AI chip · Huawei
Huawei Ascend 910B
Huawei's flagship domestic AI training accelerator, positioned as the primary Chinese alternative to Nvidia H100 after US export controls. Uses HBM3, Da Vinci architecture. Widely deployed at Chinese hyperscalers and government cloud regions.
Powers 2 tracked robots · Deployed in 1 named data center · Part of 1 rack design.
Market position
The Chinese ceiling under US export controls. Fabbed at SMIC N+2 (~7 nm equivalent), sold into ByteDance, Baidu, iFlytek, Tencent and the domestic hyperscalers. 8,192-chip Atlas 900 A3 SuperCluster is the pod-scale endpoint. Ascend 910C/910D are the follow-ons.
Efficiency and power
How much work you get per watt and per dollar, and how much power a full rack draws.
- Perf per watt
- 1.00 FP16 TFLOPS/W400 TFLOPS ÷ 400 W = 1.00
- Cooling
- waterDepends on the rack design
What fits in 64 GB
Which open-source LLMs run on one Huawei Ascend 910B, by precision. Weights only: add roughly 15% headroom for real serving. If a model does not fit at FP16, try INT8 or INT4 (smaller quality trade-off than most people expect).
| Model | Params | FP16 | INT8 | INT4 |
|---|---|---|---|---|
| Llama 3.1 8B dense | 8 B | 16 GB ✓ | 8 GB ✓ | 4 GB ✓ |
| Llama 3.1 70B dense | 70 B | 140 GB ✗ | 70 GB ✗ | 35 GB ✓ |
| Llama 3.1 405B dense | 405 B | 810 GB ✗ | 405 GB ✗ | 203 GB ✗ |
| Llama 3.3 70B dense | 70 B | 140 GB ✗ | 70 GB ✗ | 35 GB ✓ |
| DeepSeek V3 MoE (671B total, 37B active per token) | 671 B | 1342 GB ✗ | 671 GB ✗ | 336 GB ✗ |
| DeepSeek R1 MoE (671B total, 37B active per token) | 671 B | 1342 GB ✗ | 671 GB ✗ | 336 GB ✗ |
| Qwen 2.5 7B dense | 7 B | 14 GB ✓ | 7 GB ✓ | 4 GB ✓ |
| Qwen 2.5 72B dense | 72 B | 144 GB ✗ | 72 GB ✗ | 36 GB ✓ |
| Mixtral 8x7B MoE (46.7B total, 12.9B active per token) | 46.7 B | 93 GB ✗ | 47 GB ✓ | 23 GB ✓ |
| Mixtral 8x22B MoE (141B total, 39B active per token) | 141 B | 282 GB ✗ | 141 GB ✗ | 71 GB ✗ |
| Gemma 2 27B dense | 27 B | 54 GB ✓ | 27 GB ✓ | 14 GB ✓ |
| Command R+ dense | 104 B | 208 GB ✗ | 104 GB ✗ | 52 GB ✓ |
| Kimi K2 MoE (1T total, 32B active per token) | 1000 B | 2000 GB ✗ | 1000 GB ✗ | 500 GB ✗ |
Math: FP16 = params × 2 bytes; INT8 = params × 1 byte; INT4 = params × 0.5 bytes. MoE models sum every expert (full weights on disk), not the per-token active subset.
Common questions
How much does Huawei Ascend 910B cost?
Huawei Ascend 910B doesn't have a public launch list price. Vendors like this one usually price through direct sales rather than a public sheet. See list price per TFLOP for chips that do publish.
How much memory does Huawei Ascend 910B have?
64 GB of HBM2e, running at 1,600 GB/s. Straight from the vendor datasheet. See chips with the most memory for context.
How much power does one Huawei Ascend 910B draw?
400 W at the chip. A full server draws more once you add CPU, memory, networking and cooling: see the rack power number above. Compare to other chips on perf-per-watt.
What runs on Huawei Ascend 910B?
2 robot models we track use this chip, including AgiBot A2 and AgiBot Yuanzheng A2. Full list below.
Which data centers use Huawei Ascend 910B?
1 named data center runs them, including Envision Galaxy Campus. See who has the most Huawei Ascend 910B for the ranked list.
Which open-source LLMs fit on one Huawei Ascend 910B?
Of 13 open-source LLMs we track, 3 fit at FP16, 4 at INT8, and 8 at INT4 (for example Llama 3.1 8B, Llama 3.1 70B and Llama 3.3 70B). Full table above with each model's memory need.
Key facts
- Class
- Compute SoC (AI accelerator)
- Designer
- Huawei
- Safety-critical?
- No (data-center inference / training)
- Record as of
- 2026-09-27
- Most recent source
- 2024-06-01 (across 2 sources on this page)
Specifications (9 fields, click to expand)
Straight from the vendor datasheet. Dense throughput shown first; sparse (2:4) numbers in parentheses where the vendor publishes them. Full datasheet linked below.
- Process node
- SMIC N+2 (~7nm)
- TDP
- 400 W
- Memory
- 64 GB HBM2e
- Memory bandwidth
- 1,600 GB/s
- FP16 (dense)
- 400 TFLOPS
- INT8 (dense)
- 800 TOPS
- Form factor
- OAM (Atlas 300T A2)
- Announced
- 2023-08-01
- Released
- 2023-08-01
Source: vendor datasheet
Generation
Foundry & process
- Fabbed at
- SMIC (all chips SMIC makes →)
- Process node
- SMIC N+2 (~7nm)
Rack designs that use it
Whole-rack systems buyers order in bulk (NVL72, HGX, TPU pods) that ship with Huawei Ascend 910B inside.
- ×8,192Huawei Atlas 900 A3 SuperCluster (Ascend 910B)supercluster
Compare with
Benchmarks
Published performance numbers by workload. Vendor datasheet figures where noted; independent measurements otherwise.
| Workload | Value | Unit | Source | As of |
|---|---|---|---|---|
| peak tflops fp16(Reported spec from Huawei product materials) | 320 | TFLOPS | vendor | 2023-08-01 |
Robots running Huawei Ascend 910B
Every robot model publicly known to ship Huawei Ascend 910B inside it. Ranked by how widely each has been deployed.
| Model | Maker | Form factor | Role | Source | Deployments |
|---|---|---|---|---|---|
| AgiBot A2 | AgiBot | humanoid | onboard humanoid compute (China stack) | inferred | 3 |
| AgiBot Yuanzheng A2 | AgiBot | humanoid | onboard humanoid compute (China stack) | inferred | 1 |
Data centers running Huawei Ascend 910B
Who has the most? →Named data-center campuses with Huawei Ascend 910B on site. Counts shown where the operator has published them; other rows are described in general terms.
- Envision Galaxy Campusas of 2024-06-01partially energizedinferred
Huawei Ascend AI capacity at the 1.5 GW Envision Ulanqab wind-fed AI factory
Chinese AI campuses under export controls default to Huawei Ascend + domestic silicon.
Sources
- Huawei Atlas 900 A3 (Ascend 910B-based)Huawei
- Huawei AscendWikipedia
Adoption rows appear as we confirm each chip-in-robot pairing from a public source. See every chip we track for the full catalog or the Huawei page.