Speckles have been considered ubiquitous in all scattering-based coherent imaging technologies. However, as an optical-absorption-based coherent imaging technology, photoacoustic (PA) tomography (PAT) suppresses speckles by building up prominent boundary signals. We theoretically study the dependence of PAT speckles on the boundary roughness, which is quantified by the root-mean-squared value and the correlation length of the boundary height. Both the speckle visibility and the correlation coefficient between the reconstructed and actual boundaries are quantified. If the root-mean-squared height fluctuation is much greater than, and the height correlation length is much smaller than the imaging resolution, the reconstructed boundaries become fully developed speckles. In other words, speckle formation requires large uncorrelated height fluctuations within the resolution cell. The first- and second-order statistics of PAT speckles are also studied experimentally. While the amplitude of the speckles follows a Gaussian distribution, the autocorrelation of the speckle patterns tracks that of the system point spread function.