According to Agner Fog's instruction tables it has a 1 / cycle throughput, like every other shuffle type instruction. The latency is mainly determined by the memory access. A load from L1 cache takes 4 cycles, and the broadcast itself should take 1 cycle. There might also be a penalty of 1 or 2 cycles for crossing domains (I don't remember if that applies here). In any case this is a fast instruction for what it's designed to do.