XB1/PS4 Polygon pushing power

Xbox One always had 2 geometry engines.

 

Bonaire has 2.

 

I believe it's a mistake to believe either of the APUs in the new consoles are "based" on any existing thing in particular, they're custom order jobbies designed for either company that ordered them.

 

Well I dont know if having a higher cu count would effect the poly count. Clockrate does though.
The xbox one because of the gpu clock increase is capable of 1.7 billion polys a second. The Ps4 is at 1.6 billion

 

GCN doesn't distribute the geometry workload across the multiprocessors (CUs) like what Nvidia does, so there shouldn't be disparity between PS4 and XBO setup rates -- just the clock-rate difference.

 

Pushing more polygons means more shading, more texturing, more shadowing, more collision detection etc..which I don't think the Bone could afford to even if it could theoretically push more polys. So in the real world I think PS4 should be able to render more polygons anyway since it's more balanced to do so.

 

Furthermore, adding more polys would require different art (which has a production cost impact, often a big one) - unless you're simply just tesselating, and AMD chips have traditionally not excelled at this so I don't know how common tesselation will be in this console gen.

 

He's just asking about polygons, not anything else. And there's no "even if it could theoretically" . It can.

 

No he didn't, he asked:

Yes.

Triangle setup is a pointless metric by itself because it depends on so many variables.

How complicated is the vertex shader?
How many verts shaded per triangle?
How many of those tris are culled?
How many fragments are there?
How complex are the fragments?

So if we are talking meaningful real-world polygon pushing potential then yes, the PS4's GPU has the higher ceiling.

Once these machines are rooted we could probably get real numbers from benchmarking.

Anyways, I think fill rate would be the more meaningful performance metric.

 

There are way too many variables to get at answer to that question. Designing a shader that would create max throughput at any part of the pipeline wouldn't be of any practical use. The difference is negligible in either direction.

 

Agreed, but you can think it in an inverse way

Every shader is bound by some bottleneck, and thus maximizes the throughput of that part of the GPU. Shadow mapping for example is either bound by the primitive rate or the fill rate (ROPs) (not by BW since depth compression solves that). The large objects (low triangle density) are ROP bound and the small objects (high triangle density) are primitive bound (in general).

Every developer tries to balance the GPU usage. Deliberately hitting a bottleneck (maximizing throughput of only one part of the hardware while other parts are idling) is of course a stupid design decision

 

I don't know why people don't think geometry work is distributed across the CUs. Geometry related shading can execute on any CU.

Geometry engine is a marketing name encompassing all fixed function hardware related to geometry processing. This includes vertex reuse checking, tessellation, index fetching, clipping, setup, etc.

Xenos peaked at 1 triangle or 1 vertex per clock and vertex reuse determined which limited performance. At some point there could be too many ALUs in the VS or too many memory fetches and peak rate would not be maintained.

More CUs potentially means you can execute longer vertex shaders, but you would need to run benchmarks to know if it really makes a difference.