Some notes:
Vega geometry rate is up 2.6x over Fiji. Fiji 4 geometry pipeline was designed for 4 polygons per clock - Vega is 11 polygons per clock. That is a huge increase. If the clock speed is indeed 1526mhz or around for the rated TFlops - geometry restrictions will be a thing of the past. AMD did work the front end and it looks like a lot.
The backend or pixel rendering end is now connected to the L2 cache vice the memory controller, this should really speed up deferred rendering and free up the memory accesses for other operations.
CU is still 64 shaders each. AMD indicated it is made for speed - we will see.
NO anything more than 4 are only from primitive shader usage, Don't fall for the marketing hype man. We have discussed this earlier in this thread or the other vega thread. Raja's presentation of Vega, talk about RPM and polygon performance with Vega. Their 11 polygons per clock is with RPM which has to be done through primitive shaders. And that only equalizes what the competing Pascal can do. So pretty much old games, current games, and most of next gen games using current engines, not going to have primitive shaders. Cause AMD hasn't released their SDK yet. Probably won't be released till a month or two after Vega.
Yes mentioned that. but speed up err I have a different way of look at it. It saves cache thrashing so unless that is happening on a regular bases, the speed up won't be that noticeable. AMD talked about this, they can make sure it doesn't happen through drivers. Its more of a work reduction on their drivers team and some performance gains, limited