Tesla's fully autonomous driving FSD computer details released

In the early morning of April 23rd, Beijing time, Tesla's Automotive Investor Day was held in Palo Alto, California.Tesla released the long-awaited fully self-driving  3.0 hardware and the latest development progress live on the 2019 nomy Investor Day. Details of Tesla's full self-driving computer (hereinafter referred to as FSD computer), which is claimed to be the best self-driving chip in the world, have been released. Since the end of 2017, Tesla has been trying to make the first computer chip for self-driving cars, and now Tesla has finally released all the details about its FSD computer. The chip is currently being installed in every electric car on the Tesla production line. It has been running on the Tesla Model S and Model X for more than a month, and has been placed in the Model 3 for 10 days. It is indeed the most powerful chip so far. According to reports, this "best chip in the world" in Musk 's mouth is 260 square millimeters in size, has 6 billion transistors, has a dual-core neural network array, can run 36 trillion operations per second, and is manufactured using a 14-nanometer process. Compared with the previous generation of Autopilot hardware (driven by Nvidia hardware), the frame processing capacity per second has increased by 21 times, and the hardware cost per vehicle has been reduced by about 20%. Each FSD computer contains many components: 8 visual cameras, 12 ultrasonic sensors, radar , and this custom-designed fully autonomous dual-redundant FSD computer, the most important of which is Tesla's custom chip. Each FSD contains two chips, each with two accelerators specially designed to run neural networks. Neural networks are the artificial intelligence components that Tesla cars use to read road information.

The neural network accelerator on Tesla's new chip can process 2,100 frames of input images per second from the car's eight constantly running cameras. This is equivalent to 2.5 billion pixels per second. Thanks to the powerful chip lineup, the FSD computer can eventually achieve a computing power of 144TOPS. The design of the chip also reduces energy consumption, and its power consumption is 72 watts. Interestingly, the person in charge of developing this chip is Pete Bannon, a chip designer poached by Tesla from Apple. Pete Bannon has participated in the development of the iPhone 5 A5 chip and several iPhone chips. Tesla's entry into the field of chip design follows the strategy adopted by Apple and Google. Apple has long been designing its own processors for mobile devices to maximize battery life and performance, while Google customizes server chips for its artificial intelligence software. Since Tesla made up its mind to develop an autonomous driving computer for its own cars, Musk's invisible hat of "Apple fan" on his head can't be taken off.

"Lidar is finished" The sensation caused by the release of FSD computer is not limited to this. While promoting "the world's strongest" in a high-profile manner, Musk has basically criticized all opponents in the field of autonomous driving. Not only did he bring Nvidia Drive Xavier for comparison when introducing FSD, and demonstrated the ruthless crushing of the former over the latter, but he also made no secret of his attitude that artificial intelligence computer vision is better than lidar. In the post-conference session, Musk reiterated his attitude: no lidar. At the same time, he made a bold statement: "Any autonomous driving company that uses lidar is doomed to fail." Whether using lidar or computer vision, the purpose is to allow autonomous driving cars to "see clearly" the surrounding environment. At present, most companies involved in autonomous driving technology use lidar solutions, such as Google  Waymo , Baidu, UBER, Cruise, etc., including Apple's mysterious Titan Project, which has been rumored many times and mostly revolves around lidar solutions. Tesla insists on computer vision system - that is, camera + data + neural network, which has more advantages than lidar. However, this statement has also been opposed by people in the industry. The biggest premise of autonomous driving is safety, and the stability and accuracy of perception should be guaranteed under various conditions. LiDAR can accurately measure the relative distance between the edge of an object's contour and the device in the field of view. This contour information forms a so-called point cloud and draws a 3D environmental map with an accuracy of centimeters, thereby improving measurement accuracy. Autonomous driving technology is far from mature, and it is obviously unrealistic to abandon LiDAR.

Nvidia is not convinced: Is it interesting to praise one and step on the other? One day after Tesla released the FSD computer in a high-profile manner, Nvidia began to tactfully respond to its inaccurate comparison. NV said that the fixed-point computing performance of the DRIVE AGX Xavier chip cited by Tesla was 21TOPS, but it should actually be 30TOPS. At the same time, it is unfair for Tesla to compare a full circuit board (144TOPS) with two FSDs to a Xavier. What should be compared is the DRIVE AGX Pegasus that integrates two Xaviers and two GPUs, which is the complete platform for realizing autonomous driving. If we only talk about performance, the nominal performance of Pegasus is 320TOPS, which exceeds Tesla. Nvidia is now continuing to work in depth with business partners such as Volkswagen and ZF to try to dominate the commercialization of autonomous driving chips. DRIVE AutoPilot, based on the DRIVE AGX Xavier chip and its driver software platform (deep neural networks for processing perception and data from surround camera sensors), has entered two Tier 1 automotive suppliers, Continental and ZF, and is preparing to mass-produce some autonomous driving system functions in 2020. The high-end computing platform DRIVE AGX Pegasus is the chip that many companies are using for L4 autonomous driving tests. It has many partners and is considered to have broad prospects.

NVIDIA DRIVE AGX Xavier The emergence of Tesla FSD is also the biggest challenge for NVIDIA DRIVE AGX. Regarding the continued burning of money in the future, according to the industry: "Tesla FSD computer is indeed very exciting, but we cannot be too optimistic about linking it to fully autonomous driving." There is no doubt that processing power is very important for self-driving cars, but achieving fully autonomous driving is a major challenge that cannot be solved by processors alone. It takes a long time and a long road to achieve true autonomous driving. Tesla has started the research and development of autonomous driving chips. This idea breaks the traditional idea of ​​automobile manufacturing and has reference significance for future innovation. Of course, the key to choosing whose chip is finally chosen is not how high the performance is, but the maturity of commercial applications and the cost price of procurement.

For Tesla itself, the research and development, update, mass production and application of this heavy chip system will become a major project in the future, which will require countless efforts and money. Whether it can support this "world's strongest" depends on Tesla's efforts.