From lecture: DNN evaluation primarily involves doing large matrix-matrix and matrix-vector multiplications, so we can design a chip that just does these things really well without the overhead of managing a general instruction stream. More specialized hardware can also implement new instructions such as mul-add which is used very often when evaluating DNNs.
From lecture: DNN evaluation primarily involves doing large matrix-matrix and matrix-vector multiplications, so we can design a chip that just does these things really well without the overhead of managing a general instruction stream. More specialized hardware can also implement new instructions such as mul-add which is used very often when evaluating DNNs.