Actually because ROCm ships native machine code, as opposed to a bytecode, so supporting any additional architecture is a tonne of effort.

I expect support will eventually be extended to a generation or two prior GPUs, but nothing beyond that.