Huawei Ascend

Huawei Ascend (昇腾) is the artificial-intelligence computing brand of Chinese technology group Huawei Technologies. It covers a family of AI processors designed in-house by Huawei's chip design unit HiSilicon, the accompanying CANN (Compute Architecture for Neural Networks) software stack, and the Atlas line of AI servers, acceleration cards, edge devices and data-center clusters that package Ascend silicon for customers. The brand sits at the heart of Huawei's effort to build a self-sufficient AI computing platform, from chips and boards through system software and developer tooling to full data-center solutions. Ascend was introduced in 2018 as Huawei's dedicated AI chip family, built on Huawei's Da Vinci neural-network architecture. The original Ascend 310 targeted inference at the edge and in the data center, while the Ascend 910 was positioned for high-performance training. Huawei subsequently extended the line with the Ascend 910B and later generations, and it packaged the chips into Atlas 300/800/900 products and CloudEngine-based cluster solutions. The brand is closely tied to Huawei's broader 'all-scenario AI' strategy and to its HarmonyOS and Kunpeng computing families. Commercially, Ascend is sold mainly to enterprises, cloud providers, research institutions and government-related projects, largely in China, where it has become a leading domestic alternative to imported AI accelerators. Huawei Cloud offers Ascend-based compute services, and the company promotes CANN and the MindSpore framework to developers as an alternative software ecosystem. Because Huawei has been subject to U.S. export restrictions since 2019, Ascend has taken on strategic importance as a domestically designed and, increasingly, domestically manufactured AI platform. Huawei does not break out Ascend revenue separately, and the brand is not a standalone legal entity; it is a product line within Huawei's computing business. Its positioning emphasizes full-stack capability — chips, software, servers and cloud — and ecosystem building through developer programs, partner hardware and open-source software contributions.

Huawei's AI chip program grew out of its HiSilicon semiconductor design subsidiary and its broader investment in the Da Vinci neural-network architecture, which Huawei unveiled in 2018. At Huawei Connect in October 2018, the company introduced the Ascend brand with two initial chips: the Ascend 310 for inference and the Ascend 910 for training. Huawei framed Ascend as the hardware foundation of its 'all-scenario AI' strategy, intended to serve devices, edge nodes and cloud data centers from a common architecture. In 2019 Huawei released the Ascend 910 and paired it with the MindSpore open-source AI framework and the CANN operator and compiler stack, aiming to offer customers a complete alternative to established AI software ecosystems. The company also launched the Atlas product family — acceleration modules, cards, edge devices, servers and clusters — to bring Ascend silicon to market in ready-to-deploy form. The Atlas 900 cluster was presented as a large-scale training system. Huawei's placement on the U.S. Entity List in 2019 sharply raised the strategic value of Ascend. Cut off from some foreign chip supplies and contract manufacturing options, Huawei accelerated domestic sourcing and positioned Ascend as the core of a self-reliant Chinese AI computing stack. Subsequent generations, including the Ascend 910B, were adopted in Chinese data centers and offered through Huawei Cloud. Huawei also promoted Ascend-based hardware to partners and developers through ecosystem programs, and Chinese AI model developers increasingly trained and served models on Ascend clusters. By the mid-2020s Ascend had become one of the most prominent domestic AI accelerator brands in China, used in government, telecom, research and enterprise projects. Huawei continued to iterate the roadmap toward newer Ascend generations and expanded the CANN and MindSpore ecosystems, while packaging Ascend into cloud services and large cluster offerings. The brand remains a product line within Huawei rather than a separate company, and Huawei does not disclose Ascend-specific financial results.

• 2023

  Ascend 910B and later generations gain traction in Chinese data centers as a domestic AI accelerator option.

• 2020

  Atlas 900 cluster and Ascend-based cloud services are promoted to enterprise and research customers.

• 2019

  Ascend 910 and the MindSpore framework are released; Huawei is placed on the U.S. Entity List, accelerating domestic AI computing efforts.

• 2019

  The Atlas product family of AI modules, cards, servers and clusters is launched around Ascend silicon.

• 2018

  Huawei unveils the Ascend AI chip brand and Da Vinci architecture, introducing the Ascend 310 and Ascend 910.

Full-stack domestic AI computing platform — chips, software, servers and cloud — positioned as a self-sufficient alternative to imported AI accelerators, primarily for the Chinese market.

• Ascend 310

  Inference-focused AI processor for edge and data-center deployment.

• Ascend 910 / 910B

  Training-oriented AI processors, the flagship of the Ascend line.

• Atlas 900

  Large-scale AI training cluster built from Ascend-based servers.

• CANN

  Huawei's Compute Architecture for Neural Networks software stack for Ascend.

• MindSpore

  Open-source AI framework optimized for Ascend hardware.

• 2026.08.13

  Reported2026.08.13

  BAAI FlagOS completes Day0 multi-chip adaptation for Alibaba's Qwen3.8-2.4T-A95B[1]智源FlagOS完成阿里Qwen3.8-2.4T-A95B的Day0多芯片适配163.com

  On August 13, the Beijing Academy of Artificial Intelligence (BAAI) announced that its Zhongzhi FlagOS community completed Day0 multi-chip adaptation for Alibaba's open-source ultra-large MoE model Qwen3.8-2.4T-A95B, with adaptation, precision alignment and deployment validation done on nine AI chips including T-Head, Nvidia, Moore Threads, Huawei Ascend, MetaX, Kunlunxin, Hygon, Qingwei and Suiyuan, offering BF16, FP8 and INT8 precision versions.

Read more

1. 1智源FlagOS完成阿里Qwen3.8-2.4T-A95B的Day0多芯片适配163.com