02 Aug 2018
Tesla is making AI chips in-house for its self-driving cars
Up until now, Tesla had been relying on NVIDIA's Drive technology to power self-driving capabilities in its Model S, Model X, and Model 3 cars.
Tesla had been developing the chips stealthily for 2-3 years
Known as Hardware 3 or the Tesla computer, the Tesla-built AI chips had been in development "in semi-stealth mode...basically for the last 2-3 years", said CEO Elon Musk during the earnings call.
The chips are meant to replace NVIDIA Drive tech in the aforementioned car models for do all the number-crunching required for advancing self-driving capabilities of these, and future models.
Hardware 3 is much more capable, claims Musk
Tesla's motivation behind developing chips in-house was purpose-specific specialization and improving efficiency.
According to Musk, Tesla's Hardware 3 demonstrated dramatic performance improvements over the current practice of running Tesla's computer vision on NVIDIA's hardware.
Notably, whereas NVIDIA's hardware was crunching 200 frames per second, Hardware 3 demonstrated the ability to crunch 2,000 frames per second "with full redundancy and failover".
AI analyst James Wang thinks in-house development is important
4/ Apple's custom SoC effort is a key reason why the iPhone leads in performance and go-to-market. Tesla's AI chip effort will do the same—they will be able to deliver performance features *ahead* of other auto makers that must wait for the next chip from Nvidia/Intel.— James Wang (@jwangARK) August 1, 2018
Building for its own needs could be real strength
The most important thing about Hardware 3 is that it allows Tesla to build for its own specific needs at its own pace - if they think current hardware is lacking, they don't have to wait for third-parties to build the desired improvements.
With Musk claiming that Tesla's chips cost the same as bought hardware, it might turn out to be a real strength.
What drove Tesla to design its own AI chips
"The key is to be able to run the neural network at a fundamental, bare metal level. You have to do these calculations in the circuit itself, not in some sort of emulation mode, which is how a GPU or CPU would operate," explained Musk.