To maximize performance in NDWidget we need to have an async architecture that submits tasks to the GPU and fetches results afterwards. Users can follow a specific scheme such that their pytorch object, decord object, etc, submits async tasks to the GPU, other threads/processes, etc. and we use a standard interface/architecture for fetching the results and displaying in the NDWidget.

Things to checkout:

• https://docs.python.org/3/library/concurrent.futures.html#concurrent.futures.Future
• https://docs.dask.org/en/latest/futures.html
• https://docs.pytorch.org/docs/stable/futures.html