Matlab code giving wrong answers on system with i9 processor

Hello,

We have many different types of computers with Matlab installed on them. One computer in particular seems to be doing calculations wrong.

This system is an HP Zbook with an intel i9 9880H processor in it. The system has had Matlab R2017b installed on it which has been removed and reinstalled multiple times with the same exact results.

Have you seen anything like this or have any idea what may be causing it?

Thank you,

Mike

[Walter Roberson]

Are any of the other systems using the same matlab release on a core-i series processor?

[James Tursa]

Can you isolate the calculation that gives different results and post it?

[Mike Collins]

Yes, they are all on either i7 or newer Xeon processors all made by Dell, some Desktops, some laptops. This new HP Zbook with an i9 is our first test with an Intel i9 and our first instance of this issue.

Answer 1

[Mike Collins]

The issue was found to be related to the hardware. A technician had taken the hdd out of the HP Zbook and installed it into a different computer (Dell Precision). The issue was then resolved. Something with the hardware, either the processor or the board was handling floating points differently. Thank you all for your input!

[Walter Roberson]

!!

There are hardware specific drivers that can get installed, and if the Dell driver for a different processor got installed on the HP hard drive and then it was brought back to the HP, then Yes, things could go Rwong,

[Mike Collins]

That is not what happened. The HP Zbook was having the issues with HP drivers installed. The system was simply not handing floating points properly.

[Walter Roberson]

Even stranger!!

So, just to confirm, the i9 processor works in the HP with the Dell drivers ?

[Mike Collins]

No, the i9 had problems in the HP no matter what. Only when we popped the hard drive into the Dell is when we got the correct answer. The issue is hardware based with this HP Zbook and we did not find a solution for it besides use a different system. We know that it is hardware based because it is the same windows instance working on the dell with everything else staying equal except for the HP hardware.

[Walter Roberson]

Then we seem to be back to the issue "Matlab code giving wrong answers on system with i9 processor" as you do not yet have MATLAB working on an i9 system, and you have not as yet established that the i9 system gives correct floating point values outside of MATLAB.

Do you just happen to have MATLAB Coder? By converting the calculations to C / C++ code then that would help isolate the level at which the wrong answer is being produced. Alternately, if you could put together a little mex to be called from MATLAB that does the same calculation as the MATLAB level, then that could again help isolate the level at which the problems are showing up.

[Walter Roberson]

One question I would have is whether the "errors" occur on all array sizes, or only on "small" arrays, or only on "larger" arrays. I ask because for "larger" arrays, MATLAB tends to call BLAS or LAPACK or MKL functions to do highly optimized calculations, but for "smaller" arrays, MATLAB executes the operations itself.

[Star Strider]

That this appears to occur with an Intel chip brings to mind Pentium Division Bug Revisited.