I think if AMD were to do 4 stacks again, they'd need a bigger chip and interposer and would face the same or even worse problems than Fiji with HBM2 being a bit bigger and that's why the last year's leak of 7nm Vega 20 had four stacks. It looks like AMD have scrapped those plans and are going for 14nm+ which would have the same problem.
There does seem to be something of a discrepancy between Greenland as it was described in some past leaks and the Greenland that supposedly is in Vega now.
Perhaps there was a Greenland A and Greenland B, or something got changed mid-stream.
At least for now, we do not have a Vega GPU with 1/2 rate DP and 4 GMI links as a product (caveat: die shots are too blurry).
One possible advantage to the 2-stack arrangement, aside from area savings to the GPU and interposer, is that the GPU section may have been able to grow/shrink more freely than if I/O, HBM, and all the other logic surrounded it.
GCN GPUs tend to have an internal patchwork of revisions even within a generation, and potentially some elements like the CUs and uncore were switching later in the process.
On that note, it does appear from the latest diagrams that the data-side Infinity Fabric interfaces on the outside of the L2, and does not run within the L2 and CU boundary.
While GF can be seen as year behind TSMC, they're not spending their R&D to 10nm process at all unlike TSMC, and their 7nm (at least the 1st 7nm anyway) is indeed performance oriented
That might have implications as well, and some may apply to Vega.
GF's high-performance process specifically points out performance ranges akin to the POWER or high-end CPU lines, which AMD's slides concerning its older APUs and Excavator showed to have very different structure and needs from ASICs.
A lot of AMD's motivation for its interposer-based integration method is to get the GPU and CPUs back out of compromising their process features and metal stack to cater to the other.
In Vega's case, it has features like its register file that took input from design expertise derived from Zen, but is that a win if that means parts of Vega are closer to CPU-optimized than the GPU they are in?
Then again, we do have historical evidence of AMD cards gaining more performance over time compared to GeForces from every generation since GCN was first introduced. Vega being biggest change in the architecture, but still essentially GCN, is there a reason not to expect at least similar performance uplift as previous generations, if not bigger due the bigger than usual changes?
That will be interesting to watch. One caveat is that it may be less consistent and may favor the CLC and 3rd party boards with custom coolers. Potentially, it may also favor the lower-performance settings since Vega appears to be pushed even further than prior designs.
Performance optimizations that help utilization without a corresponding power decrease would just push Vega into a throttling zone, so some changes may not behave as expected.
Can sombody Test David Kanters trianglebin?
I hope someone remembers it, although it may not be possible to capture the full range of that function with this test, due to the UAV.
Maybe or maybe not? I do not recall what the test does to create the culling cases.
It may help to determine if the primitive shaders are up to snuff. If the synthetic's culling settings work by making a varying percentage of triangles back-faced, primitive shaders theoretically would not let them get to the rasterizer.
In that case, the primitive shaders and rasterizer render each other partly redundant. However, in that case the question is why the culling isn't better. Vega's polygon tests show it to be competitive or better in non-culled and half-culled cases. It's the fully culled cases that do not have the outsized benefit that Nvidia's GPUs do.
