Reports from CNBC started circulating that AMD and Tesla are working together to build a custom processor for Tesla’s autonomous Autopilot in its cars. This would reduce the reliance on NVIDIA GPU technology and would give Tesla more verticality in its product stack. CNBC claims that Tesla is far enough along on this implementation that it has samples back from AMD for evaluation.
Current Tesla vehicles, including the Model S, Model X, and presumably the Model 3, ship with semi-autonomous capabilities that are powered by NVIDIA. The hardware has been upgraded through a few iterations since the cars initial release, with the most recent changes including a discrete (stand-alone) ARM-based central processor and a discrete GP106 GPU. The GP106 GPU is the same used in the consumer gaming GeForce GTX 1060 product. That helps us understand the compute requirements for Tesla to offer higher levels of autonomy on its platforms.
Tesla brought in Jim Keller to lead its Autopilot development in January of 2016. Keller had previously been with AMD and led the development of the Zen architecture, used in modern AMD processors like Ryzen, Threadripper, and EPYC. He had previous stints at Apple designing the company’s mobile processors and another run at AMD during the height of its Athlon 64 success.
It makes sense that Keller would lead a team, rumored to be around 50 engineers strong today, to build a customized processor that can more directly and efficiently target the workloads required for self-driving technology. Removing or minimizing dependencies on external companies (including processing and sensor tech) limits the risk of being marginalized as autonomous vehicles take off in coming years.
A partnership with AMD brings with it a lot of positives for Tesla. AMD has the experience and the product portfolio to offer solutions that no one else in the industry can. The semi-custom group at AMD has been successful in attracting top projects. Both the Xbox One (and upcoming Xbox One X) from Microsoft and the PlayStation 4 from Sony utilize AMD semi-custom designs, combining AMD central processor cores with AMD graphics technology into a single chip. Each has customizations that make them unique for the customer’s targeted price and performance. Both Microsoft and Sony are repeat customers to AMD for this service.
Design specifications from Tesla would likely match or exceed the performance requirements for the new Xbox One X. Even NVIDIA has acknowledged that true autonomous driving capability will need more GPU compute horsepower than its first generation Drive PX system can provide. NVIDIA has recommended adding a second GPU to the Drive PX system, and some indications are that Tesla has started doing that with its most recent production vehicles.
Today’s AMD IP portfolio has both high performance processor and graphics technology, including the Zen central processor architecture (used in EPYC enterprise and Ryzen consumer parts) and Vega graphics architecture. Tesla could combine these to build a more powerful solution than it current integrates with NVIDIA hardware, but the balance will be around cost and performance requirements.
NVIDIA graphics processors have an advantage over AMD equivalents in power efficiency (how much compute each can accomplish in a given power envelope). While critical for some PC and gaming applications, the battery on an 85 kilowatt-hour Model S could power the 120 watts of a GP106 GPU (used in the current Tesla Autopilot system) for nearly 30 days, running 24/7. If Tesla has to sacrifice some efficiency to gain control over the processor for its cars, it will not affect range.
Even if Tesla backs away from NVIDIA for its vehicle-based compute solution, it could still utilize its graphics processors for self-driving training. Training of a neural network like the one used for autonomous driving is done in a central location with larger, powerful servers, building the data set and algorithms to enable safe and reliable driving behavior. The computer inside each Tesla is responsible for applying those algorithms to the real-time data coming from the cameras, radar, and other sensors on the vehicle.
But for NVIDIA, the size of the autonomous driving training market doesn’t scale as Tesla or other manufacturers sell more self-driving cars. Being at the heart of each car on the road is what every AI compute vendor sets as its goal.
NVIDIA has an experience advantage over AMD in building systems for autonomous driving and other AI applications, which should not be discounted. But Tesla has also been gaining that knowledge and if it decides to make its own custom processor for the job, regardless of the source, it will have the option to replace or compete with NVIDIA solutions.