[snip] Is the current hardware powerful enough to run what is needed without being rat-holed into some crazy optimizations to get acceptable performance? All that remains to be seen.
By "crazy optimizations" did you mean "too high of an error rate"?
Neural networks aren't like traditional programming, there aren't 'crazy optimizations' you just adjust your "compilation" parameters. In the case you are talking about you tune the error rate so that it can run on the current hardware. Which leaves you with one question: is that error rate acceptable? If it isn't then your only option is to revisit the architecture, but re-architecting is still different than "crazy optimizations."
Can the current hardware (HW3) work? Based on my understanding of the situation I would think that it is plausible, but unknowable until Tesla gets it working. I base the "plausible" on the performance gap between HW2 and HW3, and HW2 is currently still capable (though there are indications the limits are being hit).
If HW3 is not sufficient the question becomes will the second generation FSD chip be good enough? That is already under development and IMO may be in production before Tesla has a self driving. After all, chip design is a solved problem -- while a project can certainly slip there is no reason to not expect it to be completed. On the other hand self driving is
not a solved problem so no one really knows how long it will take. This is why I expect Waymo (or maybe even Cruise) to get approved in at least one market before Tesla. The problem for them is that their growth into new markets can reasonably be expected to be slow whereas once Tesla has working FSD there's no more development required (not that I expect them to stop) and its just a matter of convincing the regulators.
IMO the correct way to value Tesla's FSD program is as driver assistance where it has a real present value as opposed to an estimated value at some unknowable point in the future.
Looked at this way, a Tesla vehicle derives significantly greater value from its driver assistance than other vehicles and the lead appears to be increasing. This is a competitive advantage that seems likely to remain for the foreseeable future. No one else is trying seriously: Waymo and Cruise are about FSD, no driver assistance component. IIRC VW is trying to do something, but they still think software problems can be solved by hiring more developers.