Full set of video processing and image recognition IP for advanced sensing and recognition.
R-Car V4H, the central processor for ADAS and AD solutions released by Renesas in 2022, adopts 7-nanometer technology, with a deep learning performance of 34TOPS, and can recognize and process the surrounding objects at high speed through cameras, radars and lidar. The device is scheduled to start mass production in the second quarter of 2024. With the careful combination of first-class IP and professional hardware optimization, R-Car V4H has achieved an industry-leading performance-to-power ratio, which is suitable for mass production of L2+ and L3 autopilot applications.
Thanks to its high degree of integration, R-Car V4H helps customers to develop a cost-competitive single-chip ADAS electronic control unit (ECU) to support L2+ and L3 driving systems, including complete NCAP 2025 functions. R-Car V4H also supports the function of looking around and automatic parking, and has excellent 3D visual rendering effect.
Concluding remarks
The development of automobile intelligence is expected to bring trillions of market opportunities. From the current situation, L3 and above autopilot is accelerating penetration, and the demand for high-power autopilot chips is also increasing. In addition to the major enterprises mentioned above, many domestic enterprises such as CAMBRIAN, Zero Run, and Naineng have actively invested in the field of chips.
However, the research and development of autonomous driving chips is a process with huge initial investment and long payback period. Therefore, if there is no loading capacity of 1 million vehicles, it is impossible to make a profit by developing an autopilot chip with the power of a main engine factory and using it by itself.
On the whole, the smart driving chip market is still in a rapid iteration period, and China manufacturers are expected to break through, but they also have to face many uncertainties. Only by firmly grasping the certainty and market rhythm, accelerating the iteration of technology, products and tool chains, and taking the lead in forming a closed-loop business, can we further compete with the global automotive chip giants after squeezing into the table.
There are two identical NPUs in each FSD chip, which have strong computing power and are physically integrated adjacent to each other. In each calculation cycle, NPU will read 256 bytes of activation data from the built-in 32MB SRAM and combine them with another 128 bytes of weight data to enter the multiply-accumulate (MAC), and each NPU has a 96×96MAC array. After completing the MAC multiplication and accumulation operation, the data will be transferred to the Activations and Pooling parts, waiting to write the summary results in the buffer.
Tesla has also made great efforts to optimize the power consumption and cost of NNA (Neural Network Acceleration Unit) in order to make the data flow in the chip as much as possible without frequently reading and writing with memory or other modules. In the design of NNA, software is used to simplify the hardware to reduce the chip cost.
Tesla chips cover everything from autonomous driving, intelligent cockpit, domain control architecture, three electric systems, and even vehicle manufacturing and assembly. From the point of view of autonomous driving, Tesla is the only company that has realized the full-stack self-research of software and hardware at present, taking a technical route that is difficult to replicate. Tesla’s FSD chips are basically for personal use and are not for sale.
Qualcomm
Qualcomm is a mobile phone chip giant, and its entry point in the automotive field is communication and cockpit chips, with its 820A, 8155 and 8295 occupying a major share in the mid-to-high-end smart cockpit market. Previous Snapdragon Ride autopilot computing platforms, including the 5-nanometer Snapdragon 8540 and the 7-nanometer Snapdragon 9000, with a single-chip computing power of 360TOPS, have been put on the bus one after another.
At the beginning of 2023, Qualcomm introduced Snapdragon Ride Flex SoC supporting central computing architecture, which supports digital cockpit, ADAS and autopilot functions at the same time with a single SoC, providing more than 700TOPS of computing power. The first Ride Flex SoC has been sampled and is expected to be mass-produced in 2024.
At present, Qualcomm has reached cooperation with Great Wall, General Motors, BMW, Volkswagen and other manufacturers in the field of autonomous driving. In the future, the Snapdragon Ride platform will be applied to some models of the above-mentioned car factories.
In April this year, Qualcomm also completed the acquisition of Weininger, which supplemented the algorithm capability of autopilot software. Its application fields include radar, perception system, autopilot algorithm, functional safety/expected functional safety, etc.
NVIDIA
NVIDIA is a representative of the big computing chip, and has entered the field of smart cars with the evolution of GPU architecture. In NVIDIA’s NVIDIADrive series, DRIVECX faces the cockpit, and DRIVEPX faces autonomous driving; Since then, the automatic driving platforms of DRIVEPX, DRIVEPX2, DrivePXXavier, DRIVEPXPegasus and DRIVEAGX Orin have been introduced iteratively.
DRIVE Orin is a system on a chip released by NVIDIA, and it is the central computer of intelligent vehicles, which can provide 254TOPS per second, meeting the safety standards such as ISO 26262 ASIL-D. In March 2022, NVIDIA’s DRIVEAGX Orin platform began mass production and sales. The platform consists of two Orin SoC and two Ampere-based GPUs, with a maximum computing power of 2000TOPS and a power consumption of 800W W.
Orin SoC adopts 7 nm technology, which is composed of Ampere-based GPU, ARMHerculesCPU, second-generation deep learning accelerator DLA, second-generation vision accelerator PVA, video codec, and ISP with wide dynamic range. At the same time, it introduces the design of SafetyIsland. The hardware of the chip can be connected with 14 cameras, 1 laser radar, 5 millimeter-wave radars and 12 ultrasonic radars, and the sensor data is calculated 254 trillion times per second by Orin to realize high-level automatic driving function.
NVIDIA’s automobile business has formed a strong open ecology including Tier1, sensors and software manufacturers. At present, NVIDIA has dominated the high-end passenger car market of over 300,000. In the next six years, NVIDIA’s orders in hand will exceed $11 billion. According to incomplete statistics, more than 25 OEMs around the world have reached Orin cooperation with NVIDIA, including Weilai, Tucki, Ideality, Weimar, SAIC Zhiji, R Auto, Faraday Future, Lucid Group, Human Horizons Vinfast, Chinese Express and other new car-making forces; And BYD, Mercedes-Benz, Jaguar Land Rover, Volvo, Hyundai, Audi, Lotus and other traditional OEMs. In addition, Desai Siwei, Baidu Apollo, General Cruise and Google Waymo will also develop ADAS solutions based on Orin.
Mobileye
Mobileye is the earliest self-driving overlord. The "visual algorithm+chip" solution occupies about 70% market share in the L0-L2 market and was acquired by Intel in 2017.
EyeQ series chips are the core of Mobileye. Three years after mass production of EyeQ4, Mobileye launched the 7 nm FinFET process EyeQ5 in 2021. It includes multi-threaded 8-core CPU and 18-core vision processor, and its computing power is increased to 24TOPS, and its power consumption is only 10W, which can meet the driving demand of L4.
In 2022, Mobileye laid out high-level autopilot, and issued three chips in succession, including EyeQUltra for L4 autopilot, and EyeQ6L and EyeQ6H for L2 autopilot.
Compared with the previous products, the performance of the three chips has been greatly improved under the premise of controlling energy consumption. In particular, EyeQUltra, which is specially designed for autonomous driving, has 64 core processors and 12 RISC-VCPU, each with 24 threads, which can achieve high-efficiency and accurate calculation with less than 100W power consumption. It does not need the computing power and cost of integrating multiple systems, and solves the major challenges faced by the industry. EyeQUltra is expected to be fully put into production in 2025. This chip is benchmarked against Xavier of NVIDIA, but its computing power seems to be slightly lower than that of NVIDIA.
In the early days, mainstream OEMs chose EyeQ3 and EyeQ4 series in succession, which enabled Mobileye to successfully occupy the pre-assembly market. However, its black box model has shown signs of acclimatization and the market growth rate has slowed down. Previously, Mobileye EyeQ6, EyeQ6L and EyeQ6H had reached in-depth cooperation with Volkswagen Group and Ford Motor Company, and also reached an agreement with Geely. EyeQUltra is expected to provide samples in 2023 and mass-produce cars in 2024.
Heizhima intelligent
Black Sesame Intelligence is the advocator of single-chip SoC scheme. The A1000L and A1000 single-chip SoC schemes support the integration of navigation and parking without time-sharing and avoid increasing the security risks of the whole scheme.
Since 2019, the series of chips launched by Black Sesame include Huashan No.1 A500, Huashan No.2 A1000, A1000L, A1000Pro and A2000. Among them, A500 focusing on autonomous driving has a computing power of 5-10TOPS, Huashan No.2 A1000 with 16 nm technology has a computing power of 40-70TOPS, the low version A1000L has a computing power of 16TOPS, and the high version A1000Pro has a single chip computing power of 196TOPS, which supports high-level autonomous driving and can realize seamless connection from parking, urban interior to high-speed scenes. A2000 is the first 250T large computing chip in China, which adopts the core IP of domestic independent intellectual property rights in 7 nm process and meets the ASIL B level security certification standard.
In April, 2023, Black Sesame Intelligent released Wudang series cross-domain computing platform for smart cars and its first chip C1200, covering the needs of different domains of smart cars such as cockpit and smart car with higher cost performance, and realizing multi-domain integration capability. Based on 7 nm process, C1200 provides powerful general-purpose computing and general-purpose rendering computing power by using A78AE and G78AE, which support lock step. The self-developed DynamAI NN low-power neural network acceleration engine supports NOA scenarios; Built-in mature and high-performance Audio DSP module and a new generation of self-developed NeuralIQ ISP module that processes 1.5G pixels per second online. It can provide the industry’s highest MCU computing power of 32KDIPMS, handle the input of more than 12 HD cameras at the same time, and support high-speed MIPI.
Up to now, Black Sesame Intelligence has cooperated with OEMs such as Jiangqi, Dongfeng, Geely, FAW, SAIC, Hechuang Automobile and SAIC-GM-Wuling, as well as enterprises such as Bosch, Zhongke Chuangda, Soto Ruian and Baolong Group for L2 and L3 ADAS and autonomous driving perception system solutions, including Baidu Apollo’s partners.
Xinchi technology
Xinchi Technology is the world’s first "full-scene, platform-based" chip product and technical solution provider. Its products cover intelligent cockpit, intelligent driving, central gateway and MCU, covering the core chip category of the future automotive electronic and electrical architecture, aiming to achieve "four cores in one".
From 2019 to 2020, Xinchi Technology released V9L/F and V9T autopilot chips, which support ADAS and domain controller respectively. In 2022, Xinchi Technology released an autopilot chip V9P/U with a computing power of 10-200T, which has higher computing power and can support L3-level autopilot. V9 series processor is a car code chip specially designed for a new generation of intelligent driver assistance system, which integrates high-performance CPU, GPU and CV engine, and can meet the increasing demand for computing power.
In 2023, Xinchi Technology will launch a V9S autopilot chip with higher computing power, which is developed for the central computing platform architecture, with a computing power of 500-1000T, and can support L4/L5 autopilot Robotaxi.
At the 2023 Shanghai Auto Show, Xinchi Technology released a new generation of vehicle regulation processor V9P, which is specially designed for the ADAS domain controller with integrated traveling and parking. The CPU performance is 70KDMIPS, the GPU is 200GFLOPS, and the overall AI performance is as high as 20TOPS. All functions of mainstream L2+ADAS such as AEB, ACC and LKA, auxiliary parking and memory parking functions can be realized on a single chip, and the driving recorder and high-definition 360-degree viewing can be integrated.
horizon
As a large-scale pre-assembly mass production enterprise of domestic car-gauge AI chips, Horizon has successively launched China’s first car-gauge AI chip Journey 2, a new generation of AIoT intelligent application acceleration engine Sunrise 2, a high-performance car-gauge AI chip Journey 3 and a new generation of AIoT edge AI chip platform Sunrise 3.
Horizon’s third-generation car-level product Journey 5 has a single-chip computing power of up to 128TOPS, which has both large computing power and high performance, and supports the needs of 16-channel camera perception computing and multi-sensor fusion, prediction and planning control required for autonomous driving. Its core IP is 100% self-developed BPU Bayesian architecture, which can support the latest convolutional neural network and Transformer architecture and meet more than 170 different operators required for various autonomous driving scenarios. Using deep learning to accelerate ultra-large-scale parallel computing, reduce the whole computing delay and improve efficiency, and its image rate has increased from 1283 frames per second in 2021 to 1531 frames per second in 2022. With the launch of Journey 5, the horizon covers the whole vehicle smart chip scheme from L2 to L4.
At present, Horizon has reached cooperation with many OEMs and Tier 1, including Changan, FAW, SAIC, GAC, Chery, Dongfeng Motor, Zhiji, Jianghuai, Ideality, BYD, Great Wall, etc., as well as mainland China, Asia Pacific, Sagitar (parameter | inquiry) Juchuang, Winch Technology, Neusoft Ruichi, ADAYO Huayang, Desai Xiwei, etc.
Huawei
As a giant in the ICT field, Huawei’s smart car business can be traced back to 2012. In 2018, it launched the smart driving computing platform, and launched the MDC600 supporting L3 and the MDC300 supporting L4. In 2019, Huawei formally established the Smart Car Solutions Division, and in 2020, it released the MDC610 computing platform, using the Ascent 610 chip, with a single chip computing power of 200TOPS.
Huawei Shengteng 910 is an AI processor with super computing power, and its maximum power consumption is 310W W. The Leonardo da Vinci architecture developed by Huawei greatly improves its energy efficiency ratio. Its 8-bit integer precision (INT8) performance reaches 640TOPS, and its 16-bit floating-point number (FP16) performance reaches 320 TFLOPS.
The Ascension series includes training and reasoning chips. The Ascension 910 used for training has a semi-precision (FP16) computing power of 256TFLOPS, which is twice that of the industry. For reasoning, the computing power of Ascension 310 and INT8 is 16TOPS, and the power consumption is only 8W.
Aiming at the demand of automatic driving for computing platform, the MDC solution introduced by Huawei integrates the self-developed Host CPU chip, AI chip, ISP chip and SSD control chip. So far, Huawei has successively launched four series products with different computing power-MDC300, MDC210, MDC610 and MDC810, covering the L2-L5 autonomous driving ecological chain.
At present, auto-driving R&D companies such as Heduo Technology, Xidi Intelligent Driving, Neolithic, Yuanrong Qixing, Tage Zhixing, Huituo, Nezha Automobile, etc. all adopt Huawei’s MDC computing platform.
Dezhou instrument
Texas Instruments (TI) is a traditional automobile chip manufacturer with rich experience in product landing and strong supply chain management ability, and its customers are all over the automobile industry. In terms of processors, TI has launched the seventh generation, and each generation uses its own IP and patents. Twenty years’ experience in digital processors has also continuously optimized this heterogeneous processor.
TI’s mainstream intelligent driving products include TDA4 VL, TDA4 VM and TDA4 VH, with computing power of 4TOPS, 8TOPS and 32TOPS respectively. Among them, the TDA4 VM of 8TOPS has been mass-produced, which is the mainstream choice for the light-weight traveling and parking integrated domain control track.
TDA4VM series processors for ADAS and self-driving cars are based on Evolutionary Jacinto 7 architecture, which combines high-performance computing, deep learning engine and special accelerator for signal and image processing in a unique way in a target architecture with safe and compatible functions, so it is very suitable for robotics, machine vision, radar and other applications.
These devices are highly integrated, which makes the advanced automobile platform supporting centralized ECU or multiple sensor modes scalable and low-cost. The key cores include the next generation DSP with scalar and vector cores, dedicated deep learning and traditional algorithm accelerators. These devices also include the latest Arm and GPU processors for general computing, integrated next generation imaging subsystem (ISP), video codec, Ethernet hub and isolated MCU island. All these are protected by automotive safety hardware accelerators.
Renesas
Renesas is a traditional automobile chip giant, and its automobile products are mainly MCU and SoC Renesas. The early product is R-CAR V3H, which is heavily used in the intelligent forward-looking integrated machine with the help of Tier1 such as Bosch and Denso. R-CAR V3H SoC supports the safety index of ASIL C level in real-time domain, which can reduce the need to use external safety microcontroller to manage sensor fusion and final decision execution. R-Car V3H has rich and verified IP, supports perception stack, ISP with radar and/or lidar sensor fusion and up to 8 megapixel cameras, and also has a full set of video processing and image recognition IP for advanced sensing and recognition.
R-Car V4H, the central processor for ADAS and AD solutions released by Renesas in 2022, adopts 7-nanometer technology, with a deep learning performance of 34TOPS, and can recognize and process the surrounding objects at high speed through cameras, radars and lidar. The device is scheduled to start mass production in the second quarter of 2024. With the careful combination of first-class IP and professional hardware optimization, R-Car V4H has achieved an industry-leading performance-to-power ratio, which is suitable for mass production of L2+ and L3 autopilot applications.
Thanks to its high degree of integration, R-Car V4H helps customers to develop a cost-competitive single-chip ADAS electronic control unit (ECU) to support L2+ and L3 driving systems, including complete NCAP 2025 functions. R-Car V4H also supports the function of looking around and automatic parking, and has excellent 3D visual rendering effect.
Concluding remarks
The development of automobile intelligence is expected to bring trillions of market opportunities. From the current situation, L3 and above autopilot is accelerating penetration, and the demand for high-power autopilot chips is also increasing. In addition to the major enterprises mentioned above, many domestic enterprises such as CAMBRIAN, Zero Run, and Naineng have actively invested in the field of chips.
However, the research and development of autonomous driving chips is a process with huge initial investment and long payback period. Therefore, if there is no loading capacity of 1 million vehicles, it is impossible to make a profit by developing an autopilot chip with the power of a main engine factory and using it by itself.
On the whole, the smart driving chip market is still in a rapid iteration period, and China manufacturers are expected to break through, but they also have to face many uncertainties. Only by firmly grasping the certainty and market rhythm, accelerating the iteration of technology, products and tool chains, and taking the lead in forming a closed-loop business, can we further compete with the global automotive chip giants after squeezing into the table.
