A video on what makes modern processors so hard to study and predict performance.

Favorite quotes:

Your program…is going to be blips between cache misses.

The real new goal of all these hardware tricks is to run until you can get to the next cache miss.

Modern system architectures, and the timing differences between subsystems (Registers, L[1-3] caches, main memory, SSDs, HDs, and network) is just mindblowing.