Bug in 7.2.008.
DGEMM produces NaN, if the third matrix argument C (C=a*A*B+b*C) had some NaN
elements, even though the constant "b" is identically zero.
MKL 8.0 (8.0.1) seems to have no bug.
The problem was not reported to Intel: I am very disappointed in their
"Premier Service" that had never corrected a single problem I reported (of which
there were several).
Click here to download a sample Fortran file reproducing the bug.