 (4gb) and 390x (8gb) between 1440p and 4k: http://www.guru3d.com/articles_pages/amd_radeon_r9_fury_x_review,15.html

So, with multiple reviews, from multiple sites, going on multiple games, all showing that 4gb isn't a bottleneck right now for 4k, the same conclusion I mentioned earlier is easy to see. Framebuffers simply aren't enough to push the memory limits. IF console makers reliably free up ram (no dynamic freeing up of ram, you're not putting framerate critical resources into ram you may or may not have) then 4gb might not be enough in the future. But for now, and if the status quo is maintained, there is absolutely no evidence at the moment that 4gb is going to restrict you.
 
Infinisearch said:
Is nvlink is only for workstations or something?
Click to expand...
Right now, there have been images about nvlink connecting multiple GPUs and connecting GPUs with IBM PowerPC chips.

I don't think there is any upside for Intel to provide a high BW interface between its CPUs and its biggest competitor in the HPC accelerator space. They'll rather sell their own HPC chips themselves.

PCIe 4.0 is supposed to be double the bandwidth of 3 coming in @~30GB/s, dual channel DDR4 @ 3000 would saturate that at about 50% load. That's not bad, 30GB/s is about 25% of a single HBM stack bandwidth right. I don't know is 512MB/frame enough?
Click to expand...
If Pascal is going to support 1TB/s with 4 HBM2 stacks, then 30GB/s will be like sucking through a straw.

I would ignore configurations with HBM1: they are a proof of concept, but clearly not all that useful. I don't think we'll ever see another HBM1 chip other than Fiji, which doesn't even perform better than its GDDR5 counterpart. And I don't think we'll ever see 1 stack HBM2 solutions either... 2 stack HBM2 may be the first configuration that makes sense (since it exceeds a reasonable GDDR5 configuration in terms of BW), but I wouldn't even bet on that.
 
silent_guy said:
If Pascal is going to support 1TB/s with 4 HBM2 stacks, then 30GB/s will be like sucking through a straw.
Click to expand...
The question is will 30GB/s be enough for resource juggling. If in directx 12 you have developers use the graphics engine and the dma engine at the same time graphics would slow down if the DMA transfer sapped too much bandwidth. So the question becomes how much system and card link bandwidth is enough for resource juggling and how much bandwidth on the GPU side is enough to support simultaneous DMA and graphic/compute without slowing down to much. At least thats how I see it.

edit - well thats half the story at least.
 
Last edited:
Infinisearch said:
The question is will 30GB/s be enough for resource juggling. If in directx 12 you have developers use the graphics engine and the dma engine at the same time graphics would slow down if the DMA transfer sapped too much bandwidth. So the question becomes how much system and card link bandwidth is enough for resource juggling and how much bandwidth on the GPU side is enough to support simultaneous DMA and graphic/compute without slowing down to much. At least thats how I see it.
Click to expand...


Resource juggling is only used if necessary, it still drops some performance even if its not bandwidth constrained, the GPU still has some work to do to do resource juggling, and right now lots of games do do this specially MMO's with large worlds.

Lets take a 4k texture and you want to send that across the pci-e bus, its 32mb, after compression 8mb and you need to transfer it fast and of course you need the entire set. You will need a decent amount of bandwidth on the PCI-e bus if you don't want it to be a bottleneck.
 
Last edited:
Frenetic Pony said:
Good review to reference: https://www.techpowerup.com/reviews/AMD/R9_Fury_X/

Far Cry 4 and Assassin's Creed Unity seem to like the Fury X plenty, with no particular bottleneck for amount of ram. Other games, such as DA Inquisition, like the Nvidia Titan/980ti better, but don't appear to be bottlenecked by RAM either (390x v 980 non TI). Shadow of Mordor, no difference between 4gb Fury and 6gb 980ti even at 5k: http://www.guru3d.com/articles_pages/amd_radeon_r9_fury_x_review,15.html . Grand Theft Auto 5, no difference between 980 (4gb) and 390x (8gb) between 1440p and 4k: http://www.guru3d.com/articles_pages/amd_radeon_r9_fury_x_review,15.html

So, with multiple reviews, from multiple sites, going on multiple games, all showing that 4gb isn't a bottleneck right now for 4k, the same conclusion I mentioned earlier is easy to see. Framebuffers simply aren't enough to push the memory limits. IF console makers reliably free up ram (no dynamic freeing up of ram, you're not putting framerate critical resources into ram you may or may not have) then 4gb might not be enough in the future. But for now, and if the status quo is maintained, there is absolutely no evidence at the moment that 4gb is going to restrict you.
Click to expand...
Did you see links I posted?
 
silent_guy said:
My point is: if TechReport were such shilling biased paid-off baddies, they wouldn't have excluded Hawx, irrespective of whether excluding it was warranted or not.
Click to expand...

Many things could happend in so much years dude..

Just for precision, i have never tell they was biased, just this lineup of game and benchmark is really a bit strange for me. But that his choice, maybe just the game he love to play right now, at this moment.
 
lanek said:
You can remove Gta5 from the list, because it is not an TWITMP games, not a gamework titles, and not an Gaming evolved AMD.. it is maybe the more balanced PC games since a long time in respecting every hardware brand
Click to expand...
No GTA 5 is just like GTA 4, runs far better on NV hardware than on AMD. And it's NV sponsored title just like every other RockStar game.

every EA games are AMD ones .. the 2 bigger studios worldwide.
Click to expand...
Not any more they aren't, not with the death of Mantle. BF hardlines was sponsored by NV, and At E3, the new Mirror's Edge too.

The good thing ? they are not only a mess on AMD gpu's, they was offtly too a complete mess on Nvidia ones.
Click to expand...
Actually that's childish thinking, bad PC ports happen whether NV/GameWorks is involved or not, and they happen because studios give lesser care to the PC porting process than consoles, or out source the process completely to a different handler.

Also why are you bringing up the past? As of the picture was different back then? The fact of the matter is NV continues to dominate the PC gaming share today just as they did in the past.
 
Cynical as this may be: when 75% of GPU consumers buy Nvidia and the majority of new games run better on Nvidia for whatever good or bad reason, if the goal of a review is to educate how well a GPU will run a game for a consumer, it's only logical to include those games as well.

Different story if you want to use a review to judge the intrinsic performance of the architecture, of course.
 
silent_guy said:
I have no idea. ;)
Click to expand...
Neither do I, but I do find it an interesting question.
Razor1 said:
Resource juggling is only used if necessary
Click to expand...
This is true but this is a thread on if 4GB is enough. There are only two option - more ultra fast expensive VRAM or more faster system RAM and GPU interconnect combined with resource juggling, virtual texturing, and/or slower texturing from system RAM. Well there is the third option that AMD suggested with video cards with a slower RAM as well but greater in quantity.
 
fury x vs. 980ti with vram counter


Pretty big differences in grid autosport and hitman absolution, however compared to titan x.


 
gamervivek said:
fury x vs. 980ti with vram counter


Pretty big differences in grid autosport and hitman absolution, however compared to titan x.


Click to expand...

How is it possible they got half the ram usage than Nvidia gpu's ?( in gamess who dont go over 3000mb not limited by the vram capacity )... Thoses games are not new, so, cant be a developper work. or a special codes in the games who differ for Nvidia and Fury. If ofc the report of vram usage is not bugged.
 
fellix said:
The driver is doing some intensive management and swaping in the background.
Click to expand...

It rejoin the explanation of AMD about it before the launch of Fury so... should not be the case for all games, but on thoses 3, its close to half the vram usage.

can it benefit too low end gpu's who have only 2GB of ram as example and struggle with vram usage, basically this will end as they have nearly 3,5gb available ? That could be a major change in some case ..
 
Last edited:
What I think is that this will reduce the performance difference between the R390/X and the R290/X cards even further, decreasing the reasons to buy a more expensive 390 instead of a cheaper and overall faster 290X.

In the end, with DX12 multi-adapter allowing for memory pooling between GPUs and the driver sending less data to the GPU, it seems to me that getting two R9 290X 4GB will be the most cost-effective setup for driving DX12 games in high resolutions and VR helmets.

Which is why I ordered a second 290X for my desktop during the weekend :D
 
You must log in or register to reply here.
Back
Top