Does XDMA = better Hybrid Crossfire?

[NTMBK]

From Anandtech's 290X review:

Ultimately AMD’s goal with the XDMA engine is to make PCIe based Crossfire just as efficient, performant, and compatible as CFBI based Crossfire, and despite the initial concerns we had over the use of the PCIe bus, based on our test results AMD appears to have delivered on their promises.

http://www.anandtech.com/show/7457/the-radeon-r9-290x-review/4

Dual Graphics has always used PCIe to get the frames between the GPU and the APU. Does the addition of an XDMA in GCN1.1 mean we're going to see better Dual Graphics? (Because let's face it, it was a complete disaster before now.) Kaveri plus a Bonaire chip could be an interesting combination.

 

[BrightCandle]

Sure hope so. The promise of mixed GPUs has always been a great idea its just never worked.

 

[NTMBK]

Sure hope so. The promise of mixed GPUs has always been a great idea its just never worked.

I doubt it'd be enough to get you to trade in your 3930K for a Kaveri though

 

[Gloomy]

I doubt it'd be enough to get you to trade in your 3930K for a Kaveri though

Woah, there's a thought. AMD may not be in as bad a situation if GPU ***pute really takes off in games.

 

[Arkadrel]

My guess is Kaveri wont have XDMA in it, neither will the 7750's.
So it probably wont help them any with kaveri.

XDMA has a real chance of actually makeing hybrid crossfire decent though, in the future.

Whatever APU comes after the Kaveri, and whichever series of GPU cards they have out by then, will probably have it and have it working well.

By which time, running a low end pc, with a APU + amd gpu (in crossfire) will make alot more sense than going intel route.

 

[itsmydamnation]

mr dave from b3d has stated that Kaveri, Hawaii, kabini/temash and bonair all have the same IP level so i wouldn't rule it out.

 

[Tuna-Fish]

XDMA does not make the big problems in hybrid crossfire go away.

Basically, the problem is that the way modern game engines render frames, the work of a single frame cannot be easily split to two gpus. Modern games almost always use some sort of deferred rendering techniques. They effectively require that the entire frame must be present at one of the gpus at some points in the rendering process. Since moving partially rendered frames back and forth between the GPUs is going to choke any interface you can put between them, the only workable solution for crossfire or sli is alternate frame rendering, or AFR. That is, start rendering frame 1 on gpu1, once half done, start rendering frame 2 on gpu2, once frame 1 is finished, display it, then start the third on gpu1.

AFR sucks. Among other things, it makes latency worse and causes all sorts of timing issues, as it doesn't reduce the time taken to render a frame, only makes two of them render in parallel. It's especially bad for unbalanced setups, as having more power in gpu2 doesn't make gpu1 render it's frame any faster.

There is no future in hybrid crossfire, it's just a stepping stone on the way to where they can fit a good enough GPU on the same chip as the CPU. Given to how low the PS4 and XB1 will set the bar for multiplatform titles for the next few years, that's not far off.

 

[Leadbox]

On a side, yet related note, have you guys noticed how the frametime traces of XDMA Xfire looks like its dynamically tracking compared to the rigid, hard set limits of cable bridged Xfire.

This is with using the same frame pacing algorithm that came out in August
According to Pcper, AMD had actually requested that reviewers not test Xfire just yet (which explains why there's only 3 Xfire reviews) so it might get better still.
Is this how nVidia has been doing it?

 

[Arkadrel]

Is this how nVidia has been doing it?

No idea.

On a side, yet related note, have you guys noticed how the frametime traces of XDMA Xfire looks like its dynamically tracking compared to the rigid, hard set limits of cable bridged Xfire.

Noticed that too.

 

[BrightCandle]

That is a mighty interesting observation. I don't really understand why the 280X CF has that minimum value.

Pcper.com published some data right after the release of the frame pacing driver that showed a stepping at the bottom of the graph in CF. So while it tracked the changing frame rate it also had this little odd effect of discrete jumps, the data had minimums that seemed to change in groupings rarther than frame to frame. I tried to work out if it was based on time based correction or a threshold with some maths but so far I haven't yet cracked it and determined the answer. I would like to know which algorithm they chose that would produce those effects but I guess it doesn't by and large matter now that AMD with its future is banking on XDMA.

Its another piece of the puzzle but I am not seeing the answer yet.