The DMA transfers 64-bit words per each bus clock cycle between the main shared memory (RDRAM) and the internal RSP 4K DMEM (or IMEM, to transfer code).

So it's quite fast, but you need to remember that the main RDRAM is shared among the main CPU and the whole RCP (eg: it's also used as video memory for textures and frame buffers by the RDP), so contention is really high.