The Tesla chip for autopilot is already in production, according to CEO Elon Musk. The upcoming processor is designed to replace the current platform in vehicles that began shipping to the market in October 2016, and is designed to provide enough performance to collect data from existing sensors and enable full autonomous driving without driver assistance.
βA Tesla computer that supports fully autonomous driving and is already in production, such a task will only load 5% of the total computing power and 10% with maximum redundancy for reliability,β Mr. Musk said on Twitter, responding to a video in which one one of the owners is amazed by the new Navigate on Autopilot feature, which allows the car to gracefully exit the highway, but still requires the driver to maintain full attention.
It's a big step forward for the company, which has promised to bring fully autonomous driving to all of Tesla's latest vehicles eventually. Elon Musk argues that the current "Hardware 2", which includes 8 cameras, ultrasonic sensors and GPS receivers, is sufficient for fully autonomous driving at a later stage in the development of autopilot, although competitors like Waymo rely on lidar-assisted environmental scanning systems. During a keynote conference in August 2018, Tesla first announced its platform, which will replace the NVIDIA Drive PX2. In October 2018, Musk announced that the chip would appear in all of the company's new production vehicles in about six months.
The electronics are part of a package that Tesla calls "Hardware 3". At the time of the announcement, the company had already been developing the chip for three years - the task was entrusted to a team led by iPhone 5S processor developer Pete Bannon. The chip is designed to accelerate the neural network that underlies the autopilot.
While the current Drive PX2 platform can handle 20 frames per second, Tesla claims its own solution can handle 2000 frames with full redundancy to protect against crashes. This redundancy is key to making the vehicle respond safely by reducing errors. Elon Musk notes that his company's product provides for two single-chip systems (each with two neuroblocks) that work independently, for security purposes.
Two, independent system-on-chip architecture, with each SoC having two NN accelerators that can perform simultaneous health-check calculations to protect against a soft error
β Elon Musk (@elonmusk) April 8, 2019
NVIDIA's expertise in gaming graphics and highly parallel computing has proven to be extremely useful for the computational acceleration challenges associated with artificial intelligence and autopilot for cars. The Drive PX2 offers eight teraflops of performance, about six times that of the Xbox One. "I'm a big fan of NVIDIA, they're doing great stuff," Musk said during the chip's initial announcement. βBut when using a GPU, in fact, we are talking about emulation mode, and performance is limited by the bus bandwidth. Ultimately, the transfer of data between the GPU and the CPU limits the system.β
NVIDIA remained open to further collaboration with Tesla. CEO Jensen Huang said a few days after the announcement, "If it doesn't work, for whatever reason Tesla doesn't, you can give me a call and I'll be more than happy to help." Later that month, the company confirmed to Inverse that it was still working with Tesla.
Tesla sells the partial autopilot option for $3000 at the time of car purchase, or $4000 thereafter. A full autopilot costs an additional $5000 with the car or $7000 later. Mr. Musk claims that the new chip will be included in these costs. Now a more expensive package means support for features like Navigate on Autopilot, though it still requires the driver's full attention.
This year, Tesla promises to implement support for recognition and reaction to stop signs and traffic lights, as well as the ability to automatically drive on city streets, as part of a $ 5000 package. In the future, there will also be automatic change of lanes on the highway, automatic parallel and perpendicular parking, as well as remote call of a parked car to the driver. When needed, Tesla will replace NVIDIA electronics with its own solution free of charge to those who bought an expensive autopilot package.
Doesn't matter yet whether you have Tesla FSD computer or Tesla Nvidia-based computer. When it does matter later this year, Tesla will switch the computers for you.
β Elon Musk (@elonmusk) April 8, 2019
It's unclear when Tesla will be able to offer full point-to-point autopilot without any driver involvement. The company originally planned to complete work on autonomous driving between the two US coasts by the end of 2017 (primarily for trucks), but that task has been shelved in favor of developing a more versatile solution. Former infamous Google employee and co-founder of Otto (subsequently absorbed by Uber), Anthony Levandowski, claimed in December 2018 that he had achieved the goal of creating an autopilot to drive across the country before Tesla, and even published the corresponding video as evidence:
In February of this year, Elon Musk suggested that full autopilot would be safe enough by the end of next year. That's pretty soon, considering Volkswagen expects autonomous cars by 2021, and ARM's forecast for 2024 is more realistic. If Mr. Musk is right, the start of production of Tesla's specialized neuroprocessor is an important step in this direction.
Source: 3dnews.ru