Many companies are attempting to build custom hardware to do better execution (and training) of neural nets. It seems odd to me that Tesla’s nets are that much different from other people’s so as to justify their own silicon. To be better than a plain GPU, that I understand, but everybody knows that, and all the companies making chips (including Intel/MobilEye) know that.
The reason I am curious is that as we see, this chip was designed in 2016 — developing a new chip is a time consuming process, especially for companies new to the game. Your custom chip will be very good at the problem, as you saw it in 2016. When you sell it in 2019 through 2021 when your new chip is ready.
Will you beat the horizontal chip suppliers?