There has been some criticism of Tesla’s decision to only have one nVidia Drive PX 2 GPU in their Autopilot Hardware 2 system as opposed to the 2 or 4 GPUs recommended by nVidia. While one theory is that Tesla will simply swap out the current 1 GPU board for a 2 or 4 GPU board as Level 5 autonomy gets closer to release, a more likely possibility is that the Tesla Level 5 system will actually be an internet-connected, cloud-based solution.
Imagine a scenario where a car is traveling down an interstate with minimal traffic. A single GPU is more than enough to keep the car safely in lane. In fact, some of the idle GPU cycles can actually be used by other cars on the network. On the other hand, imagine the most challenging scenario for an autonomous car, where the unexpected is most likely to happen and where the stakes are highest: elementary school pickup and drop off.
In a setting like this, each car might need 4, 8, or even 16 GPUs to safely navigate in and around the school. A cloud-based solution would allow dynamic scaling of the GPUs as well as data sharing among cars that are in the area or recently passed through the area. Even the cameras, radars, and ultrasonics of parked cars would be contributing to the network. Of course the internet isn’t 100% reliable so Tesla could install solar-powered, battery-backed auxiliary control towers at schools, hospitals, stadiums, major intersections, and other critical locations. These control towers would include cameras, radar, and other sensors. Maybe in the early days of Level 5, there would even be a human or two supervising these towers. Cars passing into the auxiliary control zone would connect to a local network and have access to the additional GPUs of the control tower plus the GPUs of nearby cars, even if the internet were to be unavailable.
Another advantage of a cloud based Level 5 solution is that all GPUs don’t have to be the same. At some point there will be an nVidia Drive PX3. Or maybe another company (Tesla itself?) will create a compelling next generation GPU. While Tesla designed the motherboard to be swappable, whole fleet performance would improve even with new GPUs just at the server level or just in the newest cars.
So the answer to the question of how many GPUs are needed for Level 5 autonomy is… “as many as it takes”.