Yes, I process per-buffer, same as on CPU.

You need to find parallelism somewhere to make it worth it. This can be multiple independent channels/voices, one large simulation, one high quality simulation, a large neural network, solving PDEs, voxel simulation (https://www.youtube.com/watch?v=1bS7sHyfi58), additive synthesis, a multitude of FFTs...

Thanks for the answers!