I seem to recall that most (all?) discrete GPUs have 16 wide SIMD engines and thus use 64 wide wavefronts (4 cycles per wavefront), but that some of the lower end / embeded GPUs actually have 8 wide SIMD engines (for example the 40 cores GPUs) and thus have 32 wide wavefronts. I can't however find any references to the point at the moment and was wondering if anyone can verify / disprove the notion and can hopefully give me some references either way.
Thanks