I help maintain a Linux software project that uses MKL. For years, the MKL packages have been locked to a version from 2017, and keeping them up-to-date hasn't been high on the priority list. Hoping to realize significant performance improvements by jumping across the seven year technology gap, I set up the Intel One API repo, installed the most recent version of MKL, and measured CPU usage linking against the old vs. new MKL libraries. To my surprise, linking against the new libraries causes the MKL-using process to consume ~19% more overall CPU time, and ~3-5% higher core consumption. This is coming from a controlled test with identical settings and input data. Similar results can be reproduced on both newer (from 2024) and older (from 2017) computers with Intel CPUs. Does anyone have any clues on what might be causing this?
| MKL Version | TIME+ measured by htop |
| 2017.3 | 14.63 |
| 2024.2 | 17.35 |
