Reading that was rather vindicating. The same concern is raised by phk in his B-heap essay [1], where he criticizes Knuth for the exact same reason -- that algorithm analysis has become disconnected from the characteristics of the underlying hardware.
While phk has a point, the dig at Knuth's work feels a bit unfair. In any case, I'm curious to see whether future editions of TAOCP note this caveat in its analysis.