It should improve performance on most hardware because most LLMs are memory bandwidth bound during decode.
It should improve performance on most hardware because most LLMs are memory bandwidth bound during decode.