> running on silicon optimized to make reference counting as efficient as possible

I'm curious to understand this. Is this because of a specific instruction set support, or just the overall unified memory architecture?