We investigate asymptotic capacity limits of the Gaussian MIMO broadcast channel (BC) with spatially correlated fading to understand when and how much transmit correlation helps the capacity. By imposing a structure on channel covariances (equivalently, transmit correlations at the transmitter side) of users, also referred to as \emph{transmit correlation diversity}, the impact of transmit correlation on the power gain of MIMO BCs is characterized in several regimes of system parameters, with a particular interest in the large-scale array (or massive MIMO) regime. Taking the cost for downlink training into account, we provide asymptotic capacity bounds of multiuser MIMO downlink systems to see how transmit correlation diversity affects the system multiplexing gain. We make use of the notion of joint spatial division and multiplexing (JSDM) to derive the capacity bounds. It is advocated in this paper that transmit correlation diversity may be of use to significantly increase multiplexing gain as well as power gain in multiuser MIMO systems. In particular, the new type of diversity in wireless communications is shown to improve the system multiplexing gain up to by a factor of the number of degrees of such diversity. Finally, performance limits of conventional large-scale MIMO systems not exploiting transmit correlation are also characterized.