Why Asean is more than a collection of nations ...

This workload computes a fused multi head attention. Because it keeps the attention matrix in shared memory, it's both faster and uses less global memory. This is ...