Different Eigenvalues on Different Machines

Samuel_M_1 — Tue, 16 Sep 2014 23:34:24 GMT

Hello,

I am computing eigenvalues on different machines. With the exact same binary and the same shared libraries, I obtain different eigenvalues for near zero entries when I run on different machines. Is this expected behavior?

On a 64 core node:

   -2.7756e-17
   4.2301e-46
   3.4694e-18
   1.7984e-02
   9.4446e-01

On a 32 core node:

1.8526e-47
   3.4694e-18
   2.7756e-17
   1.7984e-02
   9.4446e-01

I suspect that MKL is auto-threading with different numbers of cores and this is causing the floating point operations to execute in different orders, but would like to make sure.

yes, You are right. this is

Gennady_F_Intel — Wed, 17 Sep 2014 09:05:57 GMT

yes, You are right. this is the expected behavior. to eliminate the problem you can try to apply CNR mode.

topic yes, You are right. this is in Intel® oneAPI Math Kernel Library

Different Eigenvalues on Different Machines

yes, You are right. this is