In the context of NCA certification, what is the most important consideration when implementing multi-gpu computing & scaling?