DeepSeek-V3.1 launch has completely ignited the domestic computing power supply chain.

On August 22, the domestic computing power industry took off. Computing chip manufacturers Cambricon (688256), Hai Guang Information (688041), and Yuntian Li Fei (688343) all surged 20CM, with Cambricon's market value exceeding 520 billion yuan and Hai Guang Information's market value surpassing 430 billion yuan.

Under the leadership of computing chips, the entire domestic computing power supply chain saw a significant rise. Semiconductor manufacturing company SMIC (688981) rose nearly 15%. Semiconductor equipment companies Northern Semiconductor (002371) and Zhiwei Company (688012) both increased by more than 6%. Huawei's computing power supply chain also performed well, with Chuanrun Co., Ltd. (002272) and Fangzheng Technology (600601) both surging 10%.

On the news front, on August 21, DeepSeek officially launched DeepSeek-V3.1. This large model uses a hybrid reasoning architecture and supports both thinking mode and non-thinking mode. It has higher thinking efficiency and can provide answers in a shorter time compared to DeepSeek-R1-0528. In addition, the new model shows significant improvements in tool usage and agent tasks. The new model has made breakthroughs in agent capabilities. According to official information, in code repair evaluation and complex task testing in command-line terminal environments, DeepSeek-V3.1 has shown significant improvement, and has made considerable progress in multiple search evaluation indicators.

Additionally, DeepSeek stated that DeepSeek-V3.1 uses UE8M0 FP8 Scale parameter precision. UE8M0 FP8 is designed for the upcoming next-generation domestic chips. This also indicates that future training and inference based on DeepSeek models are expected to use more domestic AI chips, helping to accelerate the construction of the domestic computing power ecosystem.

Currently, the main domestic computing power chip suppliers include Huawei, Hai Guang Information, Cambricon, as well as Muxi, Moer Thread, Suiyuan, Tian Shuzhixin, etc. In February this year, these computing power chips have all announced compatibility with DeepSeek's model.

However, according to the latest information from the People's Daily, the flagship AI training and inference product MTT S5000 of Moer Thread is the first batch of domestically produced GPUs that support FP8 natively and have been mass-produced. The MUSA architecture of Moer Thread natively supports hardware FP8 tensor acceleration calculations, which can well support UE8M0 FP8 Scale. Using native hardware FP8, it can achieve twice the floating-point computing power improvement compared to traditional FP16 computing, improve memory access and communication bandwidth efficiency, and increase storage capacity utilization, while optimizing the tensor expression precision.

In February this year, Moer Thread disclosed that since the start of DeepSeek's "Open Source Week," three code repositories have been opened up successively. Based on the new MUSA Compute Capability 3.1 computing architecture, Moer Thread can provide native FP8 computing capabilities and upgraded high-performance linear algebra template library MUTLASS, quickly supporting FlashMLA. Moreover, Moer Thread has optimized and implemented FP8 matrix multiplication on the new GPU architecture based on MUTLASS, supporting the corresponding functions of DeepGEMM, fully demonstrating the strong advantages of Moer Thread's MUSA architecture and full-function GPU in ecological compatibility and rapid adaptation.

Hai Guang Information previously announced that the Hai Guang DCU (Deep Computing Unit) has successfully completed the compatibility optimization with DeepSeek V3 and R1 models. DCU is a high-performance GPGPU architecture AI accelerator card launched by Hai Guang Information, dedicated to providing industry customers with an independent and controllable full-precision general AI acceleration computing solution. DCU has been widely applied in multiple fields such as education, finance, medical, government, and intelligent computing centers.

A person in the chip industry said that the latest FP8 is not supported by domestic chips at present, and support will be implemented in stages. First through software compatibility, and then gradually through hardware-native support.

Regarding the update of DeepSeek's model, which clearly supports FP8 precision and the upcoming next-generation domestic chips, CICC Research Report believes that the support of leading domestic open-source models for domestic chips is expected to promote the accelerated deployment of the domestic computing power ecosystem.

CICC also mentioned that Tencent stated at the earnings meeting that there are multiple choices for the supply channel of inference chips. Under the background of international supply chain fluctuations, domestic inference computing chips are expected to provide assistance. Huawei Ascend chips have recently participated in tenders for government, financial, and operator industry customers, proving that the competitiveness of domestic chips continues to improve.

According to public information: On August 12, the 2025 Financial AI Inference Application Implementation and Development Forum was held in Shanghai. At this forum, Huawei introduced an AI inference innovation technology - UCM Inference Memory Data Manager, aiming to promote the upgrading of AI inference experience and enhance the cost-effectiveness of inference.

Currently, artificial intelligence has entered the deep water area of development, and AI inference has become the key stage for the next explosive growth. To ensure a smooth inference experience, enterprises need to continuously increase their investment in computing power. However, how to find the optimal balance between inference efficiency and cost has become an important issue that the entire industry needs to solve urgently.

Huawei's UCM Inference Memory Data Manager includes three components: the inference engine plugin (Connector) that connects different engines and computing power, the functional library (Accelerator) that supports multi-level KV Cache management and acceleration algorithms, and the high-performance KV Cache access adapter (Adapter). Through the collaboration of the inference framework, computing power, and storage, it achieves "better experience, lower cost" for AI inference.

Aside from DeepSeek increasing its support for domestic computing power chips, there are also reports that NVIDIA has asked some component suppliers to suspend production of H20 chips.

On August 22, Foreign Ministry Spokesperson Mao Ning presided over the regular press conference. A reporter from Bloomberg asked about the report that NVIDIA had asked some component suppliers to suspend production of H20 chips. What is the Foreign Ministry's comment on the latest report about NVIDIA stopping the production of H20 chips? Mao Ning replied, "This question is recommended to you to ask the relevant Chinese authorities. As a principle, we have always believed that all parties should jointly maintain the stability and smoothness of the global supply chain."

Reporter from the People's Daily, Zhou Ling

(This article comes from the People's Daily. For more original information, please download the "People's Daily" app.)

Original: https://www.toutiao.com/article/7541388630808478234/

Statement: This article represents the personal views of the author. Welcome to express your attitude using the 【Up/Down】 buttons below.