Because in its current state Bohrium carries some computational overhead, this mostly depends on your problem size and the architecture you want to use. As a rule of thumb, switching from NumPy to Bohrium is beneficial if your set-up contains at least 1,000,000 elements (total number of elements in a 3-dimensional array, i.e., $$n_x n_y n_z$$). You can also use our benchmarks for general orientation.