Editor's Note: "New Variables" is a column launched by Automotive Heart that shares insights from frontline practitioners in smart cars. From the perspective of a witness, take you to foresee the key variables of the development of smart cars.
Contributing Author / Zhou Yanwu (Senior Expert)
Edit / Heart of the Car
If you know anything about Tesla's new generation of FSDs that will debut next year, how do you think about the older generation of FSDs today?
To conclude, the differences between the two generations of FSDs are enormous, especially the underlying hardware– which cannot be accomplished with a simple physical upgrade, let alone an OTA.
The current generation of Tesla FSD (HW3.0), the image sensor is ON Semiconductor's AR0136AT, which is a 2015 product with only 1.23 megapixels.
The new generation of FSD (HW4.0), if not surprisingly, will first be installed on the Cybertruck, which continues to delay mass production, and one of the biggest changes is that the image sensor has changed from ON Semiconductor's AR0136AT to Sony's IMX490, increasing the number of pixels to 5.43 million.
The IMX490 pixels are horizontal 2896* vertical 1876, 1/1.55 or 10.36 mm, ADC 10 frame rate is 40 frames, ADC 12 is 30 frames, sensitivity is 2280mV, dynamic range is 120 dB, and expandable to 140 dB.
AR0136AT is 1280 * 960 pixels, 1/3 optical size, IMX490 is twice as large, which is often said in the photography industry "the bottom of the first level of crushing people". The "bottom" here refers to the optical size of the sensor.
The optical size is large, and the benefits are twofold:
First, imaging is better. Because larger sensors can receive more light. The more light, the better the imaging. The higher the signal-to-noise ratio.
Second, larger sensors make it easier to get a wide angle. The so-called telephoto is easy to obtain, and wide angle is difficult to find. Even the AR0820AT, an 8.3-megapixel image sensor for autonomous driving introduced by ON Semiconductor at the 2022 CES show, has an optical size of only 1/2, which is still not as large as Sony's IMX490.
01
"Effective Distance" number game
The most direct performance improvement of the image sensor pixel is the improvement of the effective distance.
Of course, to some extent, the effective distance can also be understood as the manufacturer's self-talk. This is because there is no third-party agency in the world to detect "effective distance". However, there is currently a standard in the industry: that is, the effective distance is the distance for large vehicles with a distance error of no more than 5%.
To be less precise, if the assisted driving system recognizes that the effective distance of the truck (a few pixels is enough) is 100 meters, the effective distance to identify adults is about 10-20 meters, and the effective distance to identify children who are not more than 80 centimeters tall (requiring dozens of pixels or even hundreds of pixels) is about 8-10 meters.
All manufacturers are advertised according to the effective distance of identifying large vehicles, which has almost no reference significance.
The most effective distance correlation is the image sensor pixel. One of Mobileye's early papers describes this in detail (if you're interested, see the full text: https://www.cs.huji.ac.il/w~shashua/papers/IV2003.pdf), and here I'll just say the conclusion:
A 300,000-pixel camera with a height of 1.2 meters and a horizontal FOV of 47 degrees recognizes an effective distance of about 44 meters for trucks, about 180 meters for 1.3 million pixels, and only about 18 meters for children who are no more than 80 centimeters tall.
If it is a 5.43 megapixel camera, the effective distance can reach 80-90 meters. The degree of security has been greatly increased.
Sony's IMX324 image sensor, with a 7.42-megapixel and a horizontal FOV of 32 degrees, can see the speed limit signs 160 meters away.
Typically, the horizontal FOV of the main camera is 40-50 degrees, and the FOV for long-range cameras is 25-35 degrees. The wider the FOV, the closer the effective distance. The farther away the reverse, but the main factor is still the pixels of the image sensor.
02
Triple-camera to dual-shot or even single-shot
As shown in the figure above, the advantage of large optical size of the sensor is that a wider image information can be obtained.
Another benefit of the Sony IMX490 is its wide vertical range. The CURRENT AR0136AT used by Tesla is in 4:3 format, which is an old-fashioned design, considering that the display ratio of the monitor is generally 4:3 format, ar0126AT is not dedicated to smart driving, after all, it is a product of 2015.
The recent OV and ON Semiconductor's 8.3-megapixel image sensors are both 16:9 ratios, and only Sony's IMX490 is nearly 3:2 ratios.
Sony's design mainly takes into account the height of the traffic lights and strives to cover the traffic lights, while the OV and ON Designs are as horizontal as possible due to the optical size, so the 16:9 format is chosen.
Tesla is a three-camera design, the reason why it is done is because foV is different, the coverage range is different, in order to cover a larger range, you need to set up three cameras, Tesla is 35/50/120 degrees three FOVs, 120 degrees of camera is ultra-wide angle, need to do image distortion correction.
Considering the IMX490's ultra-long and ultra-wide coverage, the 35/50 degree dual camera can be combined into one, even if the IMX490 uses a 50-degree FOV, its effective distance is far more than the 30-degree FOV AR0136AT, and the width coverage is far more than the FOV50-degree AR0136AT.
Tesla has always been cost-oriented, and it is likely to change from three cameras to two cameras or even a single camera in the future, even if the IMX490 with a FOV of 50 degrees covers a width close to the AR016AT of 120 degree FOV.
So, what's so great about Sony's IMX490? For example, some temporary traffic intersections use LED signs for traffic lights, and some warning signs use LED displays.
The human eye looks the same due to the visual persistence characteristics, but the image sensor is different, if the frame rate of the image sensor and the frame rate displayed by the LED are different, there will be flickering, and the high brightness of the LED will be particularly sensitive, resulting in the camera not working properly.
The traditional practice is to filter out the LED display after it is detected. But sometimes the information displayed by the LED signal light is very important and cannot be filtered.
Sony, OV and ON Semiconductor have all developed corresponding solutions, Sony was the first to develop successfully, OV and ON Semiconductor also have this function on their 8.3 million pixels, of which Sony and OV technology route is the same, AND SEMI is the original technology. Sony does not need a back-end ISP to cooperate, and the LED sign can be recognized by sensors alone.
03
Conjecture of the second-generation FSD chip
If Tesla continues with the HW3.0 design, 5.36 megapixels will require at least 1,000 TOPS of AI hashrate, which is higher than Nvidia's Orin. Considering the cost factor, Tesla's second-generation FSD chip is likely not to do so.
The second generation of FSD or HW4.0 may only have the front main camera with 5.36 megapixels, and the remaining 5 are still 1.23 megapixels or upgraded to 2 million pixels.
Even so, it is initially estimated that Tesla's second-generation FSD chip will require at least 400 TOPS of AI hashrate.
Tesla's convolutional algorithm, AI actually has no AI at all, it is to rely on computational brute force, pixels doubled, computing power needs to increase at least 3 times.
Doing autopilot chips is like building blocks, and the main threshold is the capital threshold.
Domestic emerging car companies will soon launch their own autonomous driving chips. Tesla generously admitted in the first generation of FSD chips that in addition to NNA, which is an AI accelerator, the rest are purchased from third-party IP, such as CPU from ARM, GPU from ARM G71, on-chip interconnect from Arteris and so on.
As for the CPU of the second-generation FSD, it is very likely that the A72 will be replaced by the A76, or it may be the A78.
Above is the Tesla NNA bare crystal perspective view, one NNA has a bare crystal area of 47.54 square millimeters, two nearly 100 square millimeters, the main MAC has 9216 arrays, most of which give SRAM memory.
This is Tesla's consistent design, and the same is true of the Dojo chip on Tesla Technology Day in 2021, most of which give SRAM memory, because for AI computing, the main bottleneck is storage, and the most effective solution is to increase SRAM capacity.
However, the disadvantage is that SRAM memory occupies a very large area of bare crystals, and the area is proportional to the cost of the chip.
Second, SRAM is difficult to use advanced processes to increase density and reduce area.
Samsung's 14nm SRAM bit-cell size is 0.049 square microns, and 7 nanometers is 0.026 square microns. While the logic element Samsung's 14 nm transistor density is 32.5 million transistors/mm2, the 7 nm 7LPE process increases to 95.3 million transistors/mm2, an increase of nearly 3 times.
The hash rate of 400TOPS can be easily obtained by increasing the number of MACs and SRAM capacity, and does not require any skill. For Samsung's 7nm process, the bare crystal area needs to be increased by about 100 square millimeters.
Tesla's first-generation FSD chip is 260 square millimeters, and the second-generation FSD chip is expected to be 300 square millimeters, and the cost is estimated to increase by 40-50%.
04
Why did the second-generation FSD chip foundry choose Samsung?
The reason is simple.
First of all, TSMC has high prices and tight production capacity.
Secondly, the price of Samsung is much lower than that of TSMC and the production capacity is sufficient. At the same time, Tesla has already cooperated with Samsung, the first generation of FSD chips was produced in Samsung's Austin, Texas factory, and Tesla's headquarters moved to Austin in 2021.
The two sides have geographical advantages, just in the same city, cooperation and communication is much more convenient.
The table above shows the world's six largest foundries in the third quarter revenue, gross margin, operating margin and net tax margin. GolbalFoundries, which has been losing money for a long time, just went public in the U.S. in June 2021, and GolbalFoundries gross margin just turned positive in the first half of 2021.
Samsung's wafer foundry business belongs to the system LSI department, the financial data of the LSI department is public, but no financial data of the single wafer foundry business has been released, and wafer foundry is the main business of the system LSI department, but not all.
In addition to wafer foundry, Samsung Systems' LSI division also sells chips such as image sensors and cockpit SoCs supplied to Audi. Samsung Systems' LSI division is expected to have an operating margin of just 11% in Q3, 5% in Q2 and negative 5% in Q1.
Considering that the profit margin of the LSI division selling chips alone is much higher than that of wafer foundry, the operating margin of Samsung's wafer foundry business in the third quarter is certainly less than 10%, and it is expected to be only 8-9%, which is extremely significant compared with the three major foundries in Taiwan, and it is estimated that the revenue of Samsung's wafer foundry business in the third quarter is about 4.5 billion US dollars.
Samsung's wafer foundry business is large but not strong, and its physique is very weak.
In contrast, TSMC's operating profit margin is 4 times that of Samsung.
Samsung Systems LSI Division 2019-2021 WAfer Foundry Customer Distribution
Samsung Systems LSI's main customers, including Samsung's internal mobile phones and other chips, are on a downward trend year by year, with about 50% in 2021.
Some people will say that because they are internal customers, they lead to low profit margins, but this is actually not true. Samsung Systems' LSI division has an operating margin of 11% in 2020, and with more external customers joining in 2021, the operating margin has declined, and it is expected to be only 7%.
Obviously, internal customers contribute more profits, for external customers, Samsung is competing with TSMC at a very low price, the same is 7 nanometers, Samsung's price may be less than 1/3 or even 1/4 of TSMC.
Tesla's chip shipments are often hundreds of millions of dollars compared with the mobile phone field, which can be said to be insignificant, or even negligible, which leads to Tesla's chip price competitiveness is much lower than that of manufacturers such as Qualcomm.
Qualcomm and NVIDIA are the main customers of Samsung wafer foundry, other customers are mainly UNISOC, Ambarco and South Korean manufacturers, Baidu and Rockchip also have a small amount of chips, Tesla accounts for less than 1% of Samsung's wafer foundry business, and Spreadtrum's tens of millions of volume ratio, Tesla's volume is too low.
The above table is a comparison table of the performance of the 7-nanometer process of the three major factories, Intel is far ahead, but Intel does not do FOUNDC. In terms of performance comparison, TSMC also abused Samsung, which is almost the same as Intel.
Of course, Tesla chose Samsung, it can not choose the 5-nanometer process, because Samsung's 5-nanometer is not much higher than 7-nanometer, and the power consumption is also high. Tesla chose Samsung, can only choose 7 nanometers, the process is more mature.
In response to the U.S. government's call and proximity to major customers Nvidia and Qualcomm, Samsung officially announced in November 2021 that it would invest $17 billion to build an S2-2 fab in Texas.
The previous S2-1 plant had a maximum of only 11 nanometers, and the S2-2 plant could provide a 7-nanometer process. Production is expected to begin in 2023. Tesla's second-generation FSD program is manufactured at the S2-2 plant.
Before the S2-2 factory is officially put into production, Tesla will most likely not launch a second-generation FSD, let alone a Cybertruck. This is also to maximize the life cycle of the previous generation of FSDs.
If you are not in a hurry to buy Tesla, you can also wait for the release of the second-generation FSD. Wait for the party to finally usher in victory!