Release 3.0 of NAMD retains the same well-known parallel scaling capabilities of previous versions while adding improved support for GPU acceleration. A new GPU-resident mode has been introduced that more than doubles simulation performance for a single GPU and scales more efficiently across NVIDIA DGX architectures than NAMD's GPU-offload mode provided in the earlier version 2.x releases. With GPU-resident mode, NAMD is capable of simulating small systems, such as the 24k-atom DHFR using AMBER force field parameters, at a rate of more than 1 microsecond of simulated time per day. Several advanced features and enhanced sampling methodologies are available for GPU-resident mode, such as alchemical free energy methods, the Colvars collective variables module, and TCL forces.