An on-chip variability characterization system implemented in a 45-nm CMOS process is used for direct time-domain measurements of random telegraph noise (RTN) in small-area devices. A procedure for automated extraction of RTN parameters from large volumes of measured data is developed. Statistics for number of traps, NT, and single-trap amplitudes, ΔVth, are studied across device polarity, bias, and gate area. A Poisson distribution is used to model NT and a log-normal distribution is used to model ΔVth. The scaling of the two statistics across gate dimensions is discussed; the expected value of NT is shown to scale with (L −ΔL)−1, whereas the expected value of ΔVth is shown to scale with W−1(L −ΔL)−0.5. The two statistics are combined in a compact RTN probabilistic model representing the statistics of the overall ΔVth fluctuations because of RTN. This model is demonstrated to give accurate predictions of the tails of the measured RTN distributions at the 95th percentile level, which scale with W−1(L −ΔL)−1.5. A comparison between nMOS and pMOS devices shows that pMOS devices exhibit both a higher average number of traps and a larger average single-trap ΔVth amplitude, leading to a comparatively larger overall impact of RTN.