The rule of thumb is that inference speed halves with every doubling of parameter size (and obviously a doubling of memory size).

You can check out real world performances on devices here: https://llm.mlc.ai/