GCN performance

[PontiacGTX]

I have checked the Ghost recon wildlands benchmarks and I noticed something when they used Ultra Quality (UQ) the 470 is even ahead the R9 290X by a big margin when it was on very high Quality(VHQ) the 290x was faster than 480, what could have caused that a 470 that currently is slower than a 290 by a bit, outperform a 290x? and even the R9 Nano is slower here than RX 480 when the R9 nano is always as fast as 1060/Fury OR Faster

 

[Glo.]

Primitive Discard Accelerator.

Hardware features that are in Polaris architecture, and are not in previous generations of GCN.

 

[PontiacGTX]

Primitive Discard Accelerator.

Hardware features that are in Polaris architecture, and are not in previous generations of GCN.

or it can be the Vram

 

[ThatBuzzkiller]

Good chances that there's a geometry bottleneck going on with ultra quality if we take a look at how the GTX 780 is outperforming most of the GCN parts ...

Async compute with Frostbite 3's new triangle filtering method would really help ...

 

[richierich1212]

Good chances that there's a geometry bottleneck going on with ultra quality if we take a look at how the GTX 780 is outperforming most of the GCN parts ...

Async compute with Frostbite 3's new triangle filtering method would really help ...

Wildlands is a Gameworks title.

 

[ThatBuzzkiller]

Wildlands is a Gameworks title.

It is much deeper than that ...

 

[crisium]

Geometry / Tessellation most likely. Improved from GCN 2 to 4, and at only 1080p the 32 ROPs and reduced memory bandwidth from the 470 are less of a bottleneck.

 

[PontiacGTX]

Geometry / Tessellation most likely. Improved from GCN 2 to 4, and at only 1080p the 32 ROPs and reduced memory bandwidth from the 470 are less of a bottleneck.

tessellation improved mostly due to the primitive discard accelerator but tessellation in polygon throughput didnt improved too much

 

[crisium]

Don't forget GCN 3 (Tonga) had tessellation improvements over GCN 2. I am assuming these carried over to GCN 4.

You can see this with 380X vs 280X (GCN 1, but there were no GCN 2 tessellation improvements IIRC) showing similar results to the 470 vs 290 comparison the OP made.

 

[Red Hawk]

Don't forget GCN 3 (Tonga) had tessellation improvements over GCN 2. I am assuming these carried over to GCN 4.

You can see this with 380X vs 280X (GCN 1, but there were no GCN 2 tessellation improvements IIRC) showing similar results to the 470 vs 290 comparison the OP made.

The 380X had an extra advantage in tessellation over the 280X because it doubled up on hardware tessellator units (4 in the 380X vs 2 in the 280X). Since Hawaii, 4 hardware tessellators has been pretty much standard (I hope for more in Vega...)

 

[Valantar]

Am I the only one surprised by the odd CF scaling in those graphs? I'm used to seeing no gains (i.e. CF not actually being used) or even performance drops (driver overhead and buggy implementations, I suppose), but single-FPS gains? As in: it's consistently faster, but about 2% faster. So CF is doing something, it's just doing so little as for it to be barely outside of the margin of error. Am I the only one thinking that's really weird?

 

[Shmee]

It is rather strange. Of course CF scaling depends on games and drivers, so perhaps it just is not a very good game at all (yet) for mutliGPU. See that the 1080 SLI has negative scaling. My guess is just the game a bit new and un-optimized in that regard.

 

[Crumpet]

I think the gaming market is simply moving away from SLI and Crossfire, single gpu's are so fast now there's usually little point in doubling up on a GPU as you can get the same performance from a single gpu for the same money or less.