Yeah the RDNA change probably is quite a bit like that after running with GCN for so long and dropping Tera Scale fairly early on trying to build on and improve GCN’s weaknesses and being both a workstation and gaming product.
Was thinking RDNA2 would improve and build on what RDNA1 had problems with and resolve many early issues and shortcomings but it has several new additions of it’s own and it’s possibly less of a addition like what GCN 1.1 or Gen2 was to 1.0 or Gen1 as a result.
With RDNA3 rumored to be AMD dabbling into the same structure that they have for the Zen CPU’s now core complex systems and all that.
Still not much about CDNA either the compute oriented hardware that’s more of a replacement for GCN whereas RDNA’s targeting gaming more specifically.
Caught up pretty well to NVIDIA but they’re one tessellation generation behind and it shows in the performance results far as I’ve been looking at it plus other features and how some of these will pan out.
Happened a bit with the later GCN updates too in addition to the instruction set itself and features the GPU’s were overhauled or refined in various ways trying to mitigate one issue after another back and forth so while the Polaris cards should be a bit weak compared to Fury they actually carry other improvements and then Vega continued with that plus some final tweaks in the Radeon VII variant or Vega20 though technically a Mi50 type card that didn’t quite meet the workstation requirements but could be re-used.
Supposedly AMD also lost a few of the old talent and had other lay-offs since Raja to Intel and NVIDIA although since then they’ve also improved the staff situation but that takes time and then well I suppose we shall see how things progress from here on.
A lot of these planned features also didn’t quite pan out or work at all and some of it needed direct development interaction between AMD and getting the RDNA10 / 5700 (And 5600) optimizations directly into the game engine to push against NVIDIA’s Pascal and Turing GPU series which was also out earlier than when AMD had something more competitive against it.
Keep this momentum going and further fixes and quicker solutions in addition to being quicker at competing with NVIDIA and hardware launches and that could be good but it’s too early to say and just about anything could happen.
One bad decision and they’d have another period of downtime and barely being competitive and some of the current solutions and balancing acts like infinity fabric over just a larger bus width could be detrimental at high resolution modes already.
Suppose HBM was another side-step too but if HBM2E or HBM3 could make a comeback it might be useful later on but there’s also the possibility it’ll be server systems and the FirePro type cards only if anything.
EDIT: Well that’s a bit lengthier than I was thinking it’d be.
Fun to see competition between AMD and NVIDIA picking up at least, just a lot of uncertainties and possible problems although NVIDIA could hit snags too but they’ve been doing well and they have a lot of resources and the desktop market is at 80% NVIDIA 20% AMD now or more too.