GeForce + Radeon Previewing DirectX 12 Multi-Adapter with Ashes of the Singularity

[D.P.]

Another major point with scaling in this MultiGPU setup, is how they are performing their SFR.

Depending on the implementation it can make a vast difference in the overall performance.

I think they are using a Split screen method, wethere it is a fixed or variable split is unknown. I would assume they would have some kind of variable split with the engine determining the workload and trying to equalise it as much as possible. Split screen should work better in an RTS game considering the majority of the top half of the screen is not looking at sky like with an FPS or RPG etc.

SFR Supertiling should still be the best overall. Better scaling and same latency as single card.

SFR is flawed fundamentally for most games so Ashes like almost every game uses AFR which is a newer and better approach than the ancient SFR originally used in SLI from the 3DFX days.

 

[chroniclard]

Steam forums info.

"NEXT UP: Add another video card to get more performance!
Simulating a world-wide RTS war where you can zoom down on individual units with their own lighting and details and then seamlessly zooming out to see thousands fighting a war is very hardware intensive.

But we are pleased to announce that we have developed a solution for DirectX 12: You will be able to add additional video cards to increase performance. We don't mean SLI or Crossfire. We mean, just add another DirectX 12 supporting video card of any make or model.

Have an NVIDIA card? Just throw in another one or heck, an AMD card. How many times have you seen some new video card come out and decided not to buy it because your existing card was just fine and who wants to toss out a perfectly good video card?

We've given Anandtech our first build of this working for independent verification: http://www.anandtech.com/show/9740/directx-12-geforce-plus-radeon-mgpu-preview

We expect to have this to you guys in time for Beta 1 next month!"

 

[Orcvader]

From what I see this is just to make use of an old GPU you may have lying around or make use of the iGP. Would be silly buying the flagship of both AMD and Nvidia in competing generations just for this .

 

[Mauller]

SFR is flawed fundamentally for most games so Ashes like almost every game uses AFR which is a newer and better approach than the ancient SFR originally used in SLI from the 3DFX days.

SFR is not fundamentally flawed. The old implementations of running over DX are what made AFR a necessity for multi gpu. The problems revolved around gpu synchronisation which can be performed far better and more easily with a low abstraction API.

Also AFR just bloats FPS numbers due to frame queues and sucks for input latency.

Plus the implementations of SFR are nothing like 3DFX implementation, SLI meant Scan line interleaving. And their Scanline Frame Rendering split the scene into odd and even scan lines and divided the work between gpu cores. Not whole sections of a screen like modern Split Frame Rendering.

 

[Dygaza]

It'll be inteteresting what kind of advanced rendering methods we're going to se in addition to traditional AFR and SFR. Can we really see each card specialising in different areas. Like post prosessing, global illumination and all other fancy words

Interesting times ahead.

 

[D.P.]

SFR is not fundamentally flawed. The old implementations of running over DX are what made AFR a necessity for multi gpu. The problems revolved around gpu synchronisation which can be performed far better and more easily with a low abstraction API.

Also AFR just bloats FPS numbers due to frame queues and sucks for input latency.

Plus the implementations of SFR are nothing like 3DFX implementation, SLI meant Scan line interleaving. And their Scanline Frame Rendering split the scene into odd and even scan lines and divided the work between gpu cores. Not whole sections of a screen like modern Split Frame Rendering.

SFR is flawed because the work load greatly changes over the view-port in non-linear way for almost all games -that is never ever going to change unless you only play a game that enforces an fairly even workload like a top down RTS game. The problem was nothing to do with synchronization, that problem exist just as much in AFR.

SFR is not just scanline interleave, what both Nvidia and ATI did for years was to split the screen into 2-4 parts and send each part to a different GPU, and attempted to shift the dividing lines based on the workload of the previous rendered frame. This was when SLR profiles became paramount because some games behaved well when splitting the screen horizontally like an indoor FPS shooter, where ceiling and floor had similar detail and there would be characters equally likely to be left, right or center. But it was terrible for outdoor games because the top half of the screen was blue sky so the dividing line would be right near the bottom, the moment the player looked down this would be all wrong and although the dividing line would automatically shift up to divide the workload equally there was always lag.

That is why both ATI and nvidia shifted to AFR as a superior solution that prevents the kind of jerkiness and uneven workload inherent is splitting the screen. The workload is always split optimally with AFR because each GPU is rendering a complete frame and there is no guessing how much different potions will take to render.


Of course AFR increases FPS, that is the entire purpose!

 

[D.P.]

It'll be inteteresting what kind of advanced rendering methods we're going to se in addition to traditional AFR and SFR. Can we really see each card specialising in different areas. Like post prosessing, global illumination and all other fancy words

Interesting times ahead.

Not really, far too much work for most developers. Do you really expect them to work out how good the hundreds of different GPUs are for the hundreds of different rendering tasks and do some complex optimization process to match tasks to different GPUS? Let alone imagining debugging that mess.
The big developers will likely touch the surface of such capabilities though. I could see scenarios where the final frame is perhaps sent to the iGPU for post processing for example,or the iGPU does some compute tasks in the background. If you have 2 mainstream cards, say 980s, it is going to be more efficient havign both of them split the work load through AFR rather than having one do GI and the other the main rendering.

 

[D.P.]

For those that fell for the AMD propaganda about SFR,, here is a good review form 11 years ago,

http://www.hardocp.com/article/2004/11/23/nvidia_sli_upgrade_guide/6#.Vi990aK73LQ

. It was like the software was continually trying to keep up with the fast paced action and load balance correctly, but was lagging behind figuring out where the best place to load balance was.

And that was an inherently impossible problem to solve because until you render the frame you don't know where the split should be to dive the work load equally. Which is why both ATI and Nvidia ditched SFR.

 

[Dygaza]

No I don't expect small studios to invest that much in multigpu support in their housemade engines atleast. But considering that there are a lot of people updating computers all the time (especially gpu's). I do see adapting SFR support a lot more appealing for developers as traditional SLI/crossfire has ever been. Afterall everyone will have those old cards lying around , and ready to be used.

I do hope major engine developers will truly dedicate themselves into getting these new features out, and show how it's done. Their engines are afterall used in most games (unreal, cryengine, unity etc).

 

[Dygaza]

For those that fell for the AMD propaganda about SFR,, here is a good review form 11 years ago,

http://www.hardocp.com/article/2004/11/23/nvidia_sli_upgrade_guide/6#.Vi990aK73LQ

And that was an inherently impossible problem to solve because until you render the frame you don't know where the split should be to dive the work load equally. Which is why both ATI and Nvidia ditched SFR.

Valid point, but quite outdated review. Afterall we're talking about DX12 which gives devs the control , what to split, and when. Before it was done by driver.

Also keep in mind that you don't need to split everything equally (you can't). So you can't get as good scaling as you get with AFR. But you do get the smoothness, and benefit of adding cards that ain't equally as fast.

It's not easy to do, but deffinately worth the effort.

 

[David Bisset]

Surprisingly good performance, honestly I didn't believe it would work nearly that well (if at all!). Glad to see I was wrong

Edit: Talking about the OP, not caught up with the SFR/AFR argument above yet!

 

[GoogalyMoogaly]

What happens if the image quality of one vendor is different to that of the other? One may have more washed out colours or something.

Will there be a divide on screen between the bit rendered by one and the bits rendered by the other? Or if it's alternate frame will it just flicker?

On the plus side we might be able to use this to get AMD cards to output 4K 60Hz over HDMI 2.0 by adding an Nvidia card!

I do get a little bit worried that scaling may change based on the game, so now instead of having a game that runs better on Nvidia and a game that runs better on AMD we'll now have...
Game A runs better on Nvidia + Nvidia
Game B runs better on AMD + AMD
Game C runs better on Nvidia (Primary) + AMD
Game D runs better on AMD (Primary) + Nvidia

 

[Noctifer]

What happens if the image quality of one vendor is different to that of the other? One may have more washed out colours or something.

Will there be a divide on screen between the bit rendered by one and the bits rendered by the other? Or if it's alternate frame will it just flicker?

I don't think that would be the case since that depends on the output GPU, the one that delivers the frame to the screen. I think the review would mention it if something like that would happen. Wouldn't be surprised if Freesync/G-sync could work as well.

 

[Mauller]

SFR is flawed because the work load greatly changes over the view-port in non-linear way for almost all games -that is never ever going to change unless you only play a game that enforces an fairly even workload like a top down RTS game. The problem was nothing to do with synchronization, that problem exist just as much in AFR.

SFR is not just scanline interleave, what both Nvidia and ATI did for years was to split the screen into 2-4 parts and send each part to a different GPU, and attempted to shift the dividing lines based on the workload of the previous rendered frame. This was when SLR profiles became paramount because some games behaved well when splitting the screen horizontally like an indoor FPS shooter, where ceiling and floor had similar detail and there would be characters equally likely to be left, right or center. But it was terrible for outdoor games because the top half of the screen was blue sky so the dividing line would be right near the bottom, the moment the player looked down this would be all wrong and although the dividing line would automatically shift up to divide the workload equally there was always lag.

That is why both ATI and nvidia shifted to AFR as a superior solution that prevents the kind of jerkiness and uneven workload inherent is splitting the screen. The workload is always split optimally with AFR because each GPU is rendering a complete frame and there is no guessing how much different potions will take to render.


Of course AFR increases FPS, that is the entire purpose!

When I said AFR bloats FPS and increases latency. I mean that you are nnot getting real time fps. Many of the displayed frames are old which is where the latency and FPS bloating comes from.

And when I said SLI and SFR in second paragraph it was in relation to 3DFX implementation.

And they went with AFR in the end because it was easier to synchronise frames.

Plus ATI"s super-tiling based SFR ran great when it worked and the scaling was near that of AFR. And matched AFR when there was only a small frame queue used for AFR.

 

[SkeeterUK]

How come there isnt a gpu and onboard link up to test yet. Im interested to see how a 670 with my i5 3570ks hd4000 would fair.

 

[Dave2150]

The problem is even if you run high end graphics cards with a 5960X (which is a beast on compute heavy applications) you still don't get the performance you should do.

The 5960X isn't a great performer in Ashes, Skylake beats it:

 

[Dygaza]

How come there isnt a gpu and onboard link up to test yet. Im interested to see how a 670 with my i5 3570ks hd4000 would fair.

That would never work properly with AFR. iGPU is simply too slow. Would need SFR (or like) method to get real benefits out if. Just like they did demoed before with unreal engine where intel igpu was doing post prossessing.

Anyways, I doubt game like ashes will benefit from these, as it's very taxing for system ram bandwith aswell, and if you have igpu eating from that cake, you would be bottlenecked there.

 

[Mauller]

That would never work properly with AFR. iGPU is simply too slow. Would need SFR (or like) method to get real benefits out if. Just like they did demoed before with unreal engine where intel igpu was doing post prossessing.

Anyways, I doubt game like ashes will benefit from these, as it's very taxing for system ram bandwith aswell, and if you have igpu eating from that cake, you would be bottlenecked there.

For the Unreal demo, they were performing all the post processing on the iGPU. there was no SFR or AFR involved. Just pure work splitting.

 

[drunkenmaster]

The 5960X isn't a great performer in Ashes, Skylake beats it:

Errm, DX11 is very poorly threaded and at best the 6700k has 20% higher clock speeds(3.5 vs 4.2), in reality from a very quick look at reviews I would expect the 5960x to not get as high out of it's turbo and the gap between stock/turbo is also much greater on the 5960x, even greater when you consider the clockspeeds 3Ghz to 3.5Ghz is just over 16%, 4 to 4.2Ghz is 5%.

So with a 20% minimum but maybe closer to 30% clock speed difference in DX11 the 6700K is all of 15% faster than the 5960x, under DX12 it's about 4%. With far great overclocks available(due to lower base clock) which is the 'faster' chip. If both clocked to 4.5Ghz, the 5960x would be spanking the 6700k in DX12 with ease.

 

[chroniclard]

Guess when it hits beta stage I can try it out with a 980 and a 280X(afaik this is DX12?)