Automotive chips are 迎来 DeepSeek moment

Moment, this concept originated from the new republicanism of the Cambridge School in the West in the 20th century.

In Pocock's book "Machiavellian Moment," the moment is given a special meaning, describing the formation of a strong community from weakness and dispersion, referred to as the "moment."

In the process of technological development, we have experienced many moments: the "iPhone 4 moment," when Jobs unlocked a new era with the iPhone 4; the "ChatGPT moment," when OpenAI brought large models, making generative AI the center of a storm.

The emergence of ChatGPT was in 2023, and two years later, China experienced the "DeepSeek moment." Some jokingly said: "In 2023, ChatGPT will replace many jobs in the future; in 2024, ChatGPT will replace many jobs in the future. In 2025, DeepSeek will replace ChatGPT's jobs."

The arrival of DeepSeek can also be called the "DeepSeek moment." This represents the idea of making advanced technologies that are profound, difficult to understand, and expensive available in low-cost or free products.

Under the influence of DeepSeek, automotive chips have also welcomed the DeepSeek moment.

Three Ways to Integrate DeepSeek

After DeepSeek appeared, many car companies quickly announced their integration with DeepSeek. As of now, companies include: Geely, Chery, BYD, Great Wall, Leapmotor, Voyah, Dongfeng, IM Motors, Changan, SAIC-GM, GAC, FAW-Volkswagen, BAIC Jihe, Smart, Jianghuai, FAW Hongqi, FAW Bestune, SAIC Baojun, ZEEKR, etc.

However, the integration situations vary among companies. From the specific application scenarios of integrating DeepSeek:

Geely will deeply integrate its self-developed large model Xingrui with DeepSeek-R1, but no specific integration timeline has been provided. In the future, Geely's Xingrui large model may provide users with interactive experiences such as vehicle control, proactive dialogue, and proactive after-sales service.

ZEEKR's news is more specific, with the integration of the DeepSeek large model focusing on the intelligent cockpit. After integrating with its self-developed Kr AI large model, the capabilities of the voice assistant AI Eva will be enhanced.

Dongfeng Motor's subsidiaries, including Voyah, Mengshi, Yipai, Fengshen, and Nano, will successively adopt the DeepSeek large model in the near future.

In terms of specific models, the Dongfeng Mengshi 917's intelligent cockpit has completed the integration of the DeepSeek-R1 model, with some models receiving updates through OTA upgrades before the Shanghai Auto Show in April. The Voyah Zhiyin will become the first mass-produced model in the automotive industry to integrate DeepSeek, with users able to update via OTA after February 14.

IM Motors' integration method is a unified model interface platform. The intelligent cockpit will introduce domestic large models such as DeepSeek, Doubao, Tongyi Qianwen, and Zhipu AI, with a unified model interface platform where different functions call different large models. The implementation timeline has not yet been clearly communicated The way Baojun cars are integrated is through a "dual model deployment" of the DeepSeek large model and a self-developed central large model. The system backend will automatically call the corresponding large model based on applicable scenarios. Scenarios include intelligent voice assistants, intelligent recommendations, and more.

Combining the recent integration situations of DeepSeek large models announced by major car manufacturers, it can be seen that the focus of DeepSeek's integration is still in the intelligent cockpit field.

There are roughly three types of integration methods:

One group of car manufacturers, such as Geely, chooses to leverage DeepSeek's distillation capabilities to train their self-developed large models, which is equivalent to a complementary training. Through fine-tuning integration, their original models become a complete large model.

Another group of car manufacturers, like Dongfeng's Voyah, Mengshi, Yipai, Fengshen, and Nami, as well as Great Wall Motors and SAIC IM Motors, will have a unified large model interface platform that not only includes DeepSeek but also covers other common large models like Doubao and Zhipu AI.

Of course, there is also a third wave of car manufacturers that are riding the trend. For example, FAW-Volkswagen has indeed integrated DeepSeek, but not in the car, rather in content production. FAW-Volkswagen announced that its "New Media AI Content Operation Platform" has integrated DeepSeek, covering over 60 dealers and more than 1,000 accounts, reducing the workload of sales personnel. In simple terms, the production of marketing materials uses DeepSeek, which is not closely related to the car itself.

Regardless, the DeepSeek integration announced by car manufacturers is currently focused on the intelligent cockpit. DeepSeek serves more as a voice assistant, making the in-car voice assistant sound more human-like. A voice assistant capable of writing perfect essays is not particularly useful for car owners.

Previously, there was industry discussion about why so many car manufacturers announced their integration with DeepSeek but did not include "Li Auto, Nio, Xiaomi, and Ideal." Because if it’s just an upgrade of the voice assistant, then Ideal's MindGPT, Nio's NomiGPT, and Xiaomi's Xiao Ai have already achieved many of the capabilities of DeepSeek as a voice assistant.

Automotive Chips, DeepSeek Moment

However, using DeepSeek as a voice assistant is somewhat of a waste; in fact, DeepSeek's stronger potential lies in intelligent driving.

The intelligent driving solution is currently converging towards an end-to-end direction, iterating from a two-stage solution of BEV + Transformer to a one-stage solution. Previously, Waymo trained a one-stage end-to-end intelligent driving model called EMMA based on Gemini, demonstrating the potential of large language models to become specialized intelligent driving models. Moreover, the EMMA model can be self-supervised, has strong generalization capabilities, and after optimization training, it even surpasses specialized intelligent driving models in some performances, thus showcasing the possibility of training DeepSeek into an effective intelligent driving model.

DeepSeek's own COT (Chain of Thought, which refers to the thinking seconds and process displayed before DeepSeek gives an answer) capability can endow DeepSeek with performance capabilities that exceed those of EMMA Some visionary car companies have already released signals.

On February 10, BYD Chairman Wang Chuanfu stated that all new vehicles will be equipped with intelligent driving and connected to DeepSeek, with high-level intelligent driving starting to cover models priced below 100,000 yuan. Car manufacturers are beginning to install intelligent driving features, which were previously exclusive to luxury vehicles, into affordable cars priced under 80,000 yuan.

Coincidentally, brands like Changan, Geely, and Baojun are also doing similar things.

Changan announced that starting in 2025, it will completely stop selling non-intelligent new cars and will make laser radar available for models priced below 100,000 yuan; Geely will publicly release its AI intelligence strategy in early March this year.

The remarkable aspect of Deepseek-R1 is that it enhances model accuracy while significantly reducing memory usage and computational costs by redesigning the training process using a small amount of SFT data combined with multi-round reinforcement learning (which is the process of aligning the student model with the teacher model).

Transitioning to the intelligent driving field requires large-scale cloud computing clusters, including model training, simulation verification, and data closed-loop aspects, but inference tasks can be elastically scaled, constructing lightweight models. The cloud architecture can also seek optimization solutions with higher resource utilization in a heterogeneous computing architecture of CPU + GPU + FPGA. Ultimately, it paves a path for Chinese companies to break through under the constraints of limited chips.

XPeng Chairman He Xiaopeng also mentioned recently: "The DeepSeek large model has brought shockwaves to the global tech community—it not only achieves an experience comparable to OpenAI but also compresses costs to an extremely low level. In the next decade, AI will drive significant changes in the automotive industry."

In 2024, domestic intelligent driving will collectively enter the "end-to-end" era.

Many car companies are investing heavily, purchasing and even stockpiling computing power cards. Li Auto's Vice President of Intelligent Driving R&D, Lang Xianpeng, stated that with the continuous expansion of intelligent driving parameters and the future deepening of intelligent driving to L4 level, Li Auto's annual spending on computing power clusters alone reaches about 1 billion USD (approximately 7.28 billion yuan).

The "distillation method" used by DeepSeek allows for reduced reliance on high-computing power chips in non-safety domains, substituting with domestic industrial-grade or consumer-grade chips, further lowering overall costs.

Now, throwing money at computing power is no longer the only solution.

For example, on the Qualcomm 8650 platform, DeepSeek can reduce the inference response time from 20 milliseconds to 19 milliseconds, while the computing power utilization rate drops from nearly 100% to 65%. Originally, the cost of running urban NOA at 100 TOPS was about 7,000; with Deepseek's involvement, it is expected to be achievable within a cost of 5,000, and it may even allow the Horizon Journey 6E chip to run urban NOA, significantly promoting the implementation of high-level intelligent driving.

In terms of automotive computing chips, local chip manufacturers like Horizon and Black Sesame provide low-cost computing power solutions.

On the Black Sesame side, the Wudang C1200 family of chips has completed the deployment of the DeepSeek model, and the Huashan A2000 will also fully support multi-modal large models based on DeepSeek The Wudang C1200 family is a high-performance computing platform launched by Black Sesame Intelligence specifically for multi-domain integration and cockpit driving applications.

Yang Yuxin, Chief Marketing Officer of Black Sesame Intelligence, believes that "DeepSeek, as a representative of multi-modal large models, has its core value in promoting the upgrade of intelligent driving systems from 'perception-driven' to 'cognition-driven' through efficient edge-side inference capabilities, which helps lower development thresholds." The A2000 chip designed by Black Sesame Intelligence for the next generation of AI models has supported the deployment of current mainstream large models, helping automotive companies reduce algorithm adaptation costs and accelerate functional iteration through software and hardware collaborative optimization.

On the Horizon side, the Horizon Journey 6E chip is one of the most cost-effective computing chips in the current intelligent driving industry, capable of supporting intelligent driving functions such as high-speed navigation, automatic parking, and urban memory navigation, with a total price below 5,000 yuan, making it quite competitive.

According to Yu Kai, founder and CEO of Horizon, 2025 will mark a real turning point for intelligent driving, and the next three years will be crucial for competition in intelligent driving.

Huang Rui from Dongfeng Motor Research Institute believes that "the automotive industry's reliance on sensors may still exist in the short term. For high-end training chips and inference chips, as technologies represented by DeepSeek develop, the degree of reliance may decrease."

Conclusion

However, currently, over 80% of large models globally are trained based on NVIDIA chips. In the short term, it is not easy to replace NVIDIA chips with the DeepSeek + domestic chip solution. Further adaptation is also needed for intelligent driving in vehicles.

When faced with the question of whether DeepSeek can reduce AI's reliance on hardware in the automotive industry, Huang Rui from Dongfeng Motor Research Institute stated, "The automotive industry's reliance on sensors may still exist in the short term. For high-end training chips and inference chips, as technologies represented by DeepSeek develop, the degree of reliance may decrease. I believe that domestic AI companies and related practitioners will carve out a development path belonging to our country through continuous innovation in algorithms and engineering."

Author of this article: Jiu Lin, Source: Semiconductor Industry Overview, Original title: "Automotive Chips, Welcoming the DeepSeek Moment"

Risk Warning and Disclaimer

The market has risks, and investment requires caution. This article does not constitute personal investment advice and does not take into account the specific investment goals, financial conditions, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances Invest based on this information at your own risk