Telsa is reportedly developing its own AI chip for automated cars with AMD.
Sanjay Jha, CEO of GlobalFoundries, a spin-off of AMD, spilled the beans at its conference in Santa Clara, California, CNBC reports. GlobalFoundries later said Jha didn't mean to explicitly name Tesla as a partner, but was instead pointing to Tesla as an "example" of a company working with chipmakers.
But the presence of Jim Keller—former chip designer for Apple, AMD, and Honda—at Tesla has fueled speculation that Tesla and GlobalFoundries are indeed working together. A source for CNBC adds that Tesla has received samples of the first implementation of its AI processor and is running tests now.
Tesla's current vehicles use Nvidia GPUs to power its Autopilot self-driving system. By making its own chips, Tesla would be able control costs, something it will need to do to meet its CEO-imposed deadline of fully automated cars by 2019.
- Driverless Cars: When the Internet Takes the Wheel Driverless Cars: When the Internet Takes the Wheel
Fully automated cars that match up to the Level 5 standard—i.e. no human input required to drive it—will need a lot of computing power in order to run cameras, spatial sensors, and mapping processes. And if they're going to be rolling off of the production lines at scale, they're going to have to be cheap, too.
In February, GlobalFoundries VP of IoT Rajeev Rajan said his company's 22nm FDX process was instrumental in delivering the world's first insulator for an Advanced Driver Assistance Systems (ADAS) SoC, capable of handling "360-degree top view, road-sign recognition, lane departure warning, driver distraction warning, blind spot detection, surround vision, flicker mitigation for digital mirroring, pedestrian detection, cruise control and emergency braking."
While ADAS SoC's are not new, the key thing here is the costs savings. At CES 2017, Alain Mutricy, senior vice president for product management at GlobalFoundries, said the 22nm FDX process delivers similar performance to an advanced 14nm FinFET non-planar transistor at the cost of a more basic 28nm planar.