The automotive chip track is undergoing the change of "three years of Hedong and three years of Hexi".
As the growth rate of L2-level assisted driving accelerates, L2+ and even L3 become the next track for upstream chip manufacturers to compete. At the same time, the innovation of the vehicle electronic architecture and the vehicle OTA also drive car companies to consider more computing power redundancy in the chip selection of the next generation of intelligent models.
The iterative development of high-performance computing platform + software has also become a new brand label for automobile manufacturers. Chip manufacturers such as NVIDIA, Qualcomm, Xinchi Technology, and Black Sesame Intelligence are all competing for the market share of high-end intelligent driving.
In contrast, in the assisted driving track, mobileye, Renesas, TI, Xilinx and other traditional chip manufacturers are also making upward efforts. At the same time, chip manufacturers are also seeking competitiveness "plus points" on the open software platform, with the goal of helping car companies accelerate the development speed of new systems, reduce unnecessary development costs, and fully tap the "capabilities" of chips.
Such a change was not obvious a few years ago. The reason is that the transformation from the traditional black box to the white box development model, the control of car companies is gradually increasing, and the white-hot competition in the chip market is intensifying the industry change.
One
As the chip leader of the L2 level and below auxiliary driving track, Mobileye "sacrificed" a new product line at the beginning of this year, in addition to continuing to consolidate the traditional market, another goal is high-end intelligent driving.
On January 4, Mobileye unveiled its most advanced and powerful EyeQ Ultra system integration chip for autonomous driving to date at CES. Optimized for energy efficiency at 176TOPS, the EyeQ Ultra is a streamlined chip for autonomous vehicles.
176TOPS is about the sum of the computing power of 10 EyeQ5H chips, although from the data indicators, there is still a big gap from Nvidia's Orin, but Mobileye, which has not emphasized computing power, has to make changes "for the market and customers".
Also introduced with The EyeQ Ultra are two new EyeQ system integration chips for ADAS solutions, the EyeQ 6L and EyeQ 6H, based on a 7nm process and featuring cost-effective pure camera solutions.
Among them, the EyeQ 6H has about twice the hash rate of EyeQ5H, and the goal is to target the cost-effective L2+/L4 (including L3) market. The company expects the first generation of engineering samples to be available by the end of this year.
For the current automotive industry's competition for chip computing power, Mobileye said for the first time that TOPS is a very insufficient indicator of computing power. "The computational model we integrated into the EyeQ chip is very complex and far from being energy-efficient by a single metric."
Mobileye believes that they can run the entire SuperVision system on two EyeQ5 chips (in the Chinese market, first installed on The Krypton 001), which is much lower in magnitude than the hash rate or TOPS indicator of other competitors' PKs.
This is due to its innate advantages in the co-design of chip software and hardware, but as a Tier1 partner of Mobileye, whether it can give full play to this ability obviously depends on the final strength performance of different manufacturers.
Mobileye believes that at L2 and below, front-looking ADAS cameras that provide basic safety warning and control functions can reduce verification costs, improve energy efficiency, and reduce cooling costs through the close combination of software and SoCs. In other words, black box mode is more suitable.
In the company's view, the underlying ADAS is a very price-sensitive product, which means that their customers do not need to program on EyeQ chips, such as The EyeQ4 chip, in a sense to alleviate the additional cost of customers.
When it comes to multi-camera, multi-sensor systems with centralized computing platforms, starting with EyeQ5, the goal is to become a programmable platform. At present, Mobileye has released the SDK and worked with partners to develop software algorithm packages.
But the current competitive situation has also forced Mobileye to speed up. This is also evident in the speed at which the product was launched, from EyeQ4H to EyeQ5, with a gap of three years. Since last year, almost every year there will be new products landing.
Two
In the next few years, L2+ and L3 level intelligent driving technology will become an important part of the intelligent automotive industry, which is not only a technical issue, but also a matter of brand competitiveness. As a chip supplier of Bosch's third-generation smart cameras (equipped with V3H), Renesas is also intensively introducing new products to meet market demand.
R-Car V4H is the latest generation of SoC introduced by the company to meet the needs of L3-level high-end intelligent driving, and the deep learning computing power of 34TOPS is about 8 times that of V3H, which is used to process a large amount of multi-sensor data while visualizing road scenes.
V4H is based on 7nm technology (V3H is 16nm) and adds hardware accelerators for specific tasks to meet the computing power requirements of deep learning, supporting up to 16 cameras, achieving 360° surround view, and supporting millimeter wave radar and lidar and other sensors, and implementing ASIL B and D indicators in some real-time domains.
Among them, the SoC contains 4 ARM Cortex-A76 cores of 49K DMIPS, 3 lockstep mechanism ARM Cortex-R52 cores for safety-critical tasks, an image signal processor with parallel processing of machine and human vision, an image renderer and a graphics processing unit (GPU).
In terms of data interfaces, in addition to typical interfaces such as CAN, Ethernet AVB, TSN, FlexRay, etc., there are two 4G PCIe interfaces to achieve data communication with externals. At the same time, Renesas' dedicated power supply solutions can be configured.
In addition to the hardware part, the support of the software development platform is also the beginning of last year, and several chip giants want to help downstream customers (Tier 1 and OEMs) minimize the workload of hardware and software development, while reducing design complexity, cost and time to market.
The above SDK provides the full functionality of machine learning development and optimizes the performance, power efficiency, and functional safety of embedded systems. Renesas also provides a complete simulation model that allows engineers to obtain fast CNN benchmark results through cloud-based development patterns.
From the data indicators, R-Car V4H and Mobileye's EyeQ 6H are at the same level of hash rate. At present, the V3H for Bosch's third-generation camera and mobileye's basic version of the EyeQ5M (for front-view all-in-one) are also at a similar level.
Three
Judging from the current market situation, traditional manufacturers such as Mobileye and Renesas hope to meet the next-generation system needs of car companies through "upward" upgrades, while companies such as Qualcomm, NVIDIA, and Black Sesame Intelligence are holding high and directly using large computing power platforms to seize the high-end intelligent driving track.
Nvidia Orin is the first large-scale computing platform to receive large-scale mass production orders, including ideal, Weilai, Xiaopeng, Zhiji, Gaohe, Jidu, Mercedes-Benz, Volvo, Jaguar Land Rover and other brands, starting from this year, it will be mass-produced on the car.
At present, the Orin series finally launched soC has 2 versions, including 110 TOPS Orin and 254 TOPS OrinX, based on multiple Orin or OrinX combinations, domain controller computing power can reach more than 1000 TOPS.
In addition, Companies such as Qualcomm and Black Sesame Intelligence have also successively obtained the pre-installation mass production point. For example, Black Sesame Intelligence has officially announced that it has obtained a number of self-owned brand models, and the company's products have been tested and verified for more than 2 years, and are expected to achieve mass production this year.
Software-defined centralized computing has also pushed automakers to take "computing power redundancy" as a key consideration indicator in chip selection. "Through the vehicle OTA+ large computing power platform, automakers can develop more customized functions and experiences for users." Koloda CEO Wu Baiyi said.
However, chip manufacturers want to do more.
NVIDIA made an analogy, "The reason why our data center business is successful is because in addition to the hardware itself, we also provide development platforms and resources such as software and SDKs to help customers quickly land." In the process, the company found that it was entirely possible to sell software separately as a separate business, which was seen as a new stage in the chip manufacturer's business model.
Previously, NVIDIA's cooperation with Mercedes-Benz and Jaguar Land Rover involved the joint development of software. "We have the ability to share software with automakers and add value to commercialization throughout the lifecycle."
Nvidia DRIVE' open source software stack, for example, helps developers efficiently build and deploy a variety of applications, including sensing, location and mapping, planning and control, driver monitoring, and natural language processing.
In addition, NVIDIA offers a corresponding modular configuration, including NvMedia for sensor input processing, NVIDIA CUDA library for efficient parallel computing, NVIDIA TensorRT for real-time AI inference, and other developer tools and modules with access to hardware engines.
"We need to further simplify our customers' hardware design efforts, reduce functional safety verification requirements, and reduce the power consumption of the application." Naoki Yoshida, vice president of Renesas Electronics' Digital Product Marketing Division, said.
Last year, Renesas also announced the launch of the R-Car Software Development Kit (SDK), a complete software platform with a single software package that enables faster and easier software development and validation of smart cameras and autonomous driving applications used in passenger cars, commercial vehicles, and off-road vehicles.
Just recently, Renesas Electronics announced a partnership with AVL, a supplier of development, simulation, and test solutions for the automotive industry, to provide customer support for the development of electronic control units, including a growing number of domain controllers, that comply with the ISO 26262 automotive functional safety standard.
Qualcomm is not idle.
The company has been fleshing out its software ecosystem over the past few years, for example, by supporting Valeo's latest-generation Park4U (cross-level parking solution), partnering with Seeing Machines to provide embedded DMS solutions, and expanding full-stack solutions by acquiring Veoneer's software assets.
Qualcomm said the Snapdragon Ride platform will not only include hardware, but will also provide "secure middleware, operating systems, and drivers" for the chip. In addition, Qualcomm will provide positioning, perception and behavior prediction software, which are three key components of any autonomous driving system.
"For some car manufacturers, if the software development capabilities are weak and the development costs of third-party traditional Tier1 suppliers are too high, we can directly provide a complete ADAS turnkey solution." A Qualcomm spokesperson said.
Nvidia announced that it will launch the Drive Hyperion 8 complete hardware and software stack system in 2024, and the hardware partners include Continental (long-range radar), HELLA (short-range radar), Luminar (lidar), Sony (camera) and Valeo (sensor).
The mass-produced platform is also open and modular, allowing customers to combine multiple modules from computing hardware and middleware to autonomous driving, human-computer interaction and smart cockpits. The debut model is expected to be the next generation of Mercedes-Benz.
Among them, the sdk includes the DriveWorks sensor abstraction layer, Drive AV software (including deep neural networks for perception, mapping, planning and control), Drive Concierge (human-computer interaction platform) and Drive Chauffeur (for taking over driving tasks), and is open to car companies for custom development.
In the view of the Gaogong Intelligent Automobile Research Institute, in the next few years, the competition in the automotive chip track will become increasingly fierce. On the one hand, under the same performance conditions, the price will become the focus of the hardware battle. On the other hand, chip manufacturers are testing the boundaries of their own software capabilities, especially after the flattening of the supply chain, and the communication with car companies will be closer.