2021 is destined to be an extraordinary year for Tesla. Although it experienced the "roof rights protection" incident in April, this did not affect the soaring market value in the slightest. In October, after Hertz Global announced it would buy 100,000 Tesla electric cars (a news that was later denied), Tesla's market capitalization exceeded $1 trillion for the first time, surpassing 11 automakers including GM, Ford, Toyota, Volkswagen and BMW combined.
Not only that, in terms of annual sales, Tesla, as the "topic king", did not disappoint the "melon eating masses", not only ending with a perfect increase in annual sales, but also creating a new record of delivering 936,000 cars and a sharp increase in sales of 87%. In the fourth quarter alone, Tesla delivered more than 308,000 vehicles, up 71 percent from the same period last year, crushing previous wall Street analysts' predictions of about 266,000 vehicles in the fourth quarter and about 855,000 vehicles for the full year.
It is also a pure electric car, why is Tesla so good?
The strongest autonomous driving chip
The first thing to say is of course the most important FSD chip in Tesla's FSD system, and it can be said that chip research and development is the key to Tesla's automatic driving. Tesla's unique self-driving technology route makes it difficult for chips on the market to meet his needs.
Earlier in 2014, Tesla was still using Mobileye's new generation of assisted driving chip EyeQ3, but since the Tesla Model S crashed into a truck in 2016, Tesla and Mobileye parted ways and turned to the arms of NVIDIA's Drive PX 2. But even the Drive PX 2, which is known as a super car computer, has not become the "perfect self-driving chip" in Musk's eyes, and there are still three problems of high cost, high power consumption, and computing power that cannot fully meet the demand. In order to achieve Musk's ideal FSD, self-developed chips can be said to be an inevitable trend, so Tesla has also become the first OEM to develop chips specifically for automatic driving.
FSD chips are full self-driving chips, the design and planning of the chip began in 2016, that is to say, when it was still in a "sweet period" with NVIDIA, Tesla has begun to quietly develop its own chips.
The FSD chip project was led by the semiconductor industry legend Jim Keller and Pete Bannon, along with other architects, led the way, and it was officially released in April 2019. Pete Bannon revealed the process of chip development at the press conference, and we also know that the FSD chip development process went through 18 months, the first version of the tape-out in August 2017, the chip returned in December 2017, lit up, and finally delivered in batches on model S & X and Model 3 in March and April 2019, respectively.
As an autopilot product designed for Tesla, the demand for FSD chips mainly includes the following points, in order of power consumption, computing power, Barch size (latency), and security. In terms of security, the FSD chip has done a lot of redundant design, and the power supply and data channels of the two chips on the same board are independent and mutually backed up, so that any part of the board failure will not affect the basic car driving.
Image source: Tesla
From the architectural point of view, the FSD chip not only has the traditional CPU and GPU parts, but also adds specially customized NNP (Neural Network Processor), ISP and other parts, which together form a huge SoC chip. Among them, the CPU adopts the Cortex-A72 architecture, three groups, each group of 4 cores, a total of 12 cores, the highest operating frequency of 2.2GHz, this part of the processor core for general purposes of computing and tasks. In terms of GPUs, Tesla designed a GPU module that supports FP32 and FP64, with a main frequency of up to 1GHz and a maximum computing power of about 600 GFLPS.
A brief diagram of the internal architecture of Tesla FSD chip
The NNP is the most important part of the whole chip self-research. Each chip has two NPPs, each with a matrix of 96x96 MACs, 32MB of SRAM, operating at 2GHz. So the processing power of an NNP is 96x96x2 (OPs) x2 (GHz) = 36.864TOPS, single-chip 72TOPS, board 144TOPS.
When the FSD chip was first introduced, it used Samsung's 14nm process, with an area of 260 square millimeters and encapsulated about 6 billion transistors. However, there have been media reports before that Samsung Electronics has begun to produce chips for Tesla's advanced driver assistance system FSD, mainly the foundry contract of HW 4.0 chips, using a 7nm process.
How long does the FSD chip last? Since its launch in 2019, models later than the HW3.0 model 3 have switched computing chips, but the FSD chip will still carry Tesla's new visual neural network and the huge amount of computation in more complex scenarios. Even Tesla's new robot Tesla Bot in 2021, his "brain" is still this FSD chip released in 2019.
While the FSD chip may not be the most advanced autopilot chip, it is undoubtedly the most suitable chip for Tesla. Here, without sighing, the author said, "Jim Keller's shot must be a boutique."
Supercomputing to pure vision
At present, the global autonomous driving field is divided into two genres: the visual dominant route and the lidar-dominated route. There is no doubt that Tesla, which insists that "pure visual perception is the road to real-world AI", is the only visual party, dominated by cameras, with deep learning, multi-sensor redundancy, and 360° all-weather monitoring.
Previously, Tesla still took the route of visually dominated radar cooperation, but with the rapid change of autonomous driving technology, Tesla began to move towards a pure visual route. In July 2021, Tesla FSD Beta V9.0 was released in North America, and the biggest feature of this version is that it does not rely on radar sensors, and uses a pure visual automatic assisted driving technology route with cameras and AI intelligent algorithms.
Admittedly, the pure visual route with only the camera does have no small advantages. For example, cameras cost much less than lidar, which undoubtedly reduces the cost of self-driving cars, coinciding with Tesla's goal of becoming a "civilian car that runs all over the road." Not only that, the camera is more adaptable and will not be disturbed by fog, snow, rain and other weather. From the morphological point of view, the camera can also be easily integrated into the design of the car and hidden in the structure, so as not to affect the shape of the car.
But the camera also has its fatal weakness, that is, it can only feed the original image data back to the system, and there is no way to directly provide the precise distance and position of the object, so it often appears to sense the oolong event of the traveler in the unmanned environment. Therefore, if you want to take the pure visual route, Tesla must have a strong computer vision ability, which can accurately process images, and the Dojo supercomputer was born.
Dojo is derived from the Japanese language and means "dojo", which means training ground. As an indispensable part of Tesla's road to achieving FSD, Dojo specializes in the field of autonomous driving, collecting vehicle driving data from cameras on more than 1 million Tesla vehicles around the world, and continuously growing through deep learning and analysis of massive Tesla fleet data. In this training process, Dojo can automatically simulate driving, automatically find the optimal solution to the problem, so as to complete self-evolution, and finally achieve fully autonomous driving based on pure vision.
It is understood that Dojo is based on Tesla's custom computing chip D1 chip, and the training node (the smallest entity used in the supercomputer) is a 64-bit CPU that is fully optimized for machine learning workloads. Although this is the smallest computational element used in Dojo, it is capable of performing calculations that exceed 1 teraflop.
Block diagram of the training node
Different from the FSD chip, the D1 chip adopts TSMC's 7nm manufacturing process, the core area is 645mm2, the thermal design power consumption is only 400W, and it integrates four 64-bit superscalar CPU cores, with up to 354 training nodes, especially for 8×8 multiplication, supporting FP32, BFP64, CFP8, INT16, INT8 and other data instruction formats, with 362 teraflops processing power. In addition, high IO bandwidth is also one of the most important features of the D1 chip, according to Tesla, the IO bandwidth of D1 is about twice that of the most advanced network switching chips.
Judging from the current news, Dojo may be put into use in 2022, can Tesla create another miracle this year with the blessing of Dojo?
Beyond Bugatti's hero: silicon carbide
As the "king of goods" in the automotive industry, and also the first car company to realize silicon carbide, Tesla not only brings the global automotive industry to intelligent electrification, but also makes silicon carbide become the focus of global attention, fueling the development of the third generation of semiconductors.
For electric vehicles, power electronics are critical, and the inverter, as a key component of the powertrain system, can convert the DC energy stored in the battery into the three-phase ALTERNA required by the motor, and can also convert the AC power recovered by the motor into direct current during braking and provide it to the battery.
According to Shanghai Automotive News, silicon carbide is used in automotive inverters, and at the same power level, the package size of the all-silicon carbide module is significantly smaller than that of the silicon module, and the switching loss can be reduced by 75%. In the same package, the all-siC module has a higher current output capability, allowing the inverter to reach higher power. For on-board chargers and fast charging piles, silicon carbide semiconductors reduce energy loss during charging and also reduce the number of capacitors and inductors required compared to traditional silicon devices.
Judging from the current search data, many media believe that Tesla can achieve today's achievements, and silicon carbide is indispensable. And it is. IDTechEx data shows that after the adoption of SiC, the efficiency of tesla Model 3 inverter and permanent magnet motor combination is as high as 97%, which may be one of the highest efficiencies in the world (even one of the "ones" can be removed), and the endurance is about 6% higher than the previous solution. In addition, silicon carbide can also speed up charging, and the charging time can also be reduced by 50%. According to reports, the Model S Plaid has a full battery life of 640 kilometers, and can run 301 kilometers in 15 minutes on charging.
Under various advantages, silicon carbide has naturally become Tesla's darling. Since the use of the SIC MOSFET inverter in the Model 3 in 2018, the power module rear-wheel drive of the Model Y launched in 2020 has also adopted the SiC MOSFET. The model S Plaid launched in 2021 is faster than Bugatti, only 2.1 seconds can complete zero to 100 kilometers of acceleration, the maximum speed of up to 322km / h, according to industry analysts, silicon carbide is one of the key factors for Tesla to surpass Bugatti.
Like other new devices, silicon carbide inverters face high cost problems, but in Tesla's view, this may not be a problem again. Based on cost estimates for the 2018 Model 3 inverter and the 2020 Model Y second-generation inverter, the cost of Terrass's SiC inverter has fallen significantly in just the past three years, at nearly the same cost as the 2019 Nissan Leaf and Jaguar I-PACE silicon-based IGBT modules.
Image source: IDTechEx
Performance and cost go hand in hand, and the silicon carbide inverter is undoubtedly one of the behind-the-scenes credits of Tesla's success.
Write at the end
Just as Jobs defined what a "smartphone" is, Musk put the ideal of a "smart car" into reality, showing us what a "self-driving car" is. Whether it is a self-developed FSD chip or a pure visual route, or a silicon carbide on the car, Tesla is the first person to eat crabs. Mr. Lu Xun once said: "The first person to eat crab is admirable, not a warrior who dares to eat it?" ”
Although Tesla is full of controversy and contradictions, we cannot deny that it is his technological innovation that redefines the automotive industry, and this is the real secret behind Tesla's success, technological innovation is the soul of the rise of the brand.
Originally, the article should have ended here, but the author mentioned technological innovation! When it comes to how technological innovation can be less than the Chinese core.
Finally, the livestock team is quoted from "Tesla has no core technology, why is it still the world's first?" "Tesla appears to be an American brand on the surface, but 50% of its suppliers are in China." Buying a Tesla or not buying a Tesla is actually not important. What is important is whether domestic brands can create high-quality intelligent electric vehicles with high quality and low prices in the conventional competition of marketization, so that the Chinese people can enjoy the quality of life brought about by the electric vehicle technology revolution. It doesn't make sense if everyone boycotts foreign brands. It will only degrade domestic brands, and without sufficient competition, there will be no highly developed business. Historical experience has repeatedly told us that the rise of domestic brands is inseparable from the innovation of local industrial technology. Talking about patriotic marketing away from technological innovation is a mistake for the country and the people! ”
*Disclaimer: This article is original by the author. The content of the article is the author's personal opinion, semiconductor industry observation reprint is only to convey a different point of view, does not mean that semiconductor industry observation endorses or supports the view, if there is any objection, welcome to contact semiconductor industry observation.
Today is the 2928th content shared by Semiconductor Industry Watch for you, welcome to follow.
★ Material analysis of power semiconductors
★ A new round of EUV lithography machine scramble begins
★ Siege SiC substrate faucet
Wafer | integrated circuit | devices | automotive chips | storage | TSMC | AI | packages