Dresdenboy
Golden Member
AVX in Dothan?Yep. This graph that ShintaiDK posts in every thread is also complete BS, it only applies to specific server applications utilizing AVX/AVX2 etc etc.
BTW if you count AVX instruction throughput, then doing 8 float ops with one AVX instruction has the same IPC as doing 4 with SSE. This btw is a problem with using Cinebench numbers.
Cinebench - while being cache friendly and thus somewhat independent of memory subsystem - is a bad way of trying to get a clue of IPC. That it's a "singlethreaded" benchmark doesn't qualify it for that task. It measures the efficiency of SIMD calculations, loop predictors and buffers, and some stuff along the path through the FPUs.