Huawei Ascend 910B
Overview
Huawei Ascend 910B is part of the Ascend series of AI training chips, based on the Da Vinci architecture, entering mass production in 2022-2023. Under US export controls, it has become the mainstream NVIDIA alternative in the Chinese market. Equipped with 64GB HBM2e memory and 320 TFLOPS FP16 compute, it has been adopted by major Chinese AI projects at Baidu, Alibaba, Tencent, and others.
Note: The Ascend 910B has multiple versions (B1/B2/B3/B4) with slightly different specs. The core spec table uses the common figures consistent across multiple sources (FP16 320 TFLOPS / INT8 640 TOPS); the per-version compute in the version-difference table below is market-reported and unconfirmed by official specification sheets — base purchasing decisions on actual measurements.
Core Specifications
| Item | Spec |
|---|---|
| Architecture | Da Vinci (Huawei proprietary) |
| Process | TSMC 7nm+ (including SMIC localization efforts) |
| Memory | 64 GB HBM3 |
| Memory Bandwidth | 1,224 GB/s |
| FP32 | 64 TFLOPS |
| FP16 | 320 TFLOPS |
| INT8 | 640 TOPS |
| TDP | 310 W |
| Interconnect | HCCS (Huawei proprietary) |
| Form Factor | PCIe half-height half-length |
Vendor Information
| Item | Details |
|---|---|
| Manufacturer | Huawei Technologies Co., Ltd. |
| Website | https://www.hiascend.com |
| Product Page | https://www.hiascend.com/en/hardware/ascend-910-b |
| CANN Software Stack | https://www.hiascend.com/en/software/cann |
| MindSpore | https://www.mindspore.cn |
| Driver Downloads | https://www.hiascend.com/developer/download |
Key Features
- HCCS High-Speed Interconnect: NVLink alternative
- CANN Heterogeneous Compute Architecture: CUDA-like
- MindSpore Framework: Huawei's proprietary AI framework
- Domestic Ecosystem: Kunpeng + Ascend + MindSpore complete stack
Use Cases
- LLM training (tens of billions of parameters)
- Domestic replacement in the Chinese market
- Government and state-owned enterprise AI projects
- Large model inference
Related Comparisons
- Huawei Ascend 910C - Successor
- NVIDIA A100 - Previous-gen competitor
- NVIDIA H100 - Current mainstream