Eigenvalue / vector errors with 2015 MKL

[Duane_T_]

Hi:

One of tool vendors was using the 2015 MKL, possibly version 11.2.2, which produced incorrect eigenvalues and eigenvectors for large 16k x 16k matrices on our servers.  The vendor has not been able to reproduce the problem using the same code and data on their servers.  Are there any known eigenvalue / eigenvector MKL 2015 bugs for large matrices?  Could this have a hardware, processor, dependency?

I couldn't find anything in your release notes or bug fixes.  A Matlab forum mentioned an eigenvalue problem with MKL 11.2.3.

https://www.mathworks.com/matlabcentral/answers/375474-erroneous-eigenvalue-computation-for-large-matrices-in-r2016a-when-forcing-single-threaded-execution

Thanks,

Duane

This is from one our servers that shows the problem:

processor    : 0
vendor_id    : GenuineIntel
cpu family    : 6
model        : 63
model name    : Intel(R) Xeon(R) CPU E5-2667 v3 @ 3.20GHz
stepping    : 2
microcode    : 58
cpu MHz        : 3201.000

This is from our vendor's server that doesn't show the problem

processor    : 0
vendor_id    : GenuineIntel
cpu family    : 6
model        : 62
model name    : Intel(R) Xeon(R) CPU E5-2690 v2 @ 3.00GHz
stepping    : 4
microcode    : 1064
cpu MHz        : 3000.218

[Konstantin_A_Intel]

Hi Duane,

I can say that we fixed a few issues in MKL eigensolver after MKL 11.2 release. It's too hard to make this kind of retrospective analysis and to give a precise answer to your question. I'd like to avoid any kind of speculation here. We will be glad to investigate the problem if it's still reproduced with one of the latest MKL releases. Please let us know if that's the case providing some more details helping to reproduce the issue.

Regards,

Konstantin

[Duane_T_]

Hi Konstantin:

Thank you for getting back to me.  

Where can I find all of the MKL bug fixes?  Is that something that's only accessible with paid support?

The MKL release notes are here, correct?

https://software.intel.com/en-us/mkl/documentation

Thanks,

Duane

[Gennady_F_Intel]

actually we added the Fug fix List into each of Release Notes. specifically for the latest version of mkl 2018, you may take a look this follow this link - https://software.intel.com/en-us/articles/intel-math-kernel-library-intel-mkl-2018-release-notes

[Duane_T_]

Hi Gennady:

Thank you for the link.  We think we are seeing 2015 MKL eigenvalue errors with specific CPU model numbers.  I'll review the release notes.

Regards,

Duane