Why did amd abandon terascale?

[NicholasNicholasNicholas]

I really wanna know and i cant seem to find the answer. Sure, gcn works better for dx12, but that wasnt even a blip on the radar back then. 

[Mira Yurizaki]

Here's a basic summary:

• NVIDIA was basically the king of the compute market at the time and AMD wanted in on that market
• The VLIW architecture that was used in Terascale works well in graphics, but not in compute
• The reason why VLIW is great for graphics is that VLIW requires a very good compiler in order to efficiently optimize a program to take advantage of how VLIW works (i.e., execute a bunch of instructions from a program at once), and graphics are simple and predictable enough for this to happen
• The reason why VLIW is not great for compute is that compute is more variable in execution time, inputs and outputs data-wise, and there's more chance for computations needing to wait on another part of the program to finish.
• VLIW also makes it harder to debug on the assembly level, and fine tuning for performance sometimes has to get down to that level.

tl;dr, VLIW works well for embarrassingly parallel workloads like graphics. It does not work well for compute which is not necessarily embarrassingly parallel.