I haven't used GGPO but reading through the documentation, it looks like you would just use the amount of updates that your engine needs. I think that's 1 per frame. Looking at the diagram:
It looks like the prediction serves as the remote update for the most part. You would track player input locally, update with predictions, validate with real remote data, rollback when appropriate.
From the docs:
GGPO checks the quality of its prediction for previous frames every time it receives a remote input. As mentioned earlier, GGPO doesn't receive the inputs for player 2's first frame until player 1's fourth. At frame 4, GGPO notices that the inputs received from the network do not match the predicted inputs sent earlier. To resynchronize the two games, GGPO needs to undo the damage caused by running the game with incorrect inputs for 3 frames. It does this by asking the game engine to go back in time to a frame before the erroneously speculated inputs were sent (i.e. to "rollback" to a previous state). Once the previous state has been restored, GGPO asks the engine to move forward one frame at a time with the corrected input stream. These frames are shown in light blue. Your game engine should advance through these frames as quickly as possible with no visible effect to the user. For example, your video renderer should not draw these frames to the screen. Your audio renderer should ideally continue to generate audio, but it should not be rendered until after the rollback, at which point samples should start playing n frames in, where n is the current frame minus the frame where the sample was generated. Once your engine reaches the frame it was on before GGPO discovered the error, GGPO drops out of rollback mode and allows the game to proceed as normal. Frames 5 and 6 in the diagram show what happens when GGPO predicts correctly. Since the game state is correct, there's no reason to rollback.
I would see no reason to update more than once per frame, every frame.