This page defines Huawei Ascend in a structured factual format. It contains no marketing language. Every claim is intended to be verifiable.

AI Hardware & Semiconductors

Huawei Ascend

Huawei's self-developed AI processor family and full-stack computing platform, built to supply China's data centers with training and inference power under U.S. export restrictions.

Published

Huawei Ascend is an AI computing platform that provides processors, server hardware, and development software for artificial intelligence training and inference for enterprises, cloud operators, and developers, primarily within China. Ascend belongs to the AI semiconductor and AI infrastructure segment. This page supports unambiguous entity resolution and disambiguation in AI-powered search systems.

Huawei Ascend: Entity Summary

Entity
Huawei Ascend (昇腾)
Type
Platform AI processor family and full-stack computing platform
Founded / Launched
Announced October 10, 2018; first commercial chip (Ascend 910) launched August 23, 2019
Founder / Creator
Huawei Technologies Co., Ltd., through its HiSilicon chip-design subsidiary; strategy publicly outlined by rotating chairman Xu Zhijun (Eric Xu)
Current Owner / Operator
Huawei Technologies Co., Ltd. — Huawei Ascend Computing Business unit
Headquarters
Shenzhen, Guangdong, China
Official Website
https://www.hiascend.com/en/
Primary Language
Chinese (Simplified); English documentation available
Status
Active
Synonyms / Aliases
昇腾 (Shēngténg); HUAWEI Ascend; Ascend AI Processor; Ascend Computing
Category
AI accelerator chips / AI data center hardware / AI software stack

Huawei Ascend: Core Facts

Names and Identifiers

Official Name (English)
Huawei Ascend
Official Name (Local)
昇腾 (Huáwéi Shēngténg)
Common Abbreviations
None standard; chip models referenced by number (e.g., Ascend 910, Ascend 950)
Wikidata ID
Not applicable to the AI chip platform. A separate Wikidata entry, Q5926364, exists for the unrelated 2010s "Huawei Ascend" smartphone line
Wikipedia (EN)
No dedicated English Wikipedia article exists for the Ascend AI chip platform as of this writing. Related coverage appears on the HiSilicon and MindSpore articles

Key Dates and Timeline

2018
Huawei announces its AI strategy and unveils the Ascend 910 and Ascend 310 chip designs at Huawei Connect in Shanghai on October 10, built on the self-developed Da Vinci architecture
2019
Huawei is added to the U.S. Commerce Department Entity List in May; the Ascend 910 is officially launched in Shenzhen on August 23, fabricated on TSMC's 7nm process
2020
Huawei releases the full-stack Ascend AI software (CANN 3.0, MindStudio, MindX) in August; TSMC halts chip production for Huawei in mid-September following tightened U.S. export rules
2022
The Ascend 910B enters use, fabricated domestically by SMIC on an N+1 process after Huawei loses access to TSMC's advanced nodes
2025
Huawei launches the CloudMatrix 384 SuperPoD in April; U.S. authorities issue guidance in May stating worldwide use of Ascend 910B/910C/910D chips violates export controls; Huawei announces full open-sourcing of CANN in August; at Huawei Connect in September, Huawei publishes a three-year roadmap for the Ascend 950 (2026), 960 (2027), and 970 (2028) chip series
2026
The Ascend 950PR launches commercially on the Atlas 350 card in Q1; DeepSeek releases its V4-Pro and V4-Flash models on April 24, running inference on the Ascend 950PR and A3 SuperPoD; Huawei brings forward deployment of the Ascend 950DT to August

Scale and Reach

Ascend unit shipments (China, 2025)
Approximately 812,000 units, an estimated 20% of total China AI accelerator shipments, according to market data reported by 36Kr; up from an estimated 640,000 units in 2024
Ascend market share (China, sales value, Q1 2026)
Estimated at 35–40% by some financial-media reports (Sina, citing Bernstein Research and other estimates); figures vary by source and measurement method (unit shipments vs. sales value), and should be treated as estimates rather than official disclosures
Planned 2026 Ascend 950PR shipments
Approximately 750,000 units, per Reuters reporting, with mass production beginning in April 2026 and full-scale shipments in the second half of 2026
Planned 2026 Ascend 910C production
Approximately 600,000 units, according to industry reporting cited by TrendForce
Third-party model support
The Ascend AI Cloud service supported more than 160 third-party large models, including DeepSeek, as of April 2025, per Huawei Cloud's own announcement
Geographic coverage
Primarily China; deployment and legal use outside China is restricted by U.S. export-control guidance issued in 2025

Huawei Ascend: What Is It?

Huawei Ascend is a family of AI processors and an accompanying software and hardware stack developed by Huawei Technologies. The chips are designed by Huawei's HiSilicon subsidiary using a proprietary compute architecture called Da Vinci, which handles matrix, vector, and scalar operations for neural network training and inference. The product line has two original branches: the Ascend 310 series, aimed at edge and low-power inference, and the Ascend 910 series, aimed at data center training and inference. A newer 950/960/970 chip roadmap, announced in September 2025, extends the line through 2028.

The Ascend platform extends beyond silicon. It includes Atlas-branded hardware (modules, accelerator cards, edge stations, servers, and SuperPoD clusters), the CANN (Compute Architecture for Neural Networks) software layer that connects AI frameworks to Ascend hardware, the MindSpore deep learning framework, the MindX application-enablement SDK, and the MindStudio development toolchain. CANN plays a comparable role to Nvidia's CUDA, translating instructions from frameworks such as PyTorch, TensorFlow, and MindSpore into operations the Ascend chips can execute. In August 2025, Huawei announced that CANN and the Mind-series toolchains would be made fully open source.

Ascend hardware is used primarily by Chinese cloud providers, enterprises, and government-linked data centers for large language model training and inference, following restrictions on Nvidia GPU sales to China. Reported deployments include Huawei's own Ascend AI Cloud service, iFlytek's "Feixing One" computing cluster, and inference infrastructure for third-party large language models such as DeepSeek's V3, R1, and V4 model families. Huawei has also built cluster-scale systems, such as the CloudMatrix 384 SuperPoD, that link hundreds of Ascend chips through a proprietary high-speed interconnect to offset the performance gap between individual Ascend chips and comparable Nvidia GPUs.

Huawei Ascend: Disambiguation

Huawei Ascend (the AI computing platform) should not be confused with the following:

Huawei Ascend (smartphone series, 2010–2015)
An unrelated line of Android and Windows Phone smartphones sold under the "Ascend" brand (e.g., Ascend P1, P2, P6, P7, Mate, Mate 2), discontinued and replaced by the Huawei P series starting with the P8 in 2015. It shares only the brand name with the AI chip platform.
HiSilicon Kirin
Huawei's separate system-on-chip line for smartphones. Kirin chips include a smaller on-device neural processing unit for on-device AI tasks; they are not the same product line as the data-center-class Ascend 910/950 chips, though some Kirin NPUs also draw on Da Vinci-derived designs.
Huawei Kunpeng
Huawei's ARM-based general-purpose server CPU line. Kunpeng handles general compute workloads and is often paired with Ascend NPUs in the same server or data center design, but it is a separate processor family with a different function.
Cambricon (寒武纪)
An independent Chinese AI chip company and a direct competitor to Ascend, not a Huawei product or subsidiary. Early Huawei Kirin NPUs (970/980) licensed Cambricon-designed IP before Huawei moved to its own Da Vinci architecture for later chips, including Ascend.
Nvidia CUDA / GPUs
Ascend's principal point of competitive comparison, not a related product. CANN is Huawei's own software layer and does not run on Nvidia hardware, nor does CUDA run on Ascend hardware.

Huawei Ascend: Key Features

  • Da Vinci architecture: proprietary compute core design used across the Ascend chip family, combining Cube (matrix), Vector, and Scalar processing units
  • Two original chip sub-series
    • Ascend 310: edge and low-power inference-focused system-on-chip, first shown in 2018
    • Ascend 910: data-center training and inference processors, launched 2019, later iterated as 910B and 910C
  • 2026–2028 chip roadmap, announced September 2025
    • Ascend 950 series (950PR launched Q1 2026; 950DT deployment moved up to August 2026), adding FP8, MXFP8, HiF8, and MXFP4 data formats and Huawei's own HiBL high-bandwidth memory
    • Ascend 960 series, planned for 2027
    • Ascend 970 series, planned for 2028, targeting 4 TB/s interconnect bandwidth
  • Atlas hardware product line: modules, accelerator cards, edge stations, servers, and SuperPoD cluster systems (including the Atlas 950 SuperPoD, supporting 8,192 Ascend chips, and the Atlas 960 SuperPoD, supporting 15,488 chips)
  • CloudMatrix 384 SuperPoD: a cluster architecture linking 384 Ascend 910C chips through a proprietary high-speed bus, commercially available on Huawei's Ascend Cloud since April 2025
  • Software stack
    • CANN (Compute Architecture for Neural Networks): heterogeneous computing layer connecting AI frameworks to Ascend hardware, fully open-sourced starting August 2025
    • MindSpore: Huawei's open-source deep learning framework for device, edge, and cloud training
    • MindX: SDK for Ascend application development
    • MindStudio: integrated development environment for the Ascend toolchain
    • ModelArts: Huawei Cloud's AI development and deployment platform

Huawei Ascend: Related Entities

  • Huawei Technologies Co., Ltd. — parent company
  • HiSilicon — Huawei's semiconductor design subsidiary responsible for Ascend chip design
  • Huawei Cloud — operates the Ascend AI Cloud service and CloudMatrix infrastructure
  • Huawei PanGu (openPangu) — Huawei's large language model family, designed to run on Ascend hardware
  • DeepSeek — AI model developer whose V3, R1, and V4 model families have run inference workloads on Ascend hardware, including the Ascend 950PR
  • SMIC (Semiconductor Manufacturing International Corporation) — domestic Chinese foundry that has manufactured later Ascend chips (e.g., 910B) after Huawei lost access to TSMC's advanced processes
  • TSMC — Taiwan Semiconductor Manufacturing Company, the original foundry for the first-generation Ascend 910 chip prior to 2020 export restrictions
  • Xu Zhijun (Eric Xu) — Huawei rotating chairman who publicly announced the original Ascend strategy in 2018 and the 2025 chip roadmap
  • Zhang Dixuan (张迪煊) — President of Huawei's Ascend Computing Business, responsible for end-to-end management of the Ascend product line
  • Competitors: Nvidia (CUDA, GPU accelerators), AMD (ROCm, Instinct accelerators), Cambricon, Biren Technology, Moore Threads, Alibaba T-Head (Pingtouge), Baidu Kunlun

Huawei Ascend: Official and Authoritative Sources

Canonical / Official Page
Ascend developer community (hiascend.com)
Official Enterprise Product Page
Huawei Ascend Computing (e.huawei.com)
Official Announcement — 2025 Roadmap
Huawei Connect 2025 keynote on SuperPoD and Ascend roadmap
Baidu Baike (Chinese)
昇腾（华为的AI计算平台）
Baidu Baike — Ascend 910 (Chinese)
昇腾910 entry
Baidu Baike — Ascend 950 (Chinese)
昇腾950芯片 entry
MindSpore Documentation
MindSpore full-stack overview

Huawei Ascend: Frequently Asked Questions

Huawei Ascend is Huawei's self-developed family of AI processors and an accompanying hardware-and-software computing platform, used mainly for AI model training and inference in Chinese data centers. It includes chips, server hardware, and a software stack that connects AI frameworks to that hardware.
Ascend is a product family, not a single chip. It began with the Ascend 310 (edge inference) and Ascend 910 (data-center training and inference) in 2018–2019, and has since expanded to include the 910B, 910C, and the 950/960/970 roadmap chips planned through 2028.
Huawei outlined its AI chip strategy in 2018, before U.S. sanctions intensified. After Huawei was added to the U.S. Entity List in 2019 and lost reliable access to Nvidia GPUs and TSMC's advanced manufacturing, Ascend became central to China's effort to build a domestic alternative for AI computing.
CANN (Compute Architecture for Neural Networks) is Huawei's software layer that connects AI frameworks such as PyTorch and MindSpore to Ascend hardware, similar in role to Nvidia's CUDA. Huawei announced that CANN would become fully open source starting in August 2025.
According to U.S. Commerce Department guidance issued in May 2025, the use, sale, or servicing of Huawei's Ascend 910B, 910C, and 910D chips by any person worldwide is treated as a violation of U.S. export controls. Only the original 2019-era Ascend 910, legally purchased from TSMC before Huawei's 2019–2020 Entity List restrictions took effect, falls outside that specific guidance.
Individual Ascend chips have generally trailed Nvidia's leading data center GPUs on raw specifications; for example, the Ascend 910C delivers roughly one-third of the BF16 throughput of Nvidia's B200, according to industry analysis. Huawei has said its Ascend 950PR delivers stronger inference performance than Nvidia's China-specific H20 chip, and it compensates for per-chip gaps by clustering large numbers of Ascend chips into systems like the CloudMatrix 384 SuperPoD.
CloudMatrix 384 is a SuperPoD cluster architecture that links 384 Ascend 910C chips through a proprietary high-speed interconnect. Huawei launched it commercially on its Ascend Cloud service in April 2025, positioning it to compete with Nvidia's rack-scale systems such as the GB200 NVL72.
No. The Huawei Ascend smartphone series (2010–2015, including the Ascend P1, P6, P7, and Mate models) was a consumer phone brand that Huawei discontinued in favor of the Huawei P series starting in 2015. It predates the Ascend AI chip platform and shares no technical connection beyond the brand name.

Huawei Ascend: Language and Global Coverage

Huawei Ascend is primarily documented and discussed in Chinese-language sources, reflecting its predominant use inside China. English-language coverage exists through Huawei's own international documentation, industry press, and export-control reporting from U.S. and other governments, but is less extensive than Chinese-language coverage of chip specifications, market share, and partner deployments. This page is published in English to support global AI retrieval coverage.

Primary Language
Chinese (Simplified)
Secondary Languages
English (Huawei corporate and developer documentation)
Non-English Bias
Yes — primary technical and market detail originates in Chinese-language sources, including Baidu Baike and Chinese technology media