Author|Tu Ming
Edit|Chestnuts
(ChinaIT.com News) Since the advent of ChatGPT3, the world’s expectations for AI have increased unprecedentedly, and a wave of using large models to empower thousands of industries has surged.
The smart driving industry also has a “GPT” craze: first, Momo Zhixing released DriveGPT, a large generative model for smart driving, and then, Banma Zhixing released the Banma Co-Pilot automotive AI capability system based on the Alibaba Tongyi Qianwen large model. On the Ideal Family Technology Day just a month ago, Ideal Auto also announced that it will no longer rely on high-precision maps to develop urban NOA products, but will rely on the self-developed cognitive model MindGPT to improve the car’s intelligent driving effect.
In the era of large models, the intelligent driving algorithm of automobiles has been further upgraded, which puts forward higher requirements for the computing support capabilities of chips.
On the other hand, the competition in the complete vehicle market is currently fierce. Car price reduction, player reshuffle, and involution acceleration have become hot words in the automotive industry. With “cost reduction and efficiency increase” as the core, the overall automotive supply chain is under pressure, and it is transmitted to the intelligent driving chip link. An AI chip that compresses the cost to the extreme has become a rare commodity.
Taking both performance and cost into consideration, car companies “need both” and the industry’s calls for cost-effective smart driving chips are getting louder. From Tesla’s FSD chip, to Mobileye and Nvidia’s mass-produced auxiliary chips, to domestic car core companies Horizon and Xinchi, global car core companies are competing for the “king of cost performance.” Recently, the artificial intelligence perception and edge computing chip company Aixin Yuanzhi, which has just officially announced its entry into the car core track, is also one of them.
In the era of big models, what kind of chip does the automotive industry need? How to achieve the balance between high performance and low cost? Focusing on these two major issues, during the World Artificial Intelligence Conference held not long ago, “Jiazi Guangnian” interviewed Qiu Xiaoshen, founder and CEO of Aixin Yuanzhi, and Gong Huimin, president of Aixin Yuanzhi Automobile Division, and listened to their views on the current development trend of the intelligent driving industry.
Founded in 2019, Aixin Yuanzhi is mainly engaged in smart cities and intelligent driving. Its self-developed and produced SoC chip AX650N and car-standard chips M55 and M76 series can support the operation of the “Transformer” large model. It is currently one of the few chip suppliers in China that can adapt the “Transformer” large model on the edge side.
Regarding the significance of large models to the intelligent driving industry, Qiu Xiaoxin said that multi-modal large models help intelligent driving realize full AI optimization from perception to decision-making, planning, and control, while helping to cover all road scenarios and alleviate the “long tail problem” of intelligent driving. From this perspective, the large model is like the “ultimate ideal” of intelligent driving. To realize this ideal, a side-to-side AI chip that is cost-effective enough to operate large models with the highest efficiency is indispensable. In addition to the algorithm, a suitable chip is the biggest problem in the landing of the smart driving model.
In the eyes of Huang Wuling, deputy director of the Cognitive Intelligence Laboratory of the Xiongan Innovation Institute of the Chinese Academy of Sciences, “Autonomous driving technology is entering a new stage represented by multi-modal perception and cognitive understanding driven by data and knowledge.” With the emergence of large language models and large visual models, the autonomous driving industry has ushered in new opportunities for development.
A complete set of intelligent driving solutions usually includes four links: perception, planning, decision-making, and control. In the past, in the smart driving pre-installation market, perception was the first link to use the “model algorithm”, and the GPU part of the car chip was mostly used for the calculation of the “perception” link. Wang Xiaogang, the co-founder and chief scientist of SenseTime, once pointed out, “At present, most of the perceptual output is a result, some judgments are made based on rules, decisions are made, and then planning control is realized based on manual rules.”
This kind of control method is not smart enough and not precise enough, and the arrival of large models will bring changes to this situation.
Gong Huimin has a judgment on this trend, “Current in-vehicle intelligence is a multi-tasking system, and those complex tasks require a lot of small models to realize separately. The future of intelligent driving will move towards end-to-end, which can be replaced by a single large model or a small number of models, which will improve the efficiency and performance of intelligent driving.”
In addition, another problem in the intelligent driving industry is that driving scenarios cannot be fully covered. The intelligent driving market is currently divided into four levels: all-in-one, driving and parking, high-speed NOA, and urban NOA. The technology has achieved coverage of more than 90% of the scenarios, but there are still 10% of long-tail scenarios that cannot be solved. The large model can just rely on its ability in environmental cognition and understanding, intelligent decision-making, etc., to achieve the ability similar to human driving, and to cover niche scenes that have not been encountered before.
Qiu Xiaoxin summarized the advantages of large models in the field of intelligent driving into three points, “One is the ability to achieve ‘single model multi-tasking’, replacing all the parts that originally belonged to rule judgments with AI; the other is stronger generalization. There are still 10% of the scenes in intelligent driving technology that are difficult to cover, and large models can alleviate the ‘long tail problem’ of intelligent driving; the third is to further improve the accuracy and computing efficiency.”
Based on such advantages, at present, domestic intelligent driving solutions based on the “Transformer” large model are developing rapidly, and smart driving algorithm companies such as Momo Zhixing and Banma Zhixing, as well as vehicle companies such as Ideal, have “joined the car”. Among them, the most popular is the BEV large model combined with Transformer. At present, it is the consensus of domestic smart driving top players to implement BEV within this year.
The advantages of large models in the field of intelligent driving are obvious, so what are the difficulties in promoting its implementation?
Qiu Xiaoxin believes that the key is to have a cost-effective chip that can meet the needs of car companies. On the one hand, as the competition in the domestic auto industry becomes more and more serious, car companies require that L2+ autonomous driving solutions should further reduce prices in the future, and chip companies should have stronger cost control capabilities; on the other hand, the complexity of autonomous driving algorithms will require chip computing power to continue to rise. In summary, a chip that can support Transformer’s high-efficiency operation while meeting low-cost requirements will become the core demand of car companies.
For this trend, the chip architects of AISI have made a preliminary judgment as early as 2021. Since then, they have pre-embeded the support capability of the “Transformer” large model for the AIXTECH® mixed-precision NPU developed by AISI. At the same time, the first car-standard chip M55H developed by it will also pass the certification in July 2022.
This kind of predictability is also the confidence that Aixin Yuanzhi chose to enter the intelligent driving track in June this year. At present, Aixin Yuanzhi has become one of the few chip suppliers in China that can adapt the “Transformer” large model on the edge side. At the end of June, the M55H chip of AICIN has been mass-produced in two models, and the latest generation of M76 series chips has also entered the certification process.
According to Gong Huimin, its M76 series chips can fully meet the computing needs of BEV and Transformer large models in terms of model support and operator richness, and are the chips with the best performance in this segment of the market.
As the power of the large model is further revealed, the car-standard chips that can support the efficient operation of the “Transformer” will become more and more important. According to the data of Chebai Intelligence, in 2022, the global automotive AI chip market has reached 10.2 billion US dollars, which is the largest demand market for AI chips. At the moment when BEV large-scale models are constantly landing, Aixin Yuanzhi will gain greater market opportunities by virtue of its accumulation in the field of AI chips.
At present, the competition in the new energy vehicle track is becoming more and more fierce. Grasping the opportunity brought by the large model, on June 8, Aixin Yuanzhi announced its entry into the smart driving track. However, the automotive supply chain has its own set of strict hierarchy logic. Car companies pay great attention to decision-making security and are stingy about the high cost of replacing suppliers. Therefore, it is not easy for a newcomer in the industry to take root in the smart driving track.
The window opportunity for Aixin Yuanzhi to enter the smart driving track, one comes from an external industry opportunity, that is, the three-year shortage of car chips; the other comes from Qiu Xiaoshen’s prediction of the industry. She decided to deploy the smart driving industry in 2021.
Since 2020, the auto industry has been caught in a wave of core shortages, and the production capacity of auto companies has been greatly restricted. Zhang Yongwei, vice chairman and secretary-general of the China Electric Vehicle 100, said at the end of 2022, “The three-year chip shortage has reduced China’s car production by more than 2 million.”
Regarding industry predictions, Qiu Xiaoxin emphasized that for chip companies, the amount of shipments is the most critical factor affecting the choice of track. Aixin Yuanzhi has never limited its strategic layout to a certain field of smart city or intelligent driving, but regards the potential application fields related to visual perception, NPU, and computing as the market direction that the company can touch. How to make the two core IPs developed by Aixin Yuanzhi self-developed – Aixin Zhimou ® AI-ISP and Aixin Tongyuan ® mixed precision NPU to empower the industry to the greatest extent is the underlying logic of the company’s thinking.
In 2019, smart cities were the second largest AI chip market after mobile phones. By 2022, the market size of automotive AI chips will reach 10.2 billion US dollars, three times that of the mobile phone market. The market space is already huge, and the current demand for localized alternatives has just started. It is not too late for Aixin Yuanzhi to enter the smart driving track at this time. According to the statistics of Chebai Think Tank, in 2023, GAC Passenger Vehicle’s domestic chip share will reach about 20%, and SAIC Passenger Vehicle will also increase the domestic chip share to 20% in 2025. The development of China’s automotive chip track has just begun.
In the face of industry opportunities, the relative advantage of AIC is obvious. Its two core IPs, AI-ISP and AI-ISP are highly mature and have been polished in the smart city track for many years. These two technologies can be easily reused in the field of intelligent driving. The two major tracks work together, and the research and development costs of the chip are shared to the greatest extent. Therefore, the car chip of Aixin Yuanzhi will be more cost-effective.
Regarding the power of AI-ISP® AI-ISP and AI-ISP® mixed-precision NPU technology on the smart driving track, Qiu Xiaoxin said, “The technical logic of the vision processing chip applied to the camera and the intelligent driving chip are the same, both are visual processing power + computing power, which happens to be the field that AI-Smart has been constantly iterating in the past few years.”
She introduced, “In terms of visual processing capabilities, the ISP technology pioneered by Aixin applies artificial intelligence’s Deep Learning to image processing, which can clearly see things when the ambient light is bad, making the night as bright as day, and meeting the perception needs of cars in rainy, cloudy, and night times. Aixin’s self-developed mixed-precision NPU can effectively improve chip performance and power consumption. “
At present, in the field of automotive chips, Aixin Yuanzhi has formed a product sequence including three new series of chips, M55, M76, and M77. Among them, M55H is the core of its current mass production, adopting 1V1R hardware configuration, mainly for L2 level automatic driving, CMS and DMS development kits and other applications, with 8TOPS computing power. After winning the mass production of two models, the company plans to win new mass production orders in the third quarter.
The M76 series chips are aimed at the L2+ intelligent driving market, with a planned computing power of 60TOPS, while the M77 series chips are aimed at the high-end intelligent driving markets such as urban NOA. All three chips can support the Transformer algorithm, showing an increasing trend in capacity. By early next year, the M76 chip is expected to pass the vehicle-level certification, and by 2025, the M77 chip is expected to be mass-produced.
Based on the above-mentioned advantages, Aixin Yuanzhi has formed a computing power platform with a three-layer architecture of platform layer, model layer, and application layer. The platform layer is the artificial intelligence computing power platform of software + chip, the model layer is the Transformer large model, and the application layer includes the three major tracks of smart city, smart driving, and smart IOT. Based on this platform, Aixin provides car companies with software + hardware solutions, and provides a mature tool chain to help car companies get started quickly.
In the era of large-scale models, the support of the Transformer algorithm by AICIN Yuanzhi car gauge chips can make its car cores in the industry’s leading position in terms of performance and power consumption. As the large model of autonomous driving Transformer represented by BEV accelerates to the ground this year, the development advantages of AICIN Yuanzhi NPU chips will be magnified.
In the further future, with the gradual increase in the localization rate of automotive chips, car companies will begin to pursue cost-effective chips. Chip companies such as Aixin Yuanzhi, which have unique advantages in chip technology and can effectively reduce the cost of core-making by laying out multiple tracks, will gain enough room for development.
(Cover image source: provided by the interviewee)