The main focus of llama.cpp has been Apple silicon, so I suspect M1 would be more efficient. The author recently published some benchmarks: https://github.com/ggerganov/llama.cpp/discussions/4167