TSC is not per CPU but per CPU core

[ujos]

In the book there is following sentence:

Page 33 Every CPU has its own TSC

Based on my experience actually every CPU core has its own TSC. In order to get reliable duration one should bind his/her thread to the CPU core. Otherwise the duration may not be accurate.

[ujos]

Even on Windows if you call QueryPerformanceCounterfunction from the multiple threads you may get negative duration value