Horizon Robotics Vice President Yu Yinan: R&D and mass production of in-vehicle AI chips are difficult, and open source technology will help reduce industry costs

With the development of the "New Four Modernizations" in the automotive industry, the demand for computing power has doubled, and in-vehicle AI chips have increasingly become one of the core components of smart cars. Along with this, the automotive chip industry has experienced a blowout development, giving rise to a series of high-computing computing platforms represented by Huawei MDC, NVIDIA's Orin, Horizon Journey 5, Qualcomm Snapdragon Ride, etc.

Foreign brand chips have always occupied the main market share. However, with the rapid development of China's smart car industry and the impact of the "chip shortage" dilemma, self-developed in-vehicle smart chips have gradually become a must-win for the development of the domestic smart car industry. And with the progress of domestic chips, domestic automotive-grade AI chips have gradually been recognized by the market, such as Horizon Journey 2, Journey 3 and the latest high-performance, high-computing chip Journey 5.

Yu Yinan, vice president of Horizon Robotics and general manager of intelligent driving products, believes that smart car chips are like the "engine" of a car and the "Mount Everest" of contemporary hard technology. In-vehicle chips will surpass mobile phone chips to become the leader in the semiconductor field. The automotive electronic and electrical architecture will evolve from the traditional distributed architecture to the domain controller architecture and the central computing platform architecture. The AI ​​computing of smart cars will also evolve from the logic-driven software 1.0 era to the data-driven software 2.0 era, and the demand for computing power will increase significantly.

Founded in 2015, Horizon Robotics is the world's first AI chip startup. In recent years, Horizon Robotics has basically maintained the launch of a new generation of high-performance AI chips every year. Its Journey 2 chip released in 2019 has been used in many mainstream models; the new generation of Journey 3 chips in 2020 were also first mass-produced in Ideal, a leading new car-making force in China. So far, Horizon Robotics' Journey chips have been designated by more than 20 car companies and more than 50 car models, and 1 million chips were shipped last year.

The first domestically produced high-computing AI chip

With the development of automotive intelligence, the industry's requirements for the computing power and performance of automotive chips are rapidly increasing.

In this context, domestic chip manufacturers represented by Horizon Robotics have been accelerating the iteration, upgrading and optimization of their chip products in recent years. From the early Journey 2 and Journey 3 chips to the new generation Journey 5 chip launched in July 2021, Horizon Robotics' automotive chip performance has made great progress.

(Yu Kai, founder and chief scientist of Horizon Robotics)

Yu Kai, founder and chief scientist of Horizon Robotics, said that the Journey 5 chip has both high computing power and high performance. A single chip has an AI computing power of 128 TOPS, supports 16-channel camera perception computing, and can support the multi-sensor fusion, prediction, planning and control required for autonomous driving. This is also Horizon Robotics' third-generation automotive-grade product after Journey 2 and Journey 3.

Moreover, based on Horizon's R&D concept of combining software and hardware and its self-developed high-performance computing architecture BPU, Journey 5 has obvious advantages over NVIDIA's automotive chips of the same level in terms of real-time latency, processing speed and object detection. Moreover, the power consumption of Journey 5 is as low as 30W, while the official power consumption of NVIDIA's ORIN series chips is 80W. At the same time, as a representative of domestic chips, Horizon's Journey 5 chip's AI performance and energy efficiency on some typical neural networks exceed those of benchmark chips such as NVIDIA Orin-X. 

Horizon believes that the pursuit of pure computing power breakthroughs is not sustainable due to the power consumption limit of Moore's Law. At the same time, computing power does not represent the "real performance" of automotive smart chips, and chip computing efficiency also needs attention. In 2016, Horizon proposed the "New Moore's Law of AI Computing", believing that the real AI performance of a chip is determined by the combination of physical computing power, effective utilization of chip architecture, and AI algorithm efficiency. Only based on such comprehensive optimization can the performance be perceived by users.

Horizon's concept combined with the high computing power of the Journey 5 greatly improves the overall performance of the Journey 5. At present, the Journey 5 has won the mass production orders of many car companies and will become the first domestically produced AI chip with high computing power to achieve pre-installed mass production.

The mass production of automotive-grade chips is quite complex

Leading chip companies represented by Horizon Robotics have made remarkable progress, but the "chip shortage" in the automotive industry has continued since 2021. The research and development, mass production, and installation of in-vehicle AI chips are still common challenges faced by domestic chip companies.

In Yu Yinan's view, the complexity and difficulty of the series of processes from chip project establishment, including market analysis, example requirements and scenario requirements, to chip development, are very high, and it can be said that it is a long way to go. In the development process over the past few years, the biggest problem Horizon encountered was how to make the BPU computing architecture run more efficiently in the entire SOC chip architecture, including various computing IPs, the interdependence between computing units, and the flexible transfer of data. In addition to basic BPU research and development, these aspects require huge resources and manpower, and require top SOC architects to achieve.

In addition, after completing the front-end chip design, the next step is back-end simulation and chip wiring, which require the use of many tools. The next step of chip tape-out is a test of the early design skills.

Moreover, there is still a lot of important work to be done after the chip is successfully taped out, including functional safety development, information security development, a large number of chip-level tests, and chip quality optimization. When a chip is just taped out, there is still a lot of room for optimization in the chip production line process and manufacturing process, so continuous improvement is needed to reduce chip PPM and optimize the process.

Later, software personnel began to come on stage, and put algorithms, software, architecture, and many partners, including various drivers, all on board, and finally came to the vehicle manufacturer to start closed-loop development. "So it looks like a chip, an autonomous driving, and a software package, but the effort invested in the early and later stages is beyond imagination." Dr. Yu Yinan said: "The three generations of products we have made, each generation has invested much more than the previous generation. Although we have tried to reuse existing things, after continuous iterative optimization, we always find that there are still many things that need to be redone to achieve the upgrade of the entire architecture level."

Open source and sharing of technology can help the industry reduce costs

Chip R&D is complex and meticulous, and the cost is not affordable for all companies. Therefore, there is a gradual call for technology sharing and chip IP open source in the industry, in order to promote industry development through mutual cooperation within the industry. Horizon Robotics is actively practicing this industry trend.

At the 2022 China Electric Vehicle 100 Forum, Horizon proposed that based on the open "chip + algorithm + tool chain + OS" platform, it will try to further open BPU IP licensing to some vehicle manufacturers to help automakers improve their differentiated competitiveness.

Yu Yinan believes that openness will always beat closed systems. Moreover, fields such as autonomous driving are very different from the ecosystem of mobile phone apps. In the field of autonomous driving, the goals of industry players are the same, including from chip architecture to chips, to underlying perception, positioning systems, and regulatory control systems. In this case, the industry has spent a lot of resources and capabilities, and different companies have done the same thing. This kind of involution and internal consumption is meaningless.

Yu Yinan said that the openness of technology (including software and algorithm) and even chip IP will help optimize the cost of the entire industry. Horizon Robotics has always been committed to making the best and most optimal system and maintaining an open attitude, hoping to help the industry reduce costs and promote the common prosperity of the industry ecosystem.