Measuring IPC Clock-for-Clock

[OC-FTW]

One thing that has been bothering me recently is having to measure instructions per clock at the same CPU frequency for all processors. If the processor is stuck at a certain static frequency, whether it is 2.5GHz or 2.8GHz and never changes, then shouldn't IPC comparison be valid across different processors set to different static frequency?

 

[defaultluser]

One thing that has been bothering me recently is having to measure instructions per clock at the same CPU frequency for all processors. If the processor is stuck at a certain static frequency, whether it is 2.5GHz or 2.8GHz and never changes, then shouldn't IPC comparison be valid across different processors set to different static frequency?

Yes, they should. But there are two reasons most sites prefer fixed frequency runs.

1. You see different turbo frequencies for different load levels. This means each benchmark will need you to provide an average core frequency to compute IPC.

2. It's easier for the reader to compare performance at thew same frequency, rather than having to do the math (performance/clock scaling between different frequencies) in their head. This is why smart reviewers will just perform all test at a fixed frequency.