People sometimes run code written, or even compiled, by somebody else. Strange but true. ("At least I think it's strange, and I am assured it is true." -- DA)

Stop me if this is just too weird, but I believe if you rely on proprietary software then you should choose the platform where you get the best performance of that software.

There are a very finite number of HPC clusters available to any given researcher with a given grant. Each european country might have 1-10 available and you have to apply for compute time.

If you’re in a commercial setting, your company might have 1 cluster and N simulation programs.

The proprietary software mix might still perform better on AMD.

To add to that, I believe OpenBLAS surpassed MKL in benchmarks a while ago. Even if I had an Intel CPU, I would probably use OpenBLAS if I had the choice.

I benchmarked some of my large transformer networks the last few days and MKL is still 50% faster than OpenBLAS.

What's even worse in real-world applications is that OpenBLAS misbehaves when an application uses threads. This is also described in the OpenBLAS FAQ:

If your application is already multi-threaded, it will conflict with OpenBLAS multi-threading. Thus, you must set OpenBLAS to use single thread as following.

https://github.com/xianyi/OpenBLAS/wiki/faq#multi-threaded

So what are the results with libxsmm and current AMD BLAS, as that must be for small dimensions?

The reason it's serial BLAS that mainly matters is that HPC codes are usually parallelized above the BLAS; why do you want the nesting? Swapping in threaded OpenBLAS or BLIS is something you might do with basically serial stuff like vanilla R, e.g. https://loveshack.fedorapeople.org/blas-subversion.html#_add... OpenBLAS threading has been somewhat buggy, but the main problem with its OpenMP support currently seems to be that using OMP_PLACES kills it.

Reference please. This is definitely not my experience.

Definitely not mine either, though I haven't benchmarked in a few years (and would love this to be true).

MKL has PARDISO for sparse matrices which I guess has no equivalent in OpenBLAS. Though there seem to be faster open source sparse matrix solvers now.