vvid (no.delete@this.thanks.com) on August 3, 2016 12:22 am wrote:
> PM (phillip.delete@this.ravencity.de) on August 1, 2016 6:39 pm wrote:
> > David Kanter (dkanter.delete@this.realworldtech.com) on August 1, 2016 12:01 am wrote:
> > > contrast this behavior to the immediate-mode rasterizer used by AMD.
> >
> > I think you should try and run this test again on a current AMD (GCN) card.
> > The GCN cards (or at least the Hawaii XT one) behave somewhat
> > differently than the Terrascale 2 chip you tested.
>
> GCN rendering triangles in parallel and out of order.
> This is a couple of screenshots from R9-380 (4 rasterizers)
>
>
>
>

Thanks for taking the time to post this. I don't have a GCN 1.2 (AKA "Tonga") card on hand, but I really wanted to test this myself when Kanter put up the video. :(

Tonga has AMD's first color frame buffer memory compression tech, and presumably it works by rasterizing triangles in blocks in the frame buffer in order to compress tile-by-tile. That doesn't mean AMD is doing "tiling". For all we know, it could just be stepping over the triangle scan lines multiple times only rendering pixels in the box bounds. Very inefficient we would think, but FF hardware for triangle rasterization setup is so fast now, maybe it's worth it for the memory-bandwidth savings.

What's important here is that what David Kanter sees in the frame-buffer rendering on Maxwell doesn't really tell us anything new about Maxwell without additional knowledge.