Posted by Steven Vancoillie on October 03, 2012 at 11:28:28:
There is a bug present in the dcopy subroutine (and possibly other ?copy routines) in the integer*4 ACML 64-bit library as of version 4.3.0 (4.2.0 works fine), and in both the integer*4 and integer*8 MKL 64-bit libraries as of version 10.2 (I think, in any case 10.1.3.027 works fine). The bug occurs only on Opteron processors.
This bug can affect you if you (1) use opteron machines, (2) use one of the newer ACML or MKL BLAS libraries, and (3) run a big calculation AND use a large MOLCASMEM
I noticed the bug when using Cholesky decomposition, but there could be other places where this occurs.