Tesla dives into Autopilot HW3 technical detailsby Justin King
The in-house chips outperform Nvidia's Drive Xavier hardware by a factor of seven.
Tesla has provide an in-depth look at its new Autopilot v3.0 hardware, detailing technical elements to highlight the leap in performance compared to the previous HW2.5 platform.
The company's lead chip designer, Pete Bannon, explains that Tesla's in-house chip achieved a 21-fold jump in TOPS performance. The former Apple engineer said all of the previous projects he has ever worked on could only achieve a maximum of a 3X increase in performance.
Tesla began developing its own neural-network processors in 2016. At the time, other companies involved in machine-learning hardware had simply repurposed GPUs and other architectures that were not optimized specifically to run NN operations.
Notably, the presentation claims that Autopilot HW3 outperforms Nvidia's Drive Xavier chips by a factor of seven. Nvidia's chips had been used in previous Autopilot hardware and are still common among self-driving systems under development by other automakers.
"All Tesla cars being produced right now have everything necessary for full self driving," said Tesla chief Elon Musk. "All you need to do is improve the software."