The transceiver separations required for synthesizing full rank MIMO matrices in line of sight (LoS) geometries scale as the square root of the product of carrier wavelength
and range. The wavelengths at millimeter (mm) wave carrier frequencies are small therefore enable LoS spatial multiplexing with practical node form factors at ranges of 10-100 m, depending on the carrier frequency. However, such LoS MIMO links become frequency selective even with small geometric mismatches. Exact channel inversion in an
$N \times N$ MIMO system requires fractionally spaced equalization, which is practically infeasible when operating at the very high data rates (multiple Gbps) that we are interested in. In this paper, we investigate spatial oversampling (more receive antennas than transmitted data streams) with symbol rate sampling, introducing {\it designed} delay diversity across different receive antennas, as a means for removing error floors when linearly separating the spatially multiplexed streams. We study the tradeoff between the number of additional receive antennas and the complexity of temporal equalization, and argue that an attractive example architecture, compatible with form factor constraints, is one in which the number of receive antennas is double the number of transmitted data streams.