From March 15th to 17th, the 2024 China EV100 Forum was held at Diaoyutai State Guesthouse in Beijing. At the Intelligent Ecosystem Automotive Forum themed "Building a Diversified, Interconnected and User-Oriented New Ecosystem for Smart Vehicles" on the afternoon of the 17th, Axera, an artificial intelligence chip R&D and basic computing power platform company, was invited to participate. Gong Huimin, President of Axera's Automotive Business Unit, delivered a keynote speech, elaborating on the evolution of intelligent driving computing power and algorithms, as well as the latest implementation achievements of Axera Auto intelligent driving chips.

Unlike physical models that generate virtual worlds such as Sora/GPT, intelligent driving based on real physical laws is undergoing a process from real world study to real world model. The continuous development of technology has also brought about an increase in computing power and upgrades in algorithms. Gong Huimin pointed out that on the one hand, the exponential increase in cameras and radars has ushered in an era of high growth in computing power for intelligent driving. On the other hand, continuously optimized algorithms have enabled models to focus on processing elements that are more concerned by back-end modules, reducing the waste of computing power.

"By increasing computing power to enhance computational capabilities, and through algorithmic focus to enable computing power to better solve key issues in autonomous driving, we can concentrate our efforts on a single point. The ultimate goal is to achieve sufficiently intelligent autonomous driving performance in our end-side real-time application system for autonomous driving," said Gong Huimin. The "one plus, one minus" of computing power and algorithms is a process of focusing. With more focused algorithms, large computing power can enable intelligent driving end-side applications to achieve sufficiently excellent performance, that is, to break through the performance and experience bottlenecks of AI intelligent driving through sufficient "pressure intensity".

With the further development of large models and Sparse Transformer, end-to-end large models for autonomous driving have become possible. In this process, Transformer will replace the mainstream position of CNN operators and become a necessary cornerstone for building a unified model of autonomous driving.

Although Transformer is sweeping the AI field, there are not many processing platforms that can natively support Transformer at this stage. The reason, according to Gong Huimin, is that the insufficient flexibility of the original NPU design and the lack of accuracy in common PTQ deployment are affecting the implementation process of Transformer technology.

In addition, Axera M76/M77 platforms natively support various Transformer operators and variants. The flexible software and hardware co-design can also quickly adapt to new operator requirements, fully supporting the continuous evolution of the intelligent driving industry from backbone networks, BEV spatial mapping, sparse perception to multi-modal fusion and end-to-end large models. Among them, M76 is the industry's first automotive-grade SoC with native Transformer support and the only single SoC passively cooled chip for integrated parking and driving in the market. Its SwinT Transformer performance has exceeded that of Orin-X, featuring high efficiency and cost-effectiveness. The self-developed Axera Smart Compute mixed-precision NPU fully supports BEV (100%) and passive cooling for all vehicle models (100%), making it a perfect alternative to multi-chip high-speed NOA solutions. M77, on the other hand, is a true end-to-end computing chip.

Currently, the automotive industry chain is changing rapidly, and the ecological environment is becoming increasingly complex. How can ecological enterprises coexist and thrive? In this regard, Gong Huimin stated that as a chip R&D enterprise, Axera will unwaveringly focus on being a Tier 2 supplier, leaving more room for algorithm partners and Tier 1 manufacturers.

Gong Huimin introduced that Axera officially entered the automotive chip market in June 2023 and launched its automotive brand Axera Auto at the end of November. At present, it has established long-term strategic cooperative relationships with many outstanding partners. The launched automotive chip products have completed fixed-point applications in a number of important vehicle models, including front-view ADAS, electronic exterior mirrors, integrated parking and driving, and other intelligent applications where Axera's presence can be seen.

Among them, M55 has become a powerful tool to meet the future standard configuration of AEB with its excellent image quality and safety. In the latest 2023 China market passenger car standard front-view integrated machine/domain controller domestic computing solution share released by Gaogong Intelligence, Axera ranked second. The latest 2024 China ADAS SoC Vendor Leadership Matrix released by the international authoritative consulting firm Canalys also shows that Axera has entered the Leader Quadrant, becoming the fastest-growing local emerging intelligent driving SoC supplier. To meet the multi-scenario needs of intelligent driving, Axera will gradually realize larger-scale mass production and application in the second half of this year.

The tide of automotive intelligence is rolling in, and competition is becoming increasingly fierce. Gong Huimin believes that only by "being half a step slower in ideals and three steps faster in implementation" can we meet the emerging new challenges. The ultimate innovation in Axera Auto M55, M76, and M77 chips helps partners transform technology into differentiated value, into sales volume, and into profits, enabling them to have a place in industrial development. Truly achieving that the sun and moon have different lights, and day and night each have their own suitability, thus forming a well-developed automotive ecosystem.