Recently, the United States Geologic Survey Earthquake Hazards Program developed a 3D seismic velocity model of Northern California based on geology, including a detailed model of the urbanized San Francisco Bay Area. In this study, we report comparisons of observed three-component broadband (0.03–0.25-Hz) waveforms with synthetic seismograms computed with the new 3D model using an elastic finite difference method. We selected a set of 12 moderate earthquakes (Mw 4.0–5.1) that occurred within the greater San Francisco Bay Area, having well-constrained source parameters and broadband recordings with high signal-to-noise ratios. The objective of this study was to investigate how well the 3D velocity model predicts observed waveforms and to identify features of the model that may require revision to improve the waveform fits and predictions of ground motions for future events. We show for the 3 September 2000 Yountville earthquake that reported source parameters accurately predict waveforms at two close strong-motion stations (approximately 10 km from the epicenter). By comparing synthetics for the average 1D model GIL7 (Stidham et al., 1999) and the 3D structures we show that the effects of seismic wave propagation in the 3D model become important for frequencies at and above 0.1 Hz (periods less than 10 sec). Comparison of observed and synthetic seismograms for the 3D model consistently shows that the model predicts energy arriving earlier than is observed, particularly for the surface waves, indicating that the shear velocities in the upper crust must be reduced. We cross correlated the observed and synthetic waveforms and recorded the delay time and linear correlation for best alignment of the data and delayed synthetic. The results indicate that generally the 3D model predicts the observed waveforms well (mean linear correlations 0.41) and includes features that arise from the interaction of the wave field with 3D structure, especially the major sedimentary basins in San Pablo Bay, San Francisco Bay, Santa Clara Valley, and Livermore Valley. Simple conversion of the observed delay times for optimal alignment suggests shear velocities should be reduced by 4%–5% on average. Based on these findings, we conclude that the model is an excellent first step, suggesting that the overall structure of the model is accurate (i.e., the basin and discontinuity geometry). However, the velocities must be reduced to improve the observed timing of surface-wave arrivals.