When it's working correcly it's not running redundant, that's true. But we don't really how they handle/will handle cases where 1 neural processor fails. I see a few different options:Last I knew the FSD computer isn't running redundantly either as they've had to (for a couple years now) needed to use cross-node compute since they'd run out of compute on a single node quite some time ago.
1. Have a 50% smaller neural network ready to be loaded for these situations, run of the previous output for 1 cycle while the new network catches up ie run blind for ~0.06s instead of 0.03s when a neural network processor fails once.
2. Keep running the same neural network but instead of 36FPS only run it on 18FPS. Sure 0.06s reaction time is worse than 0.03s, but it still beats humans and is ~cat reaction time.
They only need to do this until they can safely stop the car on the side of the road. So the slightly increased risk of accident per mile will not be for millions of miles but for single digit miles thus the net addition of risk is insignificant in the scheme of things.
Last edited: