During earthquake rupture, faults slip at velocities of cm/s to m/s. Fault friction at these velocities strongly influences dynamic rupture but is at present poorly constrained. We study friction of simulated fault gouge as a function of normal stress (&sgr;n=25 to 70 MPa) and load point velocity (V=0.001 to 10 mm/s). Layers of granular quartz (3 mm thick) are sheared between rough surfaces in a direct shear apparatus at ambient conditions. For a constant &sgr;n, we impose regular step changes in V throughout 20 mm net slip and monitor the frictional response. A striking observation at high velocity is a dramatic reduction in the instantaneous change in frictional strength for a step change in velocity (friction direct effect) with accumulated slip. Gouge layers dilate for a step increase in velocity, and the amount of dilation decreases with slip and is systematically greater at higher velocity. The steady state friction velocity dependence (a-b) evolves from strengthening to weakening with slip but is not significantly influenced by V or &sgr;n. Measurements of dilation imply that an additional mechanism, such as grain rolling, operates at high velocity and that the active shear zone narrows with slip. Data from slow (μm/s) and fast (mm/s) tests indicate a similar displacement dependent textural evolution and comparable comminution rates. Our experiments produce a distinct shear localization fabric and velocity weakening behavior despite limited net displacements and negligible shear heating. Under these conditions we find no evidence for the strong velocity weakening or low friction values predicted by some theoretical models of dynamic rupture. Thus certain mechanisms for strong frictional weakening, such as grain rolling, can likely be ruled out for the conditions of our study. ¿ 1999 American Geophysical Union