How Huawei’s Ascend Chips Capture 79% of China’s AI Compute Market

In 2022, Huawei’s Ascend (昇腾) AI chips were deployed across 28 Chinese cities, accounting for roughly 79% of the domestic smart‑compute centre market, according to a financial news report.

Huawei Ascend Chip Portfolio

Huawei offers two main Ascend products: the Ascend 310, optimized for inference, and the Ascend 910B, designed for large‑model training. The 910B supports both FP32 and FP16 precision and delivers performance comparable to Nvidia’s A800/A100 accelerators.

Full‑Stack Ascend Compute Ecosystem

The Ascend compute industry chain is built on a full‑stack architecture that includes:

Hardware: embedded modules, board cards, small stations, servers and clusters based on the Da Vinci core.

Software: the heterogeneous CANN (Compute Architecture for Neural Networks) framework, debugging and optimization tools, the MindStudio development suite, and management utilities.

AI Frameworks: MindSpore (ranked among the top‑tier AI frameworks) and compatibility with other popular open‑source frameworks.

Application Enablement: MindX, ModelArts, HiAI and other services that provide end‑to‑end AI solutions.

Huawei Cloud Pangu Large‑Model Platform

Huawei Cloud’s Pangu 3.0 platform leverages Kunpeng CPUs and Ascend AI chips to deliver a cloud‑native AI compute service. It supports a series of foundational large models ranging from 100 billion to 1 trillion parameters, enabling industry‑specific applications in finance, government, manufacturing, mining, meteorology and railways.

HaiGuang DCU – A Competitive GPGPU Solution

HaiGuang’s DCU (Data‑Center Unit) adopts a GPGPU architecture compatible with both CUDA and AMD’s ROCm ecosystems. Released in Q3 2022, the second‑generation “Deep‑Compute II” chip doubles the performance of its predecessor and matches the capabilities of Nvidia’s A100 and AMD’s MI100 in typical AI and high‑performance‑computing workloads. Its ROCm compatibility allows low‑cost migration from CUDA‑based software stacks.

Cambricon (寒武纪) MLU Series

Founded in 2016, Cambricon focuses on AI‑core processors. Its MLU (Machine‑Learning Unit) accelerator cards are fully compatible with Baichuan 2 series models (53B, 13B, 7B) and achieve performance on par with leading international products. In January 2024, Cambricon signed a strategic partnership with HiDream.ai to jointly adapt MLU cards to the “Zhixiang multimodal large model”, confirming parity in both performance and image quality.

Jinghua (景嘉微) Jinghong Series

Jinghua announced the “Jinghong series”, a line of high‑performance AI compute modules and complete systems. These products support mixed‑precision operations (INT8, FP16, FP32, FP64), multi‑card interconnect for scalable compute, and broad compatibility with domestic and foreign CPUs, operating systems and server vendors, dramatically shortening integration cycles for AI frameworks and algorithms.

Conclusion

The Chinese AI compute market is rapidly consolidating around a few domestic vendors. Huawei’s Ascend chips dominate the smart‑compute centre market, while HaiGuang, Cambricon and Jinghua provide competitive alternatives that are increasingly compatible with global ecosystems such as CUDA and ROCm. This multi‑vendor landscape offers Chinese enterprises a range of high‑performance, full‑stack AI solutions capable of supporting next‑generation large‑model applications.