Did you check to see if your system clock was still showing the correct time of day?

I recall running a bench on an old machine, ages ago. It had a turbo button in it, but the machine design was stupid - the real-time clock update frequency was tied to the CPU frequency. So when you used turbo to make things faster, you couldn't measure the difference using the PC's clock, because it had counted faster as well.

More commonly, I've run into systems where a chunk of assembly code will disable interrupts while running a critical section; then the system clock loses updates and the benchmark measures much faster completion because very little time passed, according to the clock.