The Zhitong Finance App learned that China Merchants Securities released a research report saying that the R&D pace of Shengteng 960 is moving forward, domestic computing power is accelerating towards “calculation, connection, storage, and management” system-level collaboration. It is recommended to focus on the AI computing power industry chain and investment opportunities on semiconductor autonomous and controllable circuits. The conference's Shengteng 960 rhythm exceeded expectations, the promotion of Hi-One mass production, and the 4096-card supernode and million-card cluster plan are expected to simultaneously drive demand for computing power chips, advanced packaging and HBM, high-speed optical interconnection/NPO, network switching, and AI storage. It is recommended to focus on OEM, GPU/ASIC, network chips and optical chips, packaging and testing, equipment, components, and storage.
The main views of China Merchants Securities are as follows:
In the future, Huawei will focus on AI infrastructure to create a silicon-based black ground for a smart world
Huawei's AI strategy mainly focuses on 7 aspects; 1) The core of Huawei's AI strategy is computing power, adhering to hardware monetization; 2) “supernode+cluster” to build a solid computing power base in China and build new choices for the world; 3) building an open source and open computing power ecosystem to support native training of mainstream big models and actively embrace “100 models and thousands of states”; 4) Use the Pangu Big Model to intelligently enhance the competitiveness of its own products; 5) Provide flexible offline+offline computing power solutions to accelerate industry intelligence; 6) Innovate diverse computing power to promote AI integration, Get on the car, smart and omnipresent; 7) Surrounded by” “Use computing” to build a next-generation communication network and deliver intelligence to everyone, home, and business.
The 11-chip family in the Lingqu series was unveiled. The Shengteng 960 was ready ahead of schedule, and DT was brought forward to 27Q1
Building supernodes and supernode clusters based on Lingqu UniFiedBus, 11 key chips were developed, including: 1) computing: total computing processor (Kunpeng), intelligent computing processor (Shengteng); 2) connection: Lingqu interconnection chip, NPO high-density optical engine, network card chip; 3) storage: smart storage chip, hard disk control chip, storage control chip; 4) tube: Lingqu management chip and device management control chip. In terms of Ascend chips: 1) Ascend 960DT: FP8 computing power 2PFLOPS, bandwidth 9.6Tb/s, internet bandwidth 2.2Tb/s, original plan 27Q4 ready, now advanced to 27Q1; 2) Shengteng 960PR: FP8 computing power 2PFLOPS, 192 GB of access, 2.4 Tb/s, original plan 27Q4 ready, now advanced to 27Q3; 3) Ascend will stick to the evolutionary pace of one year. 970 and 980 chips will be launched in 2028 and 2029 70 expected FP8 computing power 3.6 PFLOPS, 288 GB of access memory, 14.4 Tb/s of bandwidth, 980 is expected to have 7.2 PFLOPS of FP8 computing power, 384 GB of access, and 38.4Tb/s of bandwidth.
With the Lingqu architecture and Hi-one optical product series, the 960 supernode will be the first product to use NPO
The industry's first mass-produced optical engine was also unveiled at the conference, using “Lingqu UnifiedBus+Hi-One” to build a supernode interconnection system that can be expanded on a large scale, pioneering the SIN-SOI silicon optical mixing process, the industry's first NPO with a built-in light source, and the industry's maximum transmission capacity of 7.2 Tbps. The industry's first supernode, the 960 supernode, was released. Each supernode has a size of 4096 cards, and can provide up to 8EFP8 computing power, up to 1PB of HBM capacity. 5,500 Hi-One are used to replace the 48,000 800G optical modules originally required, reducing power consumption of over 550 kilowatts.
The Kunpeng series has been upgraded again, and AI memory storage clusters and Agentic supernode clusters help the 10 trillion model
The Kunpeng Super Node has been upgraded. Based on the Lingqu all-optical networking, the Kunpeng Super Node supports a maximum of 4096 nodes and forms a unified memory pool of up to 256 TB, which can significantly accelerate agent performance. Based on Lingqu's AI memory storage-OceanStor M900 cluster, Huawei proposed a multi-layer KV cache architecture and launched a Huawei L3.5 layer PB-level KV cache based on Lingqu UnifiedBus and supporting “one-hop direct connection”. Based on the unified interconnection of Lingqu UnifiedBus, various interconnection protocols are unified into the Lingqu protocol, and protocol conversion costs are greatly reduced, and subsystems such as Shengteng Super Node, Kunpeng Super Node, and KV cache clusters can be equally interconnected. Through two-layer CLOS four-plane networking, the maximum cluster size can reach 512,000 cards, and in combination with multi-track topology technology, it can support up to 1 million kilogram ascending supernode clusters.
Risk warning: demand for terminals falls short of expectations; macro-environmental impact; domestic semiconductor substitution process falls short of expectations; increased industry competition, etc.