Although a simple solution, pilot-symbols-trained LMS algorithm is not capable of tracking fast varying channels in UMTS FDD downlink due to insufficient adaptation rate. Pilot-chips-trained LMS adaptation is on the other hand much more prone to noise. These two phenomena manifest themselves in the two components of the adaptation excess mean square (EMSE). A compromise can be found by considering HSDPA symbol-level Griffiths or decisiondirected equalization. These two methods enable adapting 16 times more frequently than the pilot-symbols-aided adaptation. They are also attractive for implementation since a modular approach can be adopted by exploiting either one or more of the available HSDPA code domains depending on the instantaneous channel quality and the performance requirements. © 2009 IEEE.