Paul van Gerven
23 April

Tesla announced it has been shipping cars equipped with self-driving technology powered by in-house designed chips. Eventually, the system should enable level 4 or 5 self-driving, which means the driver’s attention is only required under special circumstances or even not at all. Getting there will take time and many firmware updates, however.

Most of the IP used in the 6-billion transistor, 14nm ICs fabbed by Samsung is licensed, but Tesla designed the neural network accelerator block all by itself. CEO Elon Musk called it the “best chip in the world … by a huge margin”. He also claimed it would result in extremely safe self-driving cars and robot taxis.

Tesla uses two identical chips for redundancy. In some situations, they compare results from both processors before making a decision. Picture provided by Tesla

Tesla says the new self-driving system built around two of the AI chips provides a 21x performance gain at 80 percent of the cost compared to the NVidia-based hardware it was using before. NVidia, however, claims the comparison is unfair.

Tesla developing its own ICs is part of a larger trend. Tech behemoths such as Apple, Google and Facebook all have design operations of their own, partly to save costs by cutting out the middleman, partly to address their specific processing needs.