I reported this issue with MKL 6.1. Hopefully it is on the enhancement list for MKL 7.X.
I have a general purpose MATRIX library ( Rogue Wave Math.h++) that was easy to retrofit to use MKL. My application uses a wide variety of complex matrix sizes from 6x6 to 500x500.
For small matrices, I took a big performance hit in using MKL due to some obvious overhead in MKL calls.
I ended up having to derive special small matrix classes to call an inline zgemm.
Using Rational Quantify shows the bottleneck to clearly be zgemm in MKL.