The China Academy of Information and Communication Technology has released the first domestic standard for large automobile models.
The standard mainly covers three competency areas. Among them, scene richness focuses on evaluating the support of large car models for segmented scenarios such as smart cockpits and autonomous driving. Capability support focuses on the performance of large car models in artificial intelligence technology capabilities such as perception, understanding, reasoning, and generation. Application maturity mainly assesses the application of large car models in system ecology, deployment customization, and scenario adaptation.
GF Securities said that the PCB area, number of layers, and material performance of AI servers continue to improve, driving the increase in the value of standalone PCBs. Compared to DGX series servers, the value of a single GPU PCB in the GB200 NVL72 is expected to increase. The AI PCB market continues to grow, and the share of HDI value is increasing.
Domestic manufacturers are actively deploying AI server PCBs, and the industrial chain has ushered in new opportunities. From the “iPhone moment of AI” to the “moment of AI's fragmentation”, it is recommended to focus on the core beneficiaries of the industrial chain.
Companies related to the big AI model:
Shangtang-W (00020): At the 2024 (18th) Beijing International Automobile Exhibition, Shangtang Zhuying unveiled a series of original large-scale models and related product matrices. For the first time, it showcased the road test performance of UniAd, an end-to-end autonomous driving solution for mass production. At the same time, it also brought an AI large model cockpit product matrix with a multi-modal scene brain as the core and a new 3D interactive cockpit demonstration.
Shang Tang Zhuying's UniAd autonomous driving solution shows that whether on complex urban roads or rural roads without a central line, vehicles equipped with UniAd can efficiently and accurately complete a series of difficult operations such as turning left at a large angle onto the bridge, avoiding occupied vehicles and construction areas, and detour pedestrians to “drive like a human”.
Since it was first released in April last year, Shangtang's “Japanese New SenseNova” large model system has officially launched five major iterations. Based on more than 10TB of tokens training and covering a large amount of synthetic data, the new “Daily New 5.0” uses a hybrid expert architecture, and the context window can be effective up to about 200K during inference.