an extra avx512 fma unit means it can do twice as many z=a*b+c instructions per clock. this only really matters for matrix multiplication (for anything else you are memory bandwidth limited), but 2x matmul performance is a really big deal for hpc