At last week's Technology Innovation Day, NIO unveiled the latest NIO World Model (NWM) in the autonomous driving field, claiming to possess dual core capabilities of spatial and temporal understanding, surpassing end-to-end model deployment.
The autonomous driving chip, Shenzi NX9031, which was officially announced to have successfully taped out, is tailored for the NIO World Model. The Shenzi NX9031 is the world's first 5nm autonomous driving chip independently developed by NIO. According to NIO, one chip has the performance equivalent to four industry flagship chips (Nvidia Orin X).
In the past two years, autonomous driving chips have been a key breakthrough product direction for automakers. NIO and XPeng have been developing their own chips, with Li Auto starting slightly later. Based on the Longying No. 1, Geely's subsidiary, Xingjing Technology, has also been on the self-development path for years.
Li Bin has publicly stated that NIO purchased many Nvidia chips last year, costing the company a lot of money. Considering the procurement costs, the company decided to turn to self-developed chips. The official statement is that the Shenzi NX9031 can pay for itself in about a year.
There are many reasons for developing their own chips, but one of the main goals for the "Wei Xiaoli" (a collective term for NIO, XPeng, and Li Auto) is to free themselves from Nvidia's constraints. From industry reports, it appears that the self-developed chips are quite forward-looking and will match the latest trends such as end-to-end autonomous driving.
However, the Chinese contingent encircling Nvidia is not limited to the "Wei Xiaoli." This year, local chip suppliers are also "caught up" in the end-to-end competition. At the China Auto Forum last month, Horizon President Chen Liming clearly stated that end-to-end is currently the only feasible solution to the endgame of autonomous driving.
Lu Jianfeng, Vice President of AIChip Intelligent Vehicle Division, believes that end-to-end is the only way for advanced autonomous driving. Due to the long design and development cycle of chips, AIChip's strategy is to skip other models and focus on the One Model mode, similar to the UniAD technology architecture for NPU design.
From an industry perspective, the high cost of external procurement, the uncertain international situation, and the cost reduction benefits Tesla has previously enjoyed with its self-developed chips have all influenced domestic automakers' chip strategies and supply models.
The popularity of end-to-end large models not only catalyzed a new round of autonomous driving revolution but also accelerated the product and technological evolution of autonomous driving chips. This not only places higher demands on automakers for self-developed chips but also forces chip suppliers in the red ocean to speed up their internal competition.
The Self-Development Wave Has Arrived
Why are automakers developing their own chips?
Mastering Core Technology: Ensuring supply security and not being "strangled" by suppliers, especially powerful foreign suppliers.
Li Bin mentioned in an interview that the international impact on chip supply, due to US restrictions, has already had a real impact on China's automotive industry.
"Since last October, we haven't been able to use the world's most advanced chips for our cloud training. The autonomous driving team not only looks at cloud capabilities but also at group intelligence capabilities. While the risk for edge inference chips is currently low, we still need to be prepared for various changes."
Customization:
Industry experts told "Auto Commune"/"C-Dimension" that one of the key considerations for new automakers developing their own chips is to enhance product competitiveness through differentiation, as self-developed chips allow for customized functions.
For automakers, developing their own chips is costly but can reduce dependence on overseas chip suppliers, ensuring that "all eggs are not in one basket." Furthermore, self-developed chips can better match their own algorithms, addressing the coupling issue between algorithms and chip platforms.
In the past, Tesla's 144 TOPS computing power outperformed the 400-500 TOPS chips available on the market, mainly because the chip was designed for Tesla's own algorithms. Notably, Tesla's 144 TOPS computing power chip (Autopilot HW3.0), released in 2019, still supports end-to-end autonomous driving today.
Cost Reduction:
Li Bin stated at the press conference that NIO spent a lot of money on Nvidia chips last year. To reduce costs, NIO decided to develop its own chips, with one chip equivalent to four Nvidia chips, thereby lowering costs. According to Li Bin, the Shenzi NX9031 can pay for itself in about a year.
There are other considerations as well. Industry insiders note that promoting self-developed chips and making public commitments can positively impact the secondary market and brand perception. Additionally, self-developed chips can significantly improve the system experience, achieving strategic goals.
Notably, Tesla's early self-developed chips aimed to increase computing power and flexibility.
Reports indicate that XPeng's self-developed chip process closely follows NIO, with chips sent for tape-out, expected to return in August. Li Auto's chip development started relatively late, with the autonomous driving chip project codenamed "Schumacher," expected to complete tape-out within the year.
"A Means, Not an End"
Wu Xinzou, head of Nvidia's autonomous driving business, outlined that the development of autonomous driving can be summarized in three stages, with end-to-end being the final step.
First Stage: Completely rule-based.
Second Stage: AI large models gradually replace manual rules, completing prediction and planning.
Third Stage: Completely end-to-end large models, with AI covering the entire process from perception to decision-making.
In the third stage of autonomous driving, autonomous driving chips are highly challenging. AIChip Vice President Liu Jifeng expressed similar thoughts, stating that true end-to-end involves using large models for cloud training and validation, with the results applied to edge inference, placing significant responsibilities on chip companies.
Horizon believes that end-to-end is a means, not an end, requiring a combination of human-like experience, efficient computing, and agile delivery. End-to-end capability accumulation requires efforts in algorithm iteration, engineering foundation construction, and software-hardware integration, with software and algorithms playing a core role.
Horizon Algorithm Platform Chief Architect Mu Lisen believes that the essential capability of end-to-end lies in data iteration. Although it appears to be a forward-looking model structure, the iterative data behind it is more crucial, supporting the transition from laboratory technology to product-level maturity.
Chen Liming also acknowledged that Horizon faces difficulties with the constantly changing vehicle and sensor architectures, sensor layouts, and adoption. Despite collecting a lot of data, much of it is not high-quality or continuously usable, an issue beyond the scope of any single company to solve.
"Tesla's FSD V12.3 version was trained with 10 million sample videos, extracted from 10 billion high-quality samples. China still falls short. Moreover, the 10 billion samples were collected under a standard sensor framework, ensuring continuity for training the latest models."
Like Horizon, AIChip emphasizes its role as Tier 2, believing that the key demands for autonomous driving chips in end-to-end algorithms are high memory and multi-core large computing power.
Achieving end-to-end autonomous driving relies on critical computing chip support, including architectural innovation, core IP breakthroughs, and performance leaps.
Mu Lisen from Horizon explained to "Auto Commune"/"C-Dimension" that the technical threshold for end-to-end computing power competition lies in adapting to the computational demands brought by model structure changes and changes in operator focus.
On one hand, models will grow larger, and so will computational power; on the other hand, model structures will evolve, shifting from primarily CNN (convolutional neural networks) to mainly Transformer-based end-to-end models.
"Transformers are a broad algorithm category used in large language models (like ChatGPT) and end-to-end autonomous driving, with different operator focuses. End-to-end autonomous driving requires foundational matrix operations and additional operator support, presenting higher demands."
The Huawei camp also holds significant influence. Despite Nvidia dominating the autonomous driving chip market, China has a large contingent powered by Huawei, including brands like AITO, Avatr, Jihu, and Zhijie. Their vehicles' autonomous driving systems largely use Huawei's MDC810/MDC610 computing platforms.
With the efforts of autonomous driving chip suppliers and the accelerated rollout of self-developed chips by companies like NIO, in the coming years, the aspiration of "not being subject to Nvidia" in the domestic autonomous driving chip sector will gradually be partially realized.
