philw1776
Member
NICE VID
Probably wrong but what the hell.
No first hand knowhow here, but realizing that a human professional athete's perception to reaction performance is approximately two tenths of a second lets say 5 "frames" per second, then 25 fps if these #s are true would be 5 times faster than the best, non-distracted, 100% FOCUSED very fast reacting humans.If I remember and interpretted correctly (and I have approximately ZERO background in this stuff, so I may be incorrect):
I believe the switch to FSD Version 12 gave them back some overhead on the chips.
I can't seem to find the reference, but I have a memory that Elon(?) said with FSD 12, the system was processing all 8 camera inputs and producing vehicle control outputs at the equivalent of 51 frames per second. To me, that sounds like, if they wanted to, they could process at ~25 frames per second on just one of the two inference chips in HW3....and 25fps seems like it ought to be fast enough to handle the driving task.
My interpretation (again, plausibly very wrong) was that Tesla always intended to be end-to-end neural networks...so that is what the hardware is designed and optimized for. It took lots of development and new techniques to get there...but now that they have cut out the 300,000 lines of C++ code, the hardware is now able to be used more efficiently and make the best use of its capabilities.
Probably wrong but what the hell.