Recently, Tesla has brushed a wave of traffic in the automotive circle, such as news headlines such as "A Tesla makes 100,000 yuan" and "Profits soared by 658%". In fact, the information behind this is that Tesla released the first quarter of 2022, Tesla's total revenue reached $18.76 billion, an increase of 81% year-on-year. Gross margin of 29.1%, operating margin of 19.2%, both hit a new record high, envy and jealousy.
In the spirit of Vehicle not to eat melon masses, so I also went to check it out, why is Tesla so profitable now? I happened to see an article about the components of Tesla's autopilot system that analyzed the system Plus consulting. As a smart car, the sensors of automatic driving and domain controllers are self-evidently very expensive, for example, our previous article from the 2022 CES to see Mobileye autonomous driving product technology and strategy (who said that computing power is the only standard) revealed its goal in 2025 to reduce the full cost of autonomous driving to within 30,000 yuan, so you think about how much the cost is now? Take the pricing price of Weilai Automobile for about 400,000 yuan, if the BOM cost is half of 200,000, it is at least 15% of the cost according to the price after five years. I also want to know how expensive the smart driving hardware in smart cars is.
So the standard just borrows the System Plus Consulting Tesla AutoPilot System Parts Analysis Article from Tesla AutoPilot sensors and chips to see its high-margin secrets, perhaps can give us some thoughts.
Tesla is essentially based on a vision-led pure visual AI autopilot program without laser thunder and high-precision maps.
Therefore, the physical hardware of its Model 3 for autonomous driving is mainly:
sensor
8 cameras, 1 radar and 12 ultrasonic sensors (the detailed location of the arrangement can be found in the previous article Visual King-Xiaopeng and Tesla's autopilot solution)
8 cameras for a 360-degree field of view within a 250-meter radius. The front vision triple camera covers 250 meters, but the viewing angle is very narrow, the middle distance is 150 meters, the other front view is wide vision fisheye 60 meters, and the side and front view 2 cameras are 80 meters each, which can see the environment around the car, as well as road signs. The other 3 cameras face the side and rear of the car and can see up to 100 meters away.
12 ultrasonic sensors detect obstacles within an 8-meter radius around the car. It can work at any speed and also controls blind spots. Autopilot also uses data collected by sonar to manage automatic lane changes during overtaking.
A forward millimeter-wave radar system with a maximum detection range of 160 meters provides additional data about the surrounding environment at redundant wavelengths, which can penetrate heavy rain, fog, dust and even overtake the vehicle in front.
1 GPS is used for positioning
All eight cameras were released in 2015 by ON Semiconductor (the one that was forced to relocate its global distribution center to Singapore and Malaysia during the recent Shanghai outbreak) and have the same resolution of 1280×960 1.2 megapixel image sensors. "They cost less. They are neither new nor high-resolution. At the same time, you want to give you the next car wholesale 8 you always have to negotiate the price.
For the design of the triple camera in front of it, Tesla embeds all the CMOS sensors into the PCB board without having to deal with the SoC. Other current triple cameras such as BMW's latest 7 Series use ZF S-Cam4 with Omnivision CMOS image sensor and Mobileye'a EyeQ4 vision processor. So System Plus estimates the cost of ZF triple camera at $165, while Tesla's triple camera is $65, and Tesla's price is only 1/3 of its price.
Ultrasonic radar is not to talk about, I heard that the price of this thing is too low, as if it is the cost of less than 20 yuan.
Millimeter wave radar, Tesla uses Continental's mature radar module ARS4-B, which is a 77GHz radar chipset and 32-bit MCU provided by NXP. For forward collision warning, emergency brake assist, collision mitigation or adaptive cruise control (ACC). An important element is the simultaneous measurement of long distances (up to 250m, accuracy +/-0.2m) and short distances (up to 70m), relative velocities, and angles between two objects.
The system consists of two electronic boards, including an NXP Semiconductor microcontroller and a Broadcom Ethernet transceiver. The radio frequency (RF) board is made of an asymmetrical structure with a hybrid PTFE/FR4 substrate with a planar antenna.
NXP's 77 GHz multichannel radar transceiver chipset consists of four receivers, two transmitters, and an associated voltage-controlled oscillator (VCO) that serves as a high-frequency transmitter and receiver.
System Plus noted that while several chip companies, including MediaTek and Texas Instruments, claim to enter the automotive radar market, NXP and Infineon are two undisputed giants. Continental is a major player in the radar module suppliers. Its ARS4-B "can be found in at least 15 other vehicles, including audio Q3, Volkswagen Tiguan, Nissan Rogue, and more,"
Domain controllers
Tesla has developed a customized "liquid-cooled dual computing platform" that includes two parts, intelligent driving and smart cockpit, but they are not on the same PCB version, that is, they are just put in a box.
Those familiar with our article know that Tesla's original mobileye solution for intelligent driving, that is, hardware and software are someone else's, in HW2.5, Tesla integrated two Nvidia Parker SoCs, an Nvidia Pascal GPU and an Infineon TriCore CPU. Now in HW3.0, Tesla integrates two newly designed Tesla SoCs, two GPUs, two neural network processors, and a lockstep CPU.
The technology node used by the Tesla SoC in HW3 is 14nm, while Nvidia's HW2.5 processor uses 16nm. The number of processors has been reduced from four Tesla SoCs to two.
System Plus calculates that Tesla's HW2.5 consists of three Nvidia chips and Infineon MCUs, priced at $280. In comparison, the Tesla HW3.0, based on Tesla's two SoCs, costs $190. HW3.0 saves $90 a piece compared to HW2.5's main processing chip.
System Plus assumes that a car manufacturer spends $150 million on its own processor on design costs, with component prices unchanged and 400,000 units produced annually, and the car company can recoup its investment in four years. Obviously, Tesla's current sales do not take four years to recover the cost.
summary
So from the perspective of Tesla's autopilot hardware system, it is indeed a master of controlling costs, and of course he still maintains a high customer experience. So let me summarize the secret:
In general, it is necessary to have design thinking, the optimal cost accurately meets customer needs, not over-designed and over-engineered. This is easier said than done, and it is necessary to maintain a forward-looking vision and a scalable agility in technology.
Electronic components simplification and batch modularization, such as the generalization of the camera COMS image sensor sub-component module, which is actually the same principle as the modularization of the # vehicle architecture, so the modularization of electronic components is also a topic in the era of intelligent vehicles. Another example is that its camera does not process itself but is sent directly to the domain controller for unified processing.
Core key chip self-development, as a car main engine factory that relies on intelligent electronics, the core chip is not a very good thing in the hands of strong suppliers, especially when there are only one or two dominant ones. Similar to now NVIDIA may be the major OEMs in front of him bargaining power is estimated to be very weak.
Then when you think about it, our new forces are more in the pile of materials, fighting parameters, the Chinese people still have a way to go, come on.
Reference articles
Tesla Autopilot hardware disassembly - Junko yoshida EE times
*Reprinting and excerpting without permission is strictly prohibited - how to obtain reference materials: