domestic AI models have been launched one after another, and the demand for intelligent computing power has blown out.

IDC predicts that the scale of domestic intelligent computing power is expected to enter the level of 1 trillion floating-point computing per second in 2026, and the annual compound growth rate of intelligent computing power scale is expected to reach 52.3 from 2021 to 2026.

computing power is a new type of productivity that integrates information computing power, network carrying power and data storage power. AI chip is the core of realizing high computing power. The current mainstream AI computing power chips mainly include GPU, FPGA, ASIC, NPU, etc. GPU has absolute advantages in training load.

The core competitiveness of GPUs lies in the architecture and other factors that determine the performance advancement and computing ecological barriers. At present, the absolute share of the global GPU market is occupied by NVIDIA and AMD, domestic enterprises mainly include Huawei Haisi, Haiguang Information (688041.SH) and so on.

NVIDIA's ability to become the world's number one GPU player is mainly due to its CUDA ecosystem established under its first-mover advantage. CUDA architecture provides a simple interface for GPU programming, which greatly reduces the development threshold. Therefore, CUDA has developed rapidly and has been widely used.

At the same time, due to the closed-source nature and fast iteration of CUDA, it is difficult for latecomers to be perfectly compatible through instruction translation and other methods. Even if they are partially compatible, there will be a large performance loss, resulting in a continuous lag behind Nvidia in cost performance.

industry insiders believe that it is easier for domestic GPU to develop in the initial stage to be compatible with the existing ecology and to survive first. however, in the long run, it is still necessary to get rid of the idea of compatibility and develop its own core technology after gaining a firm foothold.

DCU of Haiguang Information is a kind of GPU, which is highly similar to CUDA in terms of ecology and programming environment. CUDA users can quickly migrate to ROCm platform of Haiguang Information at a lower cost. At present, Haiguang DCU is mainly deployed in server clusters or data centers to provide high-performance and high energy efficiency computing power for applications and support high-complexity and high-throughput data processing tasks. Haiguang Deep Calculation II has been released, realizing commercial applications in big data processing, artificial intelligence, commercial computing and other fields.

Huawei has designed a heterogeneous computing architecture CANN with functions similar to NVIDIA CUDA, an AI computing framework MindSpore to help developers incubate various AI innovative algorithms and applications, and a full-process development tool chain MindStudio for AI scenarios.

currently, Huawei's rising chips include rising 310 and rising 910. the former has an integer precision (INT8) of 16TOPS, which is mainly used in low power consumption fields such as edge computing products and mobile equipment. The latter has an integer precision (INT8) of 640TOPS, and its performance level is close to that of Avida A100.

At the same time, based on the full-stack AI computing infrastructure, industry applications and services built by the rising series of processors and basic software, Huawei is also building the rising computing industry ecology, including the rising computing hardware and software system, partners, and industry applications.

it is understood that ascension ecological partners include five categories: complete machine hardware partners, IHV hardware partners, application software partners, all-in-one solution partners and ecological operation partners. Among them, there are 13 hardware partners, 6 IHV hardware partners and about 1200 software partners. The company has more than 1.8 million AI developers, more than 1200 partners and more than 2500 industry AI solution certifications.