How significantly does the platform Matlab is run on affect numerical precision?

I'm running a differential corrections algorithm in Matlab on both OS X (10.10.2) and Linux Mint (Trusty 14.04.2, Linux 3.16.1). This algorithm uses ode113 to run a series of numerical integrations and inverts a few matrices that have a few hundred rows and columns. When I compare the output from the algorithm on the two different architectures, I get different results on the order of 1e-11. Both systems are 64-bit, and are running the same version of Matlab. I checked the output of eps on both systems, and they are the same at about 2.22e-16.

Is there some architectural difference between the Linux and OS X versions of Matlab that can explain this numerical difference?

Answer 1

[Chris McComb]

This difference could be a result of slightly different BLAS implementations.

Somewhat related.