If CUDA = Compute Unified Device Architecture
Then what is a Unified Architecture?
I’m not finding a definition or description anywhere.
Older GPUs had hardware that implemented a particular rendering pipeline. Different functional units were used for geometry transformation, shaders, etc. If you look at older OpenGL releases, the API is organized around a particular data flow dictated by the hardware at the time.
With CUDA, the decision was made to (as much as possible) instead implement the different stages of the rendering pipeline with a generic set of programmable functional units that could perform all of the rendering stages. At the same time, this architecture also unifies rendering and general purpose computing abilities efficiently into the same device.
Would this be the difference between a dedicated hardware approach, versus generic hardware that can be programmed to do the same function?
i.e. There is no shader hardware, but you can program a shader operation with the general purpose processor hardware.
So Unified might imply a single hardware approach to many different operations?
Yes.