A method of fabricating a gate structure (such as a floating gate) of a nonvolatile (e.g., flash) memory is described. After a polysilicon layer and a mask layer (e.g., silicon nitride) are formed on a semiconductor substrate, the silicon nitride layer is patterned and the polysilicon layer is partially etched. Then, a sidewall spacer is formed on sidewalls of the partially etched polysilicon layer and the patterned mask layer. The partially etched polysilicon layer is then fully etched, maintaining a partially etched shape at its top edge due to the sidewall spacer. The mask layer and the sidewall spacer are removed, to form a floating gate having a near-round edge shape. After full etching, the polysilicon layer may be heat-treated such that its top edge shape may become more rounded, fluent and/or stress- and/or strain-relieving.