Ascend is a family of artificial intelligence (AI) accelerator processors developed by Huawei Technologies. These chips are designed to handle the intensive computational demands of machine learning and deep learning workloads, particularly in data center environments. The Ascend line includes a range of products from edge devices to high-performance data center cards, aiming to provide a scalable and efficient alternative to other AI accelerators in the market.
The Ascend processors are built on Huawei's proprietary Da Vinci architecture, which is optimized for matrix multiplication and other operations central to neural network inference and training. The architecture integrates multiple AI cores, each capable of performing high-throughput computations with low power consumption. This design allows Ascend chips to support a wide array of AI models, from small embedded systems to massive large language models used in generative AI applications.
History and Development
Huawei first unveiled the Ascend series in October 2018 at its annual Huawei Connect conference in Shanghai. The initial announcement introduced two product lines: the Ascend 310 for edge computing and the Ascend 910 for data center training. The Ascend 910 was positioned as a high-performance training chip, with Huawei claiming it could deliver up to 256 teraflops (TFLOPS) of half-precision (FP16) computing power, making it competitive with offerings from NVIDIA at the time.
Since then, Huawei has expanded the Ascend family with newer generations. In 2022, the company released the Ascend 910B, an enhanced version with improved memory bandwidth and interconnect capabilities. The Ascend 910C followed in 2024, further boosting performance and efficiency. These chips are integrated into Huawei's Atlas series of AI servers and cloud services, which are marketed to enterprises and government agencies, particularly in China.
The development of Ascend is closely tied to Huawei's broader strategy of achieving technological self-sufficiency, especially after the company faced sanctions from the United States in 2019 that restricted its access to advanced semiconductor manufacturing and certain software. As a result, Huawei has invested heavily in domestic chip design and manufacturing partnerships, including with TSMC for early production and later with SMIC for more recent chips, though specific manufacturing details are often not publicly disclosed.
Architecture and Technical Specifications
The Da Vinci architecture used in Ascend processors is designed around a unified memory hierarchy and a scalable core design. Each AI core contains multiple vector units and a matrix engine, which can execute operations like convolution and matrix multiplication efficiently. The architecture supports mixed precision computing, including FP16, BF16, and INT8, allowing developers to balance accuracy and speed.
For the Ascend 910 series, key specifications include:
- Compute power: Up to 256 TFLOPS (FP16) for the original 910, with the 910B and 910C offering incremental improvements.
- Memory: High-bandwidth memory (HBM) with capacities varying by model, typically 32GB or more.
- Interconnect: Huawei's own CANN (Compute Architecture for Neural Networks) software stack and a high-speed interconnect called HCCS (Huawei Cache Coherence System) for multi-chip communication.
The Ascend 310, designed for edge inference, consumes only 8 watts and delivers up to 8 TFLOPS (FP16), making it suitable for devices like smart cameras and industrial robots. This edge chip supports on-device AI processing, reducing latency and bandwidth requirements.
Software Ecosystem
Ascend processors rely on the CANN software stack, which includes a compiler, runtime, and libraries optimized for the Da Vinci architecture. CANN supports popular deep learning frameworks such as TensorFlow, PyTorch, and MindSpore (Huawei's own framework). Developers can write models in these frameworks and then use CANN to deploy them on Ascend hardware with minimal code changes.
Huawei also provides the MindSpore framework, which is tightly integrated with Ascend and offers features like automatic parallelization and model compression. The combination of MindSpore and Ascend is marketed as a full-stack AI solution, from training to inference, and is used in various sectors including telecommunications, finance, and healthcare.
Despite the software support, the ecosystem is less mature than that of competitors like NVIDIA with its CUDA platform. This has been a barrier to broader adoption outside China, as many AI researchers and companies have built their workflows around CUDA. To mitigate this, Huawei has worked on compatibility layers and tools to ease migration, but the process remains non-trivial.
Applications and Use Cases
Ascend processors are deployed across a range of applications, primarily in data centers for training and inference of large language models and other deep learning models. In China, Ascend chips power several major AI initiatives, including the Pangu model series developed by Huawei Cloud, which covers natural language processing, computer vision, and scientific computing tasks.
Other use cases include:
- Smart cities: Video analysis and facial recognition systems.
- Autonomous driving: Inference for perception models in vehicles, though this is less prominent than NVIDIA's offerings.
- Scientific research: Simulations and data processing in fields like weather forecasting and drug discovery.
- Edge computing: Real-time AI on industrial IoT devices, such as predictive maintenance and quality inspection.
Huawei has also partnered with various Chinese universities and research institutes to promote Ascend adoption, offering hardware grants and training programs. This has helped build a domestic ecosystem that reduces reliance on foreign chips.
Market Position and Competition
The AI accelerator market is dominated by NVIDIA, which holds a majority share due to its powerful GPUs and mature software stack. Other competitors include AMD with its Instinct line, Intel with Gaudi accelerators, and specialized startups like Groq and SambaNova. In China, domestic rivals such as Baidu's Kunlun and Alibaba Cloud's Hanguang chips also compete.
Ascend's market position is strongest in China, where government policies encourage domestic technology adoption. The US sanctions have accelerated this trend, as Chinese companies seek alternatives to NVIDIA products, which are restricted from export. As a result, Ascend has seen increased demand, with Huawei reporting that its Ascend server revenue grew significantly in 2023 and 2024.
However, Ascend faces challenges in performance and software compatibility compared to NVIDIA's latest offerings. For instance, the Ascend 910C is estimated to have performance comparable to NVIDIA's A100, but not the newer H100 or H200. Additionally, the software ecosystem is less extensive, which can limit developer productivity.
Future Directions
Huawei continues to invest in Ascend, with plans for more powerful chips and broader ecosystem support. The company is also exploring advanced packaging and chiplet designs to improve performance and yield. In 2024, reports indicated that Huawei was working on the Ascend 930, which could feature even higher compute density and memory bandwidth, though official specifications have not been released.
There is also a push to integrate Ascend with quantum computing research, though this is in early stages. Additionally, Huawei is expanding its cloud services, offering Ascend-based instances on Huawei Cloud, which competes with Amazon Web Services and Google Cloud in the AI infrastructure space.
The long-term success of Ascend will depend on Huawei's ability to overcome manufacturing constraints and build a robust software ecosystem. As of 2025, the chip remains a significant player in the Chinese market, but its global impact is limited by geopolitical factors.
Conclusion
Ascend represents Huawei's ambitious attempt to create a competitive AI processor line that can rival established players. With its Da Vinci architecture, growing software support, and strategic importance in China, it has carved out a niche in the data center AI market. While it faces hurdles in performance and ecosystem maturity, ongoing development and government backing suggest it will remain a key player in the evolving landscape of AI hardware.