Free BLAS Isn’t Rubbish

2023-07-13

Note new data in BLIS repo.

There’s a myth that proprietary BLASes — at least in Intel MKL — are much better than free software ones.Only considering CPU ones here, not GPU.

See also reference in a previous item. It results in people insisting on using MKL even on AMD CPUs it doesn’t support, or isn’t meant to, in preference to AMD’s BLIS (a version of BLIS).Presumably at least no-one would deny that even reference BLAS is infinitely better than MKL on ARM and POWER, though.

(MKL was also widely used in contravention of its licence before that changed to be more permissive.) The myth is remarkably resistant to data, like the analogous one for the Intel compiler, but I’ll another opportunity to refer to some recently obtained.

Over the time I’ve known OpenBLAS, since its early days, MKL level 3 performance has only been much better than for OpenBLAS’ on new micro-architectures that no-one has taken an interest in.There was a hiatus and then backtracking on Skylake-X for a while, and it never supported Knights Landing.

BLIS is also fairly good in many cases, and possibly more robust than OpenBLAS for multi-threading. There’s also Eigen, though it’s not primarily to provide vanilla BLAS. In more specialized applications, for small matrices, the reason MKL improved was competition from libxsmm,According to its maintainer.

and BLASFEO’s limited coverage of somewhat larger ones seems fast.

Anyway, I’ll take the opportunity to refer to recent data supplementing the original,Which has since acquired new measurements, but specifically not for recent OpenBLAS.

particularly to be fair to OpenBLAS on SKX, and to add some more systems. There’s a sample result reproduced below for illustration.

By the way, ATLAS is sometimes referred to as a free optimized BLAS, but is certainly inferior to OpenBLAS on recent hardware, and appears to be dead.