• Register
  • TOS
  • Privacy
  • @NeoGAF

cobragt4001
Banned
(06-16-2013, 03:36 PM)
Only reason I decided to make this topic is I would like to know what actually keeps us from gaining extra vram when we decide to go sli or crossfire by getting another card. I understand why as of now why we can't double up on ram with videocards the way they are and how sli and crossfire works, one gpu renders even frames, the other renders odd frames and the vram is simply cloned. What would AMD or nvidia have to do to actually make multigpu setups actually double up, or triple up in the total vram pool? Would pci 3 be fast enough to actually make multigpus double up in vram or is it software that was to be changed or a combination of needed software and hardware? I thought it would be interesting to discuss this
ColdBlooder
Banned
(06-16-2013, 03:38 PM)
ColdBlooder's Avatar
Not sure but if i remember correctly, in SLI/Crossfire the ram is mirrored, not doubled. So with 2 3gb cards in crossfire you still have 3gb for games. Just both cards having the same content in the memory.

If im wrong, feel free to correct me though!
Sethos
Banned
(06-16-2013, 03:40 PM)
Sethos's Avatar

Originally Posted by ColdBlooder

Not sure but if i remember correctly, in SLI/Crossfire the ram is mirrored, not doubled. So with 2 3gb cards in crossfire you still have 3gb for games. Just both cards having the same content in the memory.

If im wrong, feel free to correct me though!

Well, that's kinda what he is saying, he's just asking if it was possible to have it double / how it could be possible.
JaseC
gave away the keys to the kingdom.
(06-16-2013, 03:40 PM)
JaseC's Avatar

Originally Posted by ColdBlooder

Not sure but if i remember correctly, in SLI/Crossfire the ram is mirrored, not doubled. So with 2 3gb cards in crossfire you still have 3gb for games. Just both cards having the same content in the memory.

If im wrong, feel free to correct me though!

You just read the thread title and not the OP, didn't you? ;)
SneakyStephan
(06-16-2013, 03:41 PM)
SneakyStephan's Avatar
Because then you wouldn't double the bandwidth like you do now and you'd be bandwidth starved at higher resolutions/framerates (which is why people crossfire/sli in the first place)

Older gpus like the voodoo sli cards I think rendered half of each frame (one horizontal line in turns I believe).
Which then also had the advantage of not having uneven frametimes (microstutter) or extra frames of vsync input lag.

So either you are bandwidth starved or you get the garbage laggy uneven framerate crap that you get today, sli is just a bad idea full stop.
n0n44m
Member
(06-16-2013, 03:41 PM)
n0n44m's Avatar
err pretty straightforward

if the full RAM pool would be available to both GPUs on both cards, then the GPU on card 1 would need to access the RAM on card 2 over the PCI-E lanes, which is infinitely slower than having the RAM on the PCB of the card itself and would thus completely destroy GPU performance

theoretically a super low latency, high bandwidth SLI/Xfire bridge could help, but that seems about as realistic as low latency Cloud Processing ...
dragonelite
Member
(06-16-2013, 03:42 PM)
dragonelite's Avatar
You need a fast shared pool of ram maybe 3D ddr4 is the solution for that for what you want OP.
Complete mobo redesign.

Dont really understand OP is it because of the 8GB of gddr5?
Because i don't see devs using that on the pc ports given how the X1 works with ddr3.
Bandwidth also need to be shared between the cpu and gpu on ps4.
GPU already have faster bandwidth then the ps4. Devs just need to make use of a streaming engine for the pc port.
Last edited by dragonelite; 06-16-2013 at 03:45 PM.
ColdBlooder
Banned
(06-16-2013, 03:43 PM)
ColdBlooder's Avatar

Originally Posted by Sethos

Well, that's kinda what he is saying, he's just asking if it was possible to have it double / how it could be possible.

Well i dont think its possible cause every gpu renders half of the same frame so they basically need to have the same data to access too. And also a big issue:

Latency.

If the gpu1 needs something from ram pool 2, the latency and "wait" for the data to arrive will propably slow the performance alot.
Relix
he's Virgin Tight™
(06-16-2013, 03:44 PM)
Relix's Avatar

Originally Posted by dragonelite

You need a fast shared pool of ram maybe 3D ddr4 is the solution for that for what you want OP.

Dont really understand OP is it because of the 8GB of gddr5?
Because i don't see devs using that on the pc ports given how the X1 works with ddr3.
Bandwidth also need to be shared between the cpu and gpu on ps4.
GPU already have faster bandwidth then the ps4. Devs just need to make use of a streaming engine for the pc port.

Where the hell is he talking about PS4 or XB1?
Sethos
Banned
(06-16-2013, 03:44 PM)
Sethos's Avatar
It all just boils down to inefficient RAM bandwidth if you were to employ any solution that would double the RAM, due to the 'expansion' design of SLI / Crossfire.
Durante
A Deadly Premonition hit his Dark Soul like a bolt of Lightning: "I can make their games better."
(06-16-2013, 03:48 PM)
Durante's Avatar

Originally Posted by cobragt4001

Only reason I decided to make this topic is I would like to know what actually keeps us from gaining extra vram when we decide to go sli or crossfire by getting another card. I understand why as of now why we can't double up on ram with videocards the way they are and how sli and crossfire works, one gpu renders even frames, the other renders odd frames and the vram is simply cloned. What would AMD or nvidia have to do to actually make multigpu setups actually double up, or triple up in the total vram pool? Would pci 3 be fast enough to actually make multigpus double up in vram or is it software that was to be changed or a combination of needed software and hardware? I thought it would be interesting to discuss this

It simply doesn't really work. Actually, the more we get into complex, custom GPU compute programs, the less traditional SLI or crossfire (with driver-managed card-level parallelism) makes sense. It has always been a huge waste of resources, but now you could be looking at adding lots of per-game inter-frame dependencies into the mix.

The good news is, with the increase in GDDR5 densities, it's really not that hard to get more than sufficient quantities of memory onto a single card.
miladesn
Member
(06-16-2013, 03:48 PM)
miladesn's Avatar
well, it would probably be possible with current system too, problem is to specifically distribute rendering and allocated memory between the cards in a game by game basis by the developer, then there is bandwidth issues and syncing issues between the two cards, it's probably a whole lot of extra work than just mirroring the memory and do some kind of half frame rendering on each, and only a very small part of the players would benefit. It's like this because it's straight forward pretty much.
HitCtrlAltDel
Junior Member
(06-16-2013, 03:49 PM)
HitCtrlAltDel's Avatar
It's pretty simple why they can't just double VRAM pool by adding another card. Because the way computer graphics work, there are a lot of unknowns as the scene is being rendered and choosing which graphics card will need a particular resource is impossible or too costly. For example: when a vertex is being transformed and placed in the world, until that transform operation happens it is unknown where on the screen it will be placed. You wouldn't know what video card to give it too. You would also run into issues with resources being needed on both cards: when primitive assembly happens and the rasterizer chooses pixels on the screen for the triangle, if all vertices are not on the same card then it would have to reach across the pcie bus to get the resource.

These are a couple reasons simplified, hopefully it answers your question.
Sethos
Banned
(06-16-2013, 03:51 PM)
Sethos's Avatar
Wait, did that ColdBlooder guy get banned ... :P
Durante
A Deadly Premonition hit his Dark Soul like a bolt of Lightning: "I can make their games better."
(06-16-2013, 03:53 PM)
Durante's Avatar

Originally Posted by Sethos

Wait, did that ColdBlooder guy get banned ... :P

It's been a while since I've seen the "not-reading-the-op" rule in effect.
JaseC
gave away the keys to the kingdom.
(06-16-2013, 03:53 PM)
JaseC's Avatar

Originally Posted by Sethos

Wait, did that ColdBlooder guy get banned ... :P

You might say it was... cold-blooded.

*crickets*
2San
Member
(06-16-2013, 03:55 PM)
2San's Avatar
Interesting topic.

How does this work for single card sli/xfire solutions, like the GTX690/HD7990?

Do they double the ram for mirroring or don't they need to do that since both gpu's are on the same board?
Sethos
Banned
(06-16-2013, 04:02 PM)
Sethos's Avatar

Originally Posted by 2San

Interesting topic.

How does this work for single card sli/xfire solutions, like the GTX690/HD7990?

Do they double the ram for mirroring or don't they need to do that since both gpu's are on the same board?

"Single-card" SLI / Crossfire works exactly like normal, if it's two 2GB cards, it's still a 2GB card.
Rolf NB
Member
(06-16-2013, 04:02 PM)
Rolf NB's Avatar

Originally Posted by cobragt4001

Only reason I decided to make this topic is I would like to know what actually keeps us from gaining extra vram when we decide to go sli or crossfire by getting another card. I understand why as of now why we can't double up on ram with videocards the way they are and how sli and crossfire works, one gpu renders even frames, the other renders odd frames and the vram is simply cloned. What would AMD or nvidia have to do to actually make multigpu setups actually double up, or triple up in the total vram pool? Would pci 3 be fast enough to actually make multigpus double up in vram or is it software that was to be changed or a combination of needed software and hardware? I thought it would be interesting to discuss this

Short answer: nothing.


Interleaving memory across the cards would eliminate mirrored data. So every time the GPU reads a kilobyte worth of textures or whatevs, half of it would come from its own memory, and the other half from the "remote" GPU.

Memory interleaving is a great technique when all memory pools are the same speed. You effectively multiply bandwidth by the amount of channels with little if any drawbacks (all contemporary "dual-channel" memory controllers in CPUs work this way). And you still get to use 100% of the storage, with no duplication.

Something like this is definitely technically feasible on GPUs, but prohibitively slow. PCIe 3.0 still only transfers 16GB/s. Local memory is 10x+ faster. Interleaving a local pool and a pool accessed through PCIe would effectively slow down all GPU memory access to the slower of the two pools. You'd get the equivalent of saddling a modern GPU with DDR3-1066 on a 128 bit bus.
2San
Member
(06-16-2013, 04:07 PM)
2San's Avatar

Originally Posted by Sethos

"Single-card" SLI / Crossfire works exactly like normal, if it's two 2GB cards, it's still a 2GB card.

So the advertised 4GB(GTX690) and 6GB(HD7990), are still respectively 2GB and 3GB in practice?
Sethos
Banned
(06-16-2013, 04:07 PM)
Sethos's Avatar

Originally Posted by 2San

So the advertised 4GB(GTX690) and 6GB(HD7990). Are still respectively 2GB and 3GB in practice.

Oh no, if they are advertised as 4GB, it just means there's two 4GB cards stuck together in reality. Just like a normal SLI / Crossfire setup.
cobragt4001
Banned
(06-16-2013, 04:12 PM)
Wow, great replies and very technical. I had had this thought in my mind for a while about the gpus and memory and what sparked the thought wasn't the Xbox or ps4. When I first build my computer, I was ignorant of the whole sli or crossfire setup from a technical stand point, I just knew, another card, better performance. Before I added my second card I was under the assumption that my vram would increase until I read about how multigpus just clone the memory so you have the amount of vram you started with with one card.
n0n44m
Member
(06-16-2013, 04:12 PM)
n0n44m's Avatar

Originally Posted by 2San

So the advertised 4GB(GTX690) and 6GB(HD7990), are still respectively 2GB and 3GB in practice?

yeah a GTX690 will only have 2 GB of RAM available
2San
Member
(06-16-2013, 04:13 PM)
2San's Avatar

Originally Posted by Sethos

Oh no, if they are advertised as 4GB, it just means there's two 4GB cards stuck together. Just like a normal SLI / Crossfire setup.

Oh all right, but how do they fit that all on a single board?

I thought PS4 was so shocking, because they used 512MB ram blocks, since it wouldn't fit otherwise. I do understand the board of the PS4 is hardly comparable to a separate GPU board.

So that does that mean you could have single card GPU with 8GB with 256MB ram block?

Originally Posted by n0n44m

yeah a GTX690 will only have 2 GB of RAM available

Ah so confusing. X.x
Last edited by 2San; 06-16-2013 at 04:16 PM.
Sethos
Banned
(06-16-2013, 04:14 PM)
Sethos's Avatar

Originally Posted by n0n44m

yeah a GTX690 will only have 2 GB of RAM available

Wait, they advertise it as 4GB? Man, I didn't know that :|
n0n44m
Member
(06-16-2013, 04:16 PM)
n0n44m's Avatar

Originally Posted by Sethos

Wait, they advertise it as 4GB? Man, I didn't know that :|

well all respectable Dutch shops put "4 GB" in their technical description

hell if you look at the Asus 690 , it's actually called "GTX690-4GD5" ... wonder what the 4 refers too :)

edit: ha from the Asus site

ASUS GTX 690: dual-GPU power enters the 28nm generation

The world's fastest graphics card with a dual-GPU and beefy 4GB GDDR5 memory

where's the beef !? ;D
Sethos
Banned
(06-16-2013, 04:17 PM)
Sethos's Avatar

Originally Posted by n0n44m

well all respectable Dutch shops put "4 GB" in their technical description

hell if you look at the Asus 690 , it's actually called "GTX690-4GD5" ... wonder what the 4 refers too :)

edit: ha from the Asus site

Never even gave that a second thought, that's really scummy :/
2San
Member
(06-16-2013, 04:18 PM)
2San's Avatar

Originally Posted by n0n44m

well all respectable Dutch shops put "4 GB" in their technical description

hell if you look at the Asus 690 , it's actually called "GTX690-4GD5" ... wonder what the 4 refers too :)

edit: ha from the Asus site



where's the beef !? ;D

All right, I understand now thanks.