Multi-GPU Resource Pooling - DX12 Demo

[MEC-777]

 

PC gaming is about to see a rather drastic change and improvement in the way multi-GPUs are implemented. No longer will you be forced to pair dGPUs from the same architecture/family or even brand. The ability to combine any two or more GPUs and have their resources (GPU cores, Vram etc.) stacked and pooled is a thing and will be possible with DX12. Whether you combine Intel or AMD iGPUs with any dGPU or Nvidia and AMD dGPUs, it won't matter. Their resources can be shared for additional gaming performance. No longer will the iGPU in your system go to waste. This is awesome news for all of us. I know this has been known as a possibility with DX12 for some time, but this is the first time I've seen a demo of it and a game dev talking about how it can be implemented. 

 

The demo in the video is using one of AMDs latest APUs paired with an R9 290X - an unlikely and formerly thought of "unbalanced" combination. But now with DX12, we not only have reduced CPU overhead - allowing this CPU/dGPU combo to actually perform optimally, but we also have the pairing of the APUs iGPU added to the total GPU capabilities of this particular system for even more performance. As mentioned, this is not exclusive to AMD hardware, they just happened to be running it on an AMD system. With the latest APUs having such strong iGPU performance, it was a good platform to demonstrate the effectiveness this can have.

 

Currently, the way multi-GPUs are combined is that each GPU (in a dual-card config) takes turn rendering each frame. Some of the disadvantages with this is that both GPUs have to be nearly identical, run at the same clock speeds and both have to hold the same data in the Vram of each card resulting in redundancies and limitations on which cards can be used/combined. The new approach made possible with DX12 is that now the devs can have control over which aspects of the scene are being rendered by which GPU. So you can have the high-end dGPU still doing most of the heavy-lifting, rendering the majority of the scene, while the iGPU takes on the remaining tasks/aspects of rendering the scene. This means the dGPU doesn't have to render the entire scene by itself, only part of it, which as you can guess, will increase frame rates. 

 

As stressed in the video, it will be up to the devs of each game to optimize and sort out which aspects/tasks can be offloaded to an iGPU and how the work will be divided. One key thing to note is that this will allow for much better crossfire/SLI or hybrid dGPU combos to be possible as the GPUs will not be taking turns rendering each frame, but each frame will be rendered in concert with each GPU taking on specific elements of the scene. This will also allow for Vram stacking, so two cards with 4GB of Vram each will amount to a total of 8GB Vram available. One card can handle all textures, while the other handles lighting etc. The potential is there, but again, it will be strongly dependant on the devs to implement it.

 

Here's the demo video:

 

https://www.youtube.com/watch?v=9cvmDjVYSNk

[EarthboundHero]

What about 'crossfiring' Intel onboard with a dedicated GPU?

[Trik'Stari]

"GOOD NEWS EVERYONE!"

[MEC-777]

What about 'crossfiring' Intel onboard with a dedicated GPU?

 

In the video they say it should be possible with any iGPU (integrated/onboard GPU) and dGPU (discreet GPU) combination. It's not brand-specific and the only limitations with be how powerful each GPU is. 

[KemoKa]

AHhhhhhhhh I wish they'd release it for hardcore people who want to test the APIs! They could have a disclaimer saying this could brick your card and everything, it'd be fine! Crowdsource your testing, companies!

[gill2k]

We can finally mix and match different GPU's???

 

I really hope devs take advantage of this.  This is f'n awesome!

[mvitkun]

@MEC-777 second time that it's been demo'd

[RexinOridle]

And here I was thinking someone SLIFired some thing.

[Fetzie]

In the video they say it should be possible with any iGPU (integrated/onboard GPU) and dGPU (discreet GPU) combination. It's not brand-specific and the only limitations with be how powerful each GPU is.

And if the developer can be bothered developing with this in mind. The potential for extremely bad implementations of this is high

[Pvt. 8Ball]

I wonder if this will improve dual card frame latency, because as it is now, at 50fps with 2 cards, you're actually getting 25fps equivalent input latency, if not more.

[deviant88]

Ive been dreaming for something like this ever since i learner what an SLi/xfire is, and im still skeptical about it, companies are just retarded man,instead of advancing and creating insane new experiences using hardware that we already have in new ways they block it, like nvidia physx,just imagine every single game that has gpu physx and shared pool memory and offloads other stuff to iGpu or a 2nd gpu, i have a gtx 460 laying around and an hd 4600 igpu now i can only imagine the 3 combined.

 

Anyway i think the iGPU co-processing is still nonesense,i still see iGPUs useless and toxic,they need to go away from dedicated desktop pc's.

Think about it for a second theres no way an iGPU that uses slow DDR3 -4 RAM can keep up with 7GHZ gddr5 or HBM,its not even close in bus width to a gpu, and HBM with 1024 bit bus its going to get even more crazy.

Just let the iGPU's go we dont need em,all we need is to be able to use cross-vendor and cross-gen gpu's.

Just the tought of buying a future hbm amd card and offloading some stuff to my old gtx 670 makes me "wet" ,not having to throw away a good card.

[Master Disaster]

Fuck hype over SWBF, BAK or any other game.

This is real hype, can't wait to see games scale across multiple CPU cores and multiple GPUs no matter what brand!!!!!

Its only been 8 years since the first X86 based dual core CPU launched.

HYPE!!!!!

[AHaskin14]

so going to have a 770, 640, and 4600 powering through games.

[Bouzoo]

Inb4 PSU makers (and MBO to some degree) earning a lot of money.

[Humbug]

Great stuff. Something we can all cheer about regardless of fanboy allegiances.

 

These oxide guys always seem to be on the cutting edge of graphics APIs. First with mantle, then with DX12, then with Vulkan.

I hope their game will be as good as their tech LOL.

[The_Strict_Nein]

This makes sense on why Nvidia are so hyping NVLink

 

They need to make a version of SLI so much better than DX12 shared resources in order to encourage people to just buy Nvidia cards.

[Humbug]

They need to make a version of SLI so much better than DX12 shared resources in order to encourage people to just buy Nvidia cards.

What oxide games showed will work on nvidia too. They just demoed it on AMD because that's what they have got working at the moment. But it's not a vendor exclusive thing...

It's not really like SLI or cross-fire, it's about have a discrete GPU share the load with your onboard GPU.

[The_Strict_Nein]

What oxide games showed will work on nvidia too. They just demoed it on AMD because that's what they have got working at the moment. But it's not a vendor exclusive thing...

It's not really like SLI or cross-fire, it's about have a discrete GPU share the load with your onboard GPU.

 

The point I'm making is that with this, it makes SLI effectively redundant.

 

NVLink is supposed to be SLI on steroids and I'm guessing for Nvidia hardware better than this implementation. If Nvidia want to push you buying two new matched cards for SLI they need to offer tangible benefits to this DX12 implementation. NVLink is their plan.

[Dietrichw]

The point I'm making is that with this, it makes SLI effectively redundant.

 

NVLink is supposed to be SLI on steroids and I'm guessing for Nvidia hardware better than this implementation. If Nvidia want to push you buying two new matched cards for SLI they need to offer tangible benefits to this DX12 implementation. NVLink is their plan.

 

I don't see NVlink coming to consumer boards. The NVlink interconnect is made to replace PCIe but intel probably won't make it work with their chipset

[Victorious Secret]

This makes sense on why Nvidia are so hyping NVLink

 

They need to make a version of SLI so much better than DX12 shared resources in order to encourage people to just buy Nvidia cards.

 

NVLink is more geared towards HPC, not consumer use. 

[MEC-777]

I wonder if this will improve dual card frame latency, because as it is now, at 50fps with 2 cards, you're actually getting 25fps equivalent input latency, if not more.

 

Not really. Both cards are working in unison at 50fps. So your input latency is still at 50fps. 

 

With this new implementation, your input latency would improve by however much your frame rate improves by.

 

 

Anyway i think the iGPU co-processing is still nonesense,i still see iGPUs useless and toxic,they need to go away from dedicated desktop pc's.

Think about it for a second theres no way an iGPU that uses slow DDR3 -4 RAM can keep up with 7GHZ gddr5 or HBM,its not even close in bus width to a gpu, and HBM with 1024 bit bus its going to get even more crazy.

Just let the iGPU's go we dont need em,all we need is to be able to use cross-vendor and cross-gen gpu's.

Just the tought of buying a future hbm amd card and offloading some stuff to my old gtx 670 makes me "wet" ,not having to throw away a good card.

 

iGPUs will never go away because the majority of users don't need or use discreet graphics. 

 

This new implementation of combining multiple GPUs does not work the way you're describing. The Vram on the dGU and the system ram being used by the iGPU does not have to be the same speed/type etc. How it will works is what ever data is necessary for whatever task the iGPU is assigned, that data is loaded to the system ram only. This cuts down on redundant data copying/transfers unlike current multi GPU methods. The issues you described would not be a problem. 

[Waruteru]

Every time I see this subject and then the "depends only on the developers" I get frustrated...

 

I'm serious(this time), last years we saw dozens examples about not so committed developers and do you think that they will support this? I don't think so...

 

And AMD and Nvidia supporting developers to do these multi gpus combos? Wouldn't believe... Life can't be this good, the world is not fair...

[Kurosaki Isshin]

"up to developers" means "up to publishers"

*sigh*

[MEC-777]

One major question I have about this is how will the drivers work and interact? AMD GPUs with Nvidia GPUs will probably be the most problematic combination and may not come to fruition. Would require both parties to work together and I can't see them collaborating on something like that. AMD would be willing, but Nvidia, I doubt it. 

 

At least we know it does work for sure combining an APU with a dGPU. 

[patrickjp93]

This makes sense on why Nvidia are so hyping NVLink

 

They need to make a version of SLI so much better than DX12 shared resources in order to encourage people to just buy Nvidia cards.

That's not the reason. NVLink is for compute tasks where bandwidth is a problem. IBM has had NVLink 1.0 implemented in their Power 8 series motherboards for a long time already.