Huawei Ascend 910C
Last reviewed
Jun 1, 2026
Sources
9 citations
Review status
Source-backed
Revision
v1 · 807 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 1, 2026
Sources
9 citations
Review status
Source-backed
Revision
v1 · 807 words
Add missing citations, update stale details, or suggest a clearer explanation.
The Huawei Ascend 910C is a data-center artificial intelligence accelerator developed by Huawei, positioned as China's leading domestic alternative to high-end NVIDIA GPUs that are restricted from sale to Chinese customers under United States export controls. Reported widely through 2024 and 2025, the Ascend 910C is most often described as a dual-die design that packages two of Huawei's earlier Ascend 910B processors together to roughly double performance. Most of its detailed specifications come from industry analysts and reporting rather than full public datasheets, so the figures below should be read as attributed estimates.
The Ascend 910C is significant less for beating NVIDIA on a per-chip basis, which analysts generally agree it does not, and more for giving China a credible path to building large AI training and inference clusters without access to the most advanced Western hardware.
According to analysis by the research firm SemiAnalysis and reporting from outlets including Reuters and the Financial Times, the Ascend 910C is built from two Ascend 910B compute dies in a single package. Estimates place its performance on the order of 780 to 800 teraflops of BF16 compute, with around 128 gigabytes of high-bandwidth memory and memory bandwidth in the range of 3.2 to 3.5 terabytes per second. These figures put a single 910C in the broad neighborhood of an NVIDIA H100 or H800 on some metrics, while trailing on efficiency and software maturity.
The chips are manufactured by the Chinese foundry SMIC on a 7-nanometer-class process. Producing advanced accelerators domestically has been constrained by yield and by limits on access to the most advanced lithography equipment, which has affected how quickly Huawei can supply the parts in volume.
Software is a central challenge. Huawei provides its own compute stack, called CANN, as an alternative to NVIDIA's CUDA. Porting and optimizing AI frameworks for CANN requires extra engineering effort, and the smaller ecosystem has historically been a barrier to adoption even among Chinese developers.
| Property (reported or estimated) | Detail |
|---|---|
| Developer | Huawei |
| Type | Data-center AI accelerator |
| Design | Dual-die, packaging two Ascend 910B dies |
| Compute | About 780 to 800 TFLOPS BF16 (analyst estimate) |
| Memory | About 128 GB HBM (analyst estimate) |
| Memory bandwidth | About 3.2 to 3.5 TB/s (analyst estimate) |
| Manufacturing | SMIC 7-nanometer-class process |
| Software stack | CANN (alternative to CUDA) |
The most discussed deployment of the Ascend 910C is the Huawei CloudMatrix 384, often abbreviated CM384, a rack-scale system that links 384 Ascend 910C accelerators using an all-optical interconnect. SemiAnalysis reported that, at the system level, CloudMatrix 384 can deliver more total compute and memory than NVIDIA's GB200 NVL72 rack, even though each individual Ascend chip is far less efficient, roughly a third as capable per chip, and the system as a whole draws substantially more power.
This is the core of Huawei's strategy: rather than win on per-chip performance or efficiency, scale up the number of chips and the bandwidth between them so that a complete cluster is competitive for training and serving large models. The trade-off is much higher energy consumption, which is more tolerable in China given relatively abundant domestic power.
The Ascend 910C exists because successive rounds of United States export controls, beginning in 2022 and tightened in later years, blocked NVIDIA from selling its most capable data-center GPUs, and then even cut-down versions like the H800 and H20, to Chinese customers. That created strong demand for a domestic substitute, and Huawei became the most prominent supplier.
Reporting through 2025 indicated growing deployment of the Ascend 910C and CloudMatrix systems by Chinese technology companies and AI developers, including for training large models. Analysts continued to note constraints on Huawei's ability to manufacture the chips at the scale required to fully replace restricted NVIDIA hardware, along with the persistent gap in software ecosystem maturity. The chip is frequently cited in debates over whether export controls slow China's AI progress or mainly accelerate its push toward self-sufficiency.