Yesterday evening, Tesla official micro-Bo said, there's now more than 600,000 cars equipped with fully automated driving chip.
Which self-developed ASIC Tesla has 6000000000 transistor per 144 trillion times to complete the calculation, 2300 can process image frames per second at the same time. Each vehicle will be equipped with two optional Tesla chip, while processing the same data, which means that the car can handle twice as much data or image aspects of computing, will greatly improve the safety performance of vehicles.
Musk is still considered a good sum of accounts, as prices continue to Tesla bicycle close to the people, as a value-added lifting machines - worth 65,000 yuan to "fully automated driving" service is approaching us.
Author | Where snow
Edit | April
Cybertruck punk sci-fi wind is not yet over, "Iron Man" Musk embarrassed dance blamed Pa list of news headlines, "nail house" on the hot search, Tesla worthy of the name.
Yesterday evening, according to the official micro Tesla claimed that there are more than 600,000 of its vehicles with a "fully automated driving chip" of these vehicles create 20 million miles with data every day, providing training for Tesla huge neural network data to help neural networks continue to "learn."
News was released, "Tesla autopilot chip" has become the hot search keywords again.
Tesla has self-developed ASIC 6000000000 transistor per 144 trillion times to complete the calculation, the image 2300 can process per second at the same time. Optional Tesla per chip will be mounted on two simultaneously process the same data. Tesla said it ensures low latency characteristics of their course of action to better ensure that the system error does not occur, greatly improved safety performance.
Last week, Tesla CEO elon Musk on Twitter said, fully automated driving will "soon" launch. In fact, this commitment should be honored at the end of last year. Musk has more than once stressed that although this feature can be achieved almost L4 level autopilot, but does not rule out a small number of cases requiring the driver manual intervention. Also, this feature need to accept government regulation.
It now appears that, although there is no Musk promised again time node, so as not to hit the face, but "soon" word may not be true, worth the wait.
"This is very valuable, it will add tens of thousands of dollars in value for Tesla, Tesla and make the car through Network taxi service for the owners to make money." Musk said.
While prices continue to Tesla cycling close to the people, as the added value of lifting machine "complete autopilot" function is approaching us. According to Tesla's official website information, "fully automated driving capabilities" option price of 46,300 yuan, but if the installation after delivery, the price will rise to 6.5 million.
1 From outsourcing to self-study: self-developed chips pave the way for Tesla L5 level autopilot
For electric vehicles, as most will do the autopilot brand, mature intelligent driving technology is the core of Tesla.
Generally, the intelligent driving assistance system covers three areas, namely, sensing layer , by the surrounding environment sensor senses a vehicle body; decision- making appropriate decisions through information collected perception layer (involving chip / algorithm); executives , by receiving real-time sensor information, and chip / decision-making algorithm derived signal in order to take action including brake driving / warnings, etc.,.
These three levels, the decision-making plays a "central" role should not be overlooked, and this is the reason Tesla has been constantly autopilot chip layout.
Currently, Tesla has released a number of autopilot function. Autopilot feature is available for highway driving, the vehicle can automatically assisted navigation driving on the highway, also automatically assist lane change on the highway. Smart Summon call feature may allow a low speed of the vehicle in the parking lot to the specified location.
Wherein Tesla Autopilot automatic driver assistance systems by first universal hardware with more level sensors L2, algorithm and chip parts, and then the SIM card data and applications remotely managed via OTA software upgrade unlocking the mobile communication air interface, automatic channel change to achieve specific functions, adaptive cruise like.
Tesla software and hardware both have an absolute right to control, so that it could continue to push technology iterations to achieve functional and open up new features. Software and hardware to move forward on two legs alternately, pushing the overall function of steady improvement.
The hardware side, a total of three upgrades, HW1.0 to HW2.0 primarily by increasing the number of sensors and depth of learning the perception dramatically. HW2.0 to HW3.0 carried out two major chip upgrade for the basic realization of fully automated driving L5 level required computing power.
Tesla using the first generation hardware HW1.0 Mobileye visually recognized when a chip, the information collecting image recognition techniques rely mainly on the stage of Mobileye, Mobileye camera data from the roof of the vehicle and a peripheral first radar radar only provide assistance information.
HW2.0 using the NVIDIA Drive PX 2 chip, the entire board is not high degree of integration, there is plenty of blank space.
HW2.5 unprecedented increase overall chip integration, increased CPU 4 on the basis of the structure prior to the motherboard, leap integration operator brings jump force, reaches the theoretical force level calculated Drive PX 2.
To HW3.0, Tesla began to enable self-developed automatic driving chip FSD, "fully autonomous computer" (full self-driving computer, hereinafter referred to as FSD), in computing layer have complete control over the force. FSD dual chip design, the force reaches a count 144TOPS, performance comparison HW2.5 increased 21-fold, while the power consumption is only increased by 25%. Drive Xavier theoretical performance of Nvidia's value far more than 21 TOPS, up to seven times.
This marks on the chip and algorithm layer, Tesla currently has a certain first-mover advantage --HW3.0 been fully able to cope with the autopilot L5 level required amount perception layer data and computing power, become Benitez pull autopilot core competitiveness of technology.
2 deconstruction FSD chip advantage: complete autopilot how to achieve?
First, the biggest feature FSD board design is a dual chip design for redundancy, to reduce the failure risks ribbon, while improving the safety and accuracy of image processing.
According to Musk's statement, FSD chip board made completely redundant, meaning that each of the functional areas HW3.0 can be damaged, and the entire hardware can still maintain normal operation.
Meanwhile, two processors are provided inside the motherboard, two chips on the same board power supply and a data channel is independent and mutual backup. Two chips on the same data analysis, mutual authentication, comparative analysis, and then come to a final conclusion, which greatly improves the safety and accuracy of image processing.
Secondly, FSD core strengths is a powerful image processing and high-speed data transmission capability.
, The FSD consists of a processor is responsible for processing data from a single general-purpose processor, a central processing point of view the CPU, a GPU is responsible for graphics processing, two responsible for learning and prediction depth neural processing unit and a built-in image processor NPU ISP composition.
GPU and ISP constitutes a smart driving AI chip protagonist, GPU unit for the graphics processing unit, the processor core work is to help complete the graphics and animation rendering, allowing users to obtain valid information on the screen. The image processor ISP effect mainly three primary colors RGB raw data generated camera image information is converted into complex.
In addition, the FSD data transfer rate of data is far greater than 8 Tesla equipped cameras are now generated by the sensor, then Tesla is upgraded to L5 level autopilot reserved enough space data transmission.
It is worth mentioning that the neural processing units NPU FSD huge storage capacity of the chip, the bandwidth fast. NPU is responsible for making deep learning process based on image data generated by the ISP model - but before that, the data will be stored in the SRAM.
A static random access memory can be simply understood as it is much faster than the speed of memory operation, but also much higher cost of the memory chip, is typically used in processing chip-level cache 1-3. FSD now has 32MB of cache, a huge capacity SRAM summed up as the biggest advantage of similar chips on the chip FSD comparison market.
Further, the chip Tesla Chief Engineer Pete Bannon said at least automatic driving process cache to achieve bandwidth of 1TB / sec, and the SRAM chips FSD can actually provide 2TB / sec of bandwidth.
3 "Made in China" from the "China Design": profit margins continue to uplift
As the world's largest-selling electric car prices, the 2018 Tesla delivered a total of 245,000 electric vehicles, electric vehicles worldwide in 2018 accounted for total sales of 12.3%, higher than BYD / Beijing Automotive New car prices and other traditional energy sources.
Thus, after breaking production bottlenecks, Tesla's technology advantages into sales advantage. After Tesla achieve localization of this year, reducing space to improve, raw materials, labor, freight, customs duties have discount space, gross margin is expected to further improve.
In addition to profit margins have increased substantially, according to foreign media reports, Ark Investment Management Investment Management analyst Gini Tasha (Tasha Keeney) today said that in the next five years, Tesla stock price per share is expected to reach $ 6,000.
As the self-study Tesla chips, got rid of relying on Nvidia, Mobileye and other international giants of control has always been, other core materials batteries, the sensor also has some discount space.
Dominant scheme using visual sensor, a millimeter wave radar / vehicle camera / total ultrasonic radar function achieved equivalent to or higher than the level of the laser radar. From the viewpoint of price comparison, a millimeter wave radar / vehicle camera / ultrasonic radar significant discount compared to laser radar, and therefore cost effective compression sensor.
Seen from the figure, the highest proportion of the cost of raw material costs, according to Morgan Stanley estimates, the current raw materials account for about 50% of the operating costs of the future with the production of Tesla's center of gravity shift to China, human-powered transport cost reduction, raw materials accounted for ratio will further increase to 70%.
2020 January 7, Tesla in Shanghai officially delivered the first batch of domestic model3, priced from 353,800 yuan to 29 million cases of breath. From start to delivery, it took only a year's time.
January 15, a "create original" China wind "Tesla, let the most beautiful Chinese art into the future of Tesla" - brief Careers behind the slogan is great ambitions of its layout in China, Tesla announced , Shanghai super factory was completed 10 months, welcomed the long-awaited "made in China" Tesla, for the realization of recruitment from "made in China" Tesla will set up a design center in China to change "Chinese design" of outstanding design talent, to create original "China wind" Tesla, and available worldwide.
Musk Imagine the Chinese market territory is "the future maybe there will be more Model series produced here, there may even be designed by the Chinese Tesla was born," but what geometry opportunity, perhaps Tesla still has a long way to go
