It's probably worth saying that ~10% differences in BLAS performance are probably lost in the noise of run time variance on an HPC system, especially as things don't typically spend most of their time in BLAS. Also, you don't have licence issues with BLIS and OpenBLAS. For instance when I last looked at cp2k, a major materials science copyleft application, it didn't have a licence exception for linking with MKL.