Tesla has revealed more information about the ‘Full Self-Driving Computer’, which was announced earlier this month that it is in production. It is a self-designed chip, a neural accelerator, which, according to the company, is ‘objectively the best chip in the world’.
During the presentation to investors, it was claimed that the chip enables a 21-fold improvement over Nvidia’s previous-generation Tesla Autopilot hardware. Compared to the Tesla Autopilot 2.5, the hardware would also cost 20 percent less. The chips, which should enable the fully self-driving functions via an over-the-air update, are made by Samsung using a 14nm finfet process. According to Musk, Tesla has concluded that there is no chip built from the ground up for specific neural networks. Based on this, it was decided to develop a chip ourselves to enable the self-driving functionality.
Pete Bannon, the man who heads up the Autopilot hardware at Tesla, said during the presentation that the chip can process 2100 images per second. These images come from the eight continuously operating cameras that Tesla cars have on board. This equates to 2.5 billion pixels per second. This input must be processed by the image signal processor, which goes at 1 gigapixel per second through a 24bit pipeline. Tone mapping and noise reduction can also be applied here. The lpddr4 and the 128-bit memory bus provide a peak bandwidth of 68GB/s. For this image information there is the neural network processor which consists of 9216 units. These run at 2GHz and reach 36 tops per chip, so 72 tops in total since there are two accelerators. Furthermore, there is support for h.265.
During the presentation, Tesla compared its new chip with Nvidia’s Drive Xavier chip. Tesla’s chip comes out at 144 tops and Nvidia’s at 21 tops, the car company claims. However, Nvidia has some criticism on this and does not think it is a fair comparison. A spokesperson for Nvidia told Marketwatch that only the Nvidia Drive AGX Pegasus would have had a correct comparison. According to Nvidia, this one reaches 320tops. In addition, the spokesperson reports that the Xavier chip does not reach 21 but 30 tops and is in fact only made for certain auxiliary systems and not for fully autonomous driving.
These new Tesla chips are in all Model S and X cars manufactured since March and Model 3 units produced since April 12. Existing Tesla owners who purchase the ‘Full Self-Driving package’ in the coming months will still receive the hardware. The computers are mounted behind the glove compartment and consist of two chips, so that the system continues to work if one of the two gives up the ghost. To this end, a separate power supply has also been used for each of the chips.
In addition to the details about the Full Self-Driving Computer, Musk has reiterated that Tesla continues to focus on cameras and radar for autonomous driving. Lidars, which many other companies see as crucial to fully autonomous driving, are once again rejected by Musk. “Anyone who depends on lidar is doomed to fail.” He finds it expensive and unnecessary. The CEO does acknowledge that it is still a significant challenge to get the software to the point where full autonomy is a fact, but that should be a fact with the robot taxis by the end of 2020.
The claim that the Full Self-Driving Computer would be 20 percent cheaper than the previous hardware is striking. Elon Musk announced last week that the price of the Full Self-Driving upgrade for Tesla cars will rise steadily after May 1. The increase would be around three thousand dollars.