There's plenty of optimisations still possible if you optimise for specific GPU families. Eg see Hipfire[1] and q38rocm/HaloPFX[2], they're so much more faster than llama.cpp for consumer AMD GPUs.
OK, yes, this I do agree with but I was mostly factoring these in (because I did that research a couple of weeks ago while speccing out a hypothetical R9700 box)