Are the cards physically in the same machine, or are they in a remote chassis and connected through the network?
I think on P2, Amazon's cards are remote, and there is pretty significant latency when using them for time-sensitive computing.
What kind of performance can we expect from GCP compared to AWS P2?