Dual GPU cards may finally be able to combine their VRAM.

[PerfectTemplar]

Mantle only + needs to be optimized by game devs.

 

So there will probably be 1 or 2 games next to battlefield 5 which supports the technology and that's it. Not so excited....

If DirectX 12 could do this, i would be soooo happy.

They're talking about the influences of mantle on the industry. Basically saying that the fact they can make mantle achieve this means they want the entire industry to be influenced and also be able to achieve it across multiple APIs. Also, it would be a feature of the API that should be inherent in applications that utilise the API.

[Mister Snow]

My guess is this has something to do with how Mantle crossfire works.

 

In Beyond Earth mantle crossfire uses a different rendering technique than AFR (I believe it's similar to SuperTiling described here http://techreport.com/review/8826/ati-crossfire-dual-graphics-solution/3) and I suspect that if you do this properly it may be possible to not require both cards to have the same framebuffer contents.  The reason it typically isn't used is that it's more complicated to implement, but maybe Mantle/DX12 remove enough overhead on drawcalls to make it feasible instead of AFR.  

 

It might be that they are using HSA method of communication by which they pass along pointers to memory segments instead of copying data so GPU 1 could access data on a memory chip of GPU 2 and the other way around. 

[Gunjob]

I find it amusing how many people think this applies to SLI/Crossfire configurations. This is for dual GPU based cards such as the R9 295x2 and HD 7990.

 

Instead of consuming both memory spaces for the same resources the developer can optimize it as a single memory pool.

I was being hopeful more than anything, but if it only applies to a dual GPU card, then they just got alot more appealing if this gets adopted in DX12 games.  

[AxelRantila]

Why is it amd that always pushes for more innovation while invidia looks for ways to just be effecient like really come on they did get sound through hdmi and now they want to put freesync on displayport come on invidia do some new stuff besides gameworks

Nvidia gave us G-sync. If it weren't for that, how much interest would  AMD have in free-sync.

Oh, and Nvidia cards ships with HDMI 2.0

[ChrisxIxCross]

Sounds awesome! This is the kind of technology we need with people pushing higher resolutions/ game engines becoming more and more advanced. I can imagine Star Citizen at 4K Ultra settings will need at least 8GB of vram.

[Agent181]

Nvidia gave us G-sync. If it weren't for that, how much interest would  AMD have in free-sync.

Oh, and Nvidia cards ships with HDMI 2.0

Oh you mean that thing that really doesnt work and brings up the price of a monitor up to 2x its normal price also amd cards are old we are barely getting hdmi 2.0 on them in the next release

[TheRetiredSlave]

Errrr, dumb question:How do they get around the inherent fact that the VRAM pool of one GPU isn't connected to the other GPU? I mean, sure the second GPU can access the first one's memory, but how can it do it without the inherent latency being a problem?

[Keudn]

so my 2GB 770 can sli without being pointless?

[Brainless906]

I hope AMD makes this proprietary, maybe locking it to Mantle.

 

Have a kick at Nvidia's proprietary butt.

 

No thanks, open standers plz. even if its one of those things that ends up performing better on amd cards (like mantle would presumably run better on an AMD gpu vs a NVIDIA one, when/if we get to that point).

 

No locked down bullshit. 

[Brainless906]

Aww it'd be great for once to give Nvidia a taste of its own medicine.

 

Actually if they make Mantle completely open and then bake the tech into it...that would force nvidia to take a closer look at building something on mantle. Or they'll just build there own system in response. Honestly the tech will come to both sides of the fence eventually, i'd just prefer not to segment the market more than it already is over arbitrary in game features. I just think if they baked it into something of their own, even if 100% open it'll make nvidia down create thar version just because pure ridiculous pride wouldnt let them use anything built by AMD.

[The Benjamins]

Aww it'd be great for once to give Nvidia a taste of its own medicine.

it would help Nvidia more than hurt them. AMD is not in a good financial spot and needs more adoption, so they need new tech to be open source so they can get free RnD. also its a API thing, so it can be in mantle and Microsoft can put it in DX12, also mantle is open to other companies to use too.

[Tim Drake]

Cool, now SLI 970's can have 7GB!

Where do you live so I can punch you in the face?

 

I'm joking but still, not funny

[Brainless906]

it would help Nvidia more than hurt them. AMD is not in a good financial spot and needs more adoption, so they need new tech to be open source so they can get free RnD. also its a API thing, so it can be in mantle and Microsoft can put it in DX12, also mantle is open to other companies to use too.

 

Even if its open though, or like mentioned before baked into mantle which is open, i highly doubt Nvidia would touch it at all purely out of a pride thing.

[The Benjamins]

Even if its open though, or like mentioned before baked into mantle which is open, i highly doubt Nvidia would touch it at all purely out of a pride thing.

if its added to DX12, Nvidia has no say if DEVs use it or not.

[Thony]

I think this should be pointed out from the article:

"just because you are running Windows 10 and DX12 and/or Mantle API does not mean that your multiple GPU configuration is now stacking memory. The capability is present in these low overhead APIs but they will not come into affect until devs specifically optimize the game as such."

So there will be 2 games a year that will stack the vram. Yaaaay, many people excited after hearing the news but when it depemds on publishers its just a dissapointment because they don't care.

[EChondo]

Who ever said that the 970 doesn't have 4GB of usuable vram?

I was joking man, just trying to have a little fun from the misconceptions happening recently haha.

 

Where do you live so I can punch you in the face?

 

I'm joking but still, not funny

Southern California, come fight me! I'll defend my right to joke around!

 

I'm joking but still, learn to laugh

[othertomperson]

Didn't they already try this before, back when AFR and SFR were alternative dual GPU technologies and ultimately SFR performed worse than a single GPU, hence it being abandoned?

 

If this works and is efficient, it can't be a bad thing. And if it ends up giving value to dual mid-range GPUs like 760s and 270Xs, high end gaming could become an awful lot cheaper so that'd be welcome too.

 

 

I was joking man, just trying to have a little fun from the misconceptions happening recently haha.

 

Southern California, come fight me! I'll defend my right to joke around!

 

I'm joking but still, learn to laugh 

 

A lot of people on this forum genuinely believe what you said is true. I think that's the issue that people are having.

[Wyzzy Moon]

So you could have more or the same amount of VRAM as system RAM and it wouldn't even be that ridiculously unbalanced

[Dabombinable]

Now if they also allowed people to offload task other than PhysX to a second GPU, I'll be happy. Without using SLI, I see enormous performance boosts in PhysX games just by having my GTX 650Ti (2GB OC version) set as the CUDA card.

[EpicAdom]

Now, how good is that extra ram.

What?

[Tim Drake]

I was joking man, just trying to have a little fun from the misconceptions happening recently haha.

 

Southern California, come fight me! I'll defend my right to joke around!

 

I'm joking but still, learn to laugh

Jokes been overused to hell.

[Fetzie]

i always wondered why they havent done this before. i mean it just turning it into a shared memory pool rather then 2 duplicated sets. the obliviously a lot of harder things to overcome. but strange its never been done before.

I think it's at least partially to do with the bandwidth available via the SLI/CFX bridge. Now that AMD pushes the data via the PCI-E lanes that is no longer an issue.

Mantle only + needs to be optimized by game devs.

So there will probably be 1 or 2 games next to battlefield 5 which supports the technology and that's it. Not so excited....

If DirectX 12 could do this, i would be soooo happy.

It is something that is more likely to be leveraged by workstation applications such as Photoshop or CAD.

[EChondo]

A lot of people on this forum genuinely believe what you said is true. I think that's the issue that people are having.

Well only 2 people have called me out on it, but nearly 35 people have "liked" it, so I'm assuming the 35 of them got the joke.

 

Jokes been overused to hell.

I haven't seen anyone joke over it?  :huh: If there have been then I am sorry, don't mean to overuse a joke, at least I didn't say Keep on Diggin'  :lol:

[Thracks]

Non-mirrored VRAM isn't the only cool thing you can do with mGPU in Mantle. You can also create an asymmetric CrossFire configuration between GPUs of substantially different performance envelopes, then load-balance the workloads between them so each card runs at peak performance. For example, you could have the dGPU do geometry and lighting, where an APU can do physics and AI work.

 

You can also asynchronously schedule compute and graphics workloads at very fine granularity, which is tremendously helpful to the latency of a VR solution. 

 

You could do split-frame rendering, which treats mGPU as a single faster GPU. No AFR latency penalty, no queue depth. In every way it will behave like one GPU with the resources of two.

 

Mantle can do its own command buffer prediction, where the command buffer is the stream of work sent to the CPU by the game, which the CPU reorders for the GPU. Typically the CPU would have to calculate scenarios for the command buffer, working ahead of the game to predict what will happen for peak performance. All of this ahead-of-time calculation has to be shared with the GPU, with the GPU replying back with what it's currently doing. You can eliminate all of that round trip performance junk if the GPU is able to make its own estimations based on the knowledge is has of its current rendering state. This applies to any sort of graphics config if the developer chooses to use it.

 

You can do per-eye rendering with mGPU, with each GPU giving full performance of a GPU to one eye in a VR or stereo3D configuration.

 

There's so much more lurking in that API that is so far beyond what we think of as "rules" in the PC industry about how CPUs and GPUs should work together, or MGPU must be done.

 

I'm pleased to see people are interested in this kind of stuff in Mantle.

[Dabombinable]

Non-mirrored VRAM isn't the only cool thing you can do with mGPU in Mantle. You can also create an asymmetric CrossFire configuration between GPUs of substantially different performance envelopes, then load-balance the workloads between them so each card runs at peak performance. For example, you could have the dGPU do geometry and lighting, where an APU can do physics and AI work.

 

You can also asynchronously schedule compute and graphics workloads at very fine granularity, which is tremendously helpful to the latency of a VR solution. 

 

You could do split-frame rendering, which treats mGPU as a single faster GPU. No AFR latency penalty, no queue depth. In every way it will behave like one GPU with the resources of two.

 

Mantle can do its own command buffer prediction, where the command buffer is the stream of work sent to the CPU by the game, which the CPU reorders for the GPU. Typically the CPU would have to calculate scenarios for the command buffer, working ahead of the game to predict what will happen for peak performance. All of this ahead-of-time calculation has to be shared with the GPU, with the GPU replying back with what it's currently doing. You can eliminate all of that round trip performance junk if the GPU is able to make its own estimations based on the knowledge is has of its current rendering state. This applies to any sort of graphics config if the developer chooses to use it.

 

You can do per-eye rendering with mGPU, with each GPU giving full performance of a GPU to one eye in a VR or stereo3D configuration.

 

There's so much more lurking in that API that is so far beyond what we think of as "rules" in the PC industry about how CPUs and GPUs should work together, or MGPU must be done.

 

I'm pleased to see people are interested in this kind of stuff in Mantle.

Would all of these features work on the R9 290X? Because I might be replacing my faulty GTX 970 with one after the RMA is sorted.