Erroneous eigenvalue computation for large matrices in R2016a when forcing single threaded execution

Why does MATLAB R2016a produce wrong eigenvalues for large matrices (N>7799) when forced to run as a single threaded application? 

Example code: 

%Diagonalization of a tridiagonal matrix.
t=1; %off-diagonal elements
tic;
%creates tridiagonal matrix of size NxN
T=gallery('tridiag',N,t,0,t);
T=full(T);
T(1,N)=t;
T(N,1)=t;

[V,D]=eig(T,'vector');%compute eigenvalues and store in vector

fprintf('Invoking "eig()" yields eigenvalues:\n')

fprintf('max: %2.4f \n',max(D))

fprintf('min: %2.4f \n',min(D))

toc;
exit

Execution Command: 

taskset -c 0 /opt/MATLAB/R2016a/bin/matlab -nodisplay -nodesktop -nosplash -nojvm -r "N=7800;test"

The above command yields wrong eigenvalues. The correct max and min eigenvalues are supposed to be 2 and -2 respectively.

However, the following multithreaded implementation produces accurate results: 

 

taskset -c 0,1 /opt/MATLAB/R2016a/bin/matlab -nodisplay -nodesktop -nosplash -nojvm -r "N=7800;test"

 

Answer 1

[MathWorks Support Team]

CAUSE:

This error is due to a bug in the Intel Math Kernel Library 11.2.3 that was shipped with MATLAB R2016a. This is a bug that is present only in MATLAB R2016a.

SOLUTION: 

Please upgrade to a different release of MATLAB in order to avoid this error.