Huawei Ascend 910
Product Overview
Huawei Ascend 910 is the first cloud training chip of Huawei's Ascend series, released in August 2019 in Shenzhen alongside the edge inference chip Ascend 310, marking the formal establishment of Huawei's "Da Vinci architecture" full-scenario AI compute system.
The Ascend 910 uses the TSMC 7nm+ EUV process (the last advanced node Huawei obtained before sanctions); at launch it claimed to be "the world's most powerful AI processor" with 256 TFLOPS FP16, outperforming the contemporary NVIDIA V100 and matching early A100 levels. It laid the architectural foundation for the entire 910B / 910C / 920 / 950 series.
📌 In many materials, "910A" refers to this first-generation 910 (32 Da Vinci Max cores @ 1.0GHz); it belongs to the 910 family along with the later 910B/C/D. This card describes it uniformly as Ascend 910 (1st gen / 910A).
Core Specifications
| Parameter | Value |
|---|---|
| Release | 2019-08 (Shenzhen) |
| Architecture | Da Vinci Max |
| Process | TSMC 7nm+ EUV |
| AI Cores | 32 Da Vinci Max cores |
| FP32 | Not disclosed |
| FP16 | 256 TFLOPS |
| INT8 | 512 TOPS |
| Memory | 32 GB HBM2 |
| Memory Bandwidth | 1.5 TB/s (896 GB/s, varies by board) |
| TDP | 310 W (official release figures) |
| Interconnect | HCCS (Huawei in-house, 1st generation) |
| Systems | Atlas 300T / Atlas 800 (training servers) |
| Video Decoding | 128-channel full HD H.264/H.265 |
Evolution Comparison with the Later 910 Family
| Model | Release | Process | FP16 | Memory | TDP | Positioning |
|---|---|---|---|---|---|---|
| Ascend 910 (1st gen) | 2019 | TSMC 7nm+ | 256 T | 32GB HBM2 | 310W | First-generation cloud training |
| Ascend 910B | 2021-22 | SMIC 7nm N+2 | 256-414 T | 64GB HBM2e | 310W | Domestic mainstay |
| Ascend 910C | 2025 Q1 | SMIC 7nm N+2 | 800 T | 128GB HBM3 | ~310W | Chiplet dual-die flagship |
| Ascend 920 | 2025 H2 | SMIC 6nm | 900+ T | ~96GB | 400W | Next-generation bandwidth king |
Key Features
- Da Vinci Max architecture: 32-core many-core AI Core design, leading single-die compute density of its 2019 era
- Mixed precision: full FP16 / FP32 / INT8 support, for training and inference
- Origin of the full-stack ecosystem: paired with MindSpore + CANN to build Huawei's full AI stack
- Historic significance: Huawei's only AI chip on an advanced TSMC EUV process (pre-sanctions)
Vendor Information
| Parameter | Value |
|---|---|
| Manufacturer | Huawei Technologies Co., Ltd. (HiSilicon) |
| Website | https://www.hiascend.com |
| Debut | 2019-08 |
| Architecture | Da Vinci Max |
Use Cases
- ✅ Cloud AI training (1st gen, models under 100B parameters)
- ✅ Research HPC platforms
- ✅ Starting point of Huawei's full-stack AI infrastructure
- ❌ No longer the mainstay (superseded by 910B/C)
Related Comparisons
- Huawei Ascend 910B - Domestic improved version
- Huawei Ascend 910C - Chiplet dual-die flagship
- NVIDIA V100 (contemporary benchmark; no standalone spec page yet)
- NVIDIA A100 - Performance-class reference