Originally posted by bridgman
View Post
Originally posted by humbug
View Post
1. GPU hardware architectures have converged on "scalar SIMD" as a consequence of needing to support compute as well as graphics. For graphics-only the VLIW SIMD model was arguably more efficient since essentially all of the work involved short vectors (typically 3- or 4-element) plus a scalar or two, but the vector size for compute varied widely and was usually very large, so a scalar instruction set ended up being more versatile although it did require relatively more control logic (program counters etc..) for a given number of ALUs.
2. The biggest one IMO is the move away from older graphics APIs to newer ones like Vulkan and DX12 (probably Metal too, although I haven't looked at it much). OpenGL had become both sufficiently large and sufficiently old that there were just too many different ways to use the API, particularly with NVidia encouraging application developers to use compatibility profiles where the lack of standards more or less ensured a degree of vendor lock-in.
Comment