2024 China EV100 Forum: Terry Gong Emphasizes XMOTIVE’s Native Support for Transformers, Positioning It as a Future Automotive Chip Solution
From March 15 to 17, 2024, the China EV100 Forum was held at the Diaoyutai State Guesthouse in Beijing. Axera, a key player in AI chip development and foundational computing platforms, participated in a focused session on March 17, themed “Creating an Integrated and User-Friendly New Intelligent Vehicle Ecosystem.” At this session, Terry Gong, President of Axera’s Automotive BU, delivered a keynote address, in which he discussed the advancements in computing power and algorithms essential for smart driving technologies and shared the latest deployment successes of Axera’s smart driving chips.
Enhanced Computing Power and Refined Algorithms
The impact of the Transformer architecture as a cornerstone is emerging.
Unlike models like Sora/GPT that generate virtual worlds, smart driving, which adheres to real physical laws, is undergoing a transformation from real-world study to real-world model. Continuous technological advancements have also led to enhanced computing power and upgraded algorithms. Gong pointed out that, on one hand, the exponential increase in cameras and radars has ushered in an era of high growth in computing power for smart driving. On the other hand, continuously optimized algorithms allow models to focus on elements that are more relevant to back-end modules, reducing waste in computing power.
“By enhancing computing power, we supercharge what our systems can handle, and by refining algorithms, we make sure this power is focused on the crucial aspects of self-driving technology. The end goal is to achieve truly smart autonomous driving performance in real-time terminal applications,” said Gong. He emphasized that enhancing computing power and refining algorithms is a process of focusing resources. With greater computing power and more streamlined algorithms, smart driving applications at the terminal can achieve exceptional performance, effectively overcoming the challenges of AI-driven autonomous driving through concentrated “pressure.”
Advancements in large models and Sparse Transformers have paved the way for end-to-end large models in autonomous driving. This shift will see Transformer taking over from CNN as the main operator, becoming the key building block for creating unified models in self-driving technology.
Native Support for Transformers
XMOTIVE Develops Future Automotive Chip Solutions
Although the Transformer model is sweeping through the AI field, there are currently only a few processing platforms that can natively support it. According to Gong, the key factors hindering the implementation of Transformer technology are the insufficient flexibility of existing NPU designs and the lack of precision in commonly used PTQ deployments. Additionally, relying on CPU subgraph computations exacerbates inference speed loss. Therefore, the industry urgently needs more suitable computational platforms.
The complexity of implementing Transformers requires chip platforms to enhance inference efficiency while also offering more flexible deployment capabilities. In response to these requirements, Axera, leveraging its self-developed Axera Tongyuan AI processor , has launched the M76/M77 series chip platforms for future smart driving applications.
According to Terry Gong, the M76/M77 series chips achieve industry-leading Transformer inference efficiency through native support for Transformer operators and optimized data I/O. Furthermore, the Axera Tongyuan NPU provides versatile support for deploying mixed precision models, facilitating the rapid implementation of PTQ-optimized Transformer models using Axera’s toolchains, while ensuring that the precision matches the required deployment criteria.
Additionally, the M76/M77 chips natively support various Transformer operators and their variants. The flexible co-design of hardware and software allows for quick adaptation to new operator requirements, fully supporting the smart driving industry’s continuous evolution from backbone networks and BEV (Bird’s Eye View) maps to sparse perception, multimodal fusion, and end-to-end large models. Specifically, the M76 is the industry’s first SoC designed with native Transformer support. It is also the only single-SoC chip featuring passive cooling, tailored for integrated driving and parking functions. Its Swin Transformer performance surpasses that of Orin-X, offering high efficiency and cost-effectiveness. Axera’s self-developed Axera Tongyuan NPU provides 100% support for BEV and passive cooling across all vehicle models, making it a perfect replacement for multi-chip high-speed NOA solutions. Meanwhile, the M77 is a truly end-to-end computing chip.
Stay Committed to Being a Tier 2 Supplier
Expediting the Implementation of Autonomous Driving
In the rapidly evolving automotive industry with its increasingly complex ecosystem, how can companies within this ecosystem achieve mutual prosperity and symbiosis? Regarding this question, Gong responded that as a chip designer, Axera is committed to excelling as a Tier 2 supplier, thereby providing more space for algorithm partners and Tier 1 suppliers.
Gong noted that Axera officially entered the automotive chip market in June 2023 and launched its automotive brand, XMOTIVE, at the end of November 2023. Axera has already established long-term strategic partnerships with many outstanding partners. The automotive chip products it has launched are now designated for several key vehicle models, featured in intelligent applications such as ADAS front-view all-in-one solutions, CMS, and domain controllers for driving and parking.
Specifically, the M55, with its excellent image quality and security features, has become a powerful tool for meeting future AEB standards. According to the latest report by the GaoGong Intelligent Automobile Research Institute on the market share of front-view integrated systems and domain controller computing solutions for passenger cars in China in 2023, Axera secured the second position. The China ADAS SoC Leadership Matrix 2024 report released by the international authoritative consulting firm Canalys also shows that Axera has entered the Champion’s quadrant, becoming the fastest-growing local SoC supplier in the smart driving sector. To meet the diverse needs of smart driving scenarios, Axera plans to gradually achieve larger-scale mass production of its chips in the second half of 2024.
As the push for smarter vehicles accelerates and competition intensifies, Gong believes that only by being “cautious in planning and swift in action” can new challenges be met. The cutting-edge innovation in XMOTIVE’s M55, M76, and M77 chips helps partners turn technology into unique value, sales, and profit, securing their place in the industry’s growth. This approach ensures that “the sun and the moon shine differently, and day and night each have their own advantages,” thereby fostering a robust automotive ecosystem.
