On the occasion of the Tesla Autonomy Day , an invitation-only event held in Palo Alto, California, Tesla revealed all the details about his Full Self-Driving computer , previously known as Autopilot 3 . It has been announced that FSD is now in production and that it is located within all the Model S and Model X models produced in March and in the Model 3 vehicles produced starting from 12 April. As the name itself underlines, it conveys the promise of total autonomous driving entirely managed through the hardware present in the car. We are talking about a computer with two separate SoCs that manage two neural networks for a reason of calculation redundancy.
During the conference, Elon Musk and Pete Bannon , Tesla's hardware and Autopilot project manager, explained how the new chip works and why the company decided to produce its own technology, abandoning NVIDIA Drive Xavier , which was at the base of the previous generation of Autopilot.
Musk claims to have felt the need to use a chip that had been developed from scratch for the neural network , which did not correspond to previously existing technologies. Hardware was then associated with a custom designed software on it. Musk has been defining this technology as "the best chip in the world" for months, and during the event showed slides that highlight an important performance gap compared to previous generation NVIDIA technologies. FSD would be able to process 21 times more frames per second than the previous Tesla Autopilot hardware.
Bannon, who worked with Apple among other companies, also explained how power problems were solved. FSD actually consumes more than the previous generation (72W against 57W) but still remains within a range of consumption that allows Tesla to power it without having to install a larger capacity battery. Bannon also stated that the production of the new hardware, compared to the previous generation hardware, costs 20% less for each individual car .
Musk added that Samsung is producing Tesla's new chip in Austin, Texas, and is not expecting any problems with the supply. Furthermore, production savings are directly invested in the next generation of chips, which Musk expects to perform three times better than the one just introduced.
The FSD die has a size of 260 square millimeters and has 6 billion transistors . Naturally the main objective of the system is to manage the flow of data coming from the sensors installed on the self-driving car through a computational model based on two separate neural networks . The new Full Self-Driving computer is based on a logic focused on redundancy, also with regard to the power supply and the calculations: the system, in fact, performs the calculations separately twice using the two neural networks on which it is based and compares them before providing to steer the machine. "The probability that this computer fails is less than the likelihood that a human being will fail," Musk said.
We are talking about a SoC built with a 14 nanometer FinFET- type construction process , equipped with 32 MB of SRAM. The GPU runs at 1 Ghz and is capable of processing 600 gigaflops per second, both in FP16 and in FP32 . As for the CPU part, we have 12 ARM A72 cores operating at 2.2 Ghz: this is the micro-architecture that follows ARM A57, already a reference point in its generation, and that implements the ARMv8-A set of instructions 64-bit.
Installing an FSD computer on a Tesla costs the customer $ 5,000 and allows the vehicle to change lane automatically, to manage the car at motorway junctions and to proactively take the exits. Later in the year, according to Tesla's promises, the cars equipped with FSD will be able to manage themselves automatically in the cities , to recognize traffic lights and stops, and to act accordingly.
When activating the new service powered by FSD, and called Navigate on Autopilot , the car displays a message reminding the driver that he is still fully responsible for the vehicle. If the driver removes his hands from the steering wheel, he also risks disabling the system.
Despite this, Musk points out that the new Tesla with FSD from the hardware point of view are ready to drive completely autonomously . FSD, in fact, receives information on what surrounds the car from eight cameras that have a 360-degree view around the vehicle , which work together with a front radar and short-range ultrasonic sensors . According to Musk, therefore, it would be enough to push a button to release a software update that makes all these cars completely autonomous.
According to other experts it takes at least 10 years for these vehicles to be managed in a totally disengaged way compared to being human. Also because Tesla vehicles do not use LIDAR sensors . These sensors, used by other automakers such as Waymo, GM and Ford , emit a beam of light to detect the distance to which an object is located. According to some, not preparing the vehicle for a LIDAR system would reduce accuracy. This is not the case for Musk because LIDAR would be unnecessarily expensive and cumbersome to install on Tesla vehicles.
Musk's statements at the Tesla Autonomy Day event (see it below in its entirety, about 4 hours) are doing a lot of discussion in various environments on the network, especially the feasibility of full autonomous driving right away. The promises are certainly very important: we'll see if the first tests show a concrete benefit for drivers linked to the presence of the new computer.