Huawei Ascend 910B vs 寒武纪 思元690(国产 AI 训练/推理芯片): Spec Comparison & Buyer's Guide
In AI infrastructure selection, Huawei Ascend 910B and 寒武纪 思元690(国产 AI 训练/推理芯片) are two accelerators frequently compared. This article contrasts them item by item — architecture, compute, memory, power, and release cadence — to help you quickly judge which fits training or inference workloads.
Spec Comparison Table
| Vendor | Huawei Ascend 910B | 寒武纪 思元690(国产 AI 训练/推理芯片) |
|---|---|---|
| Vendor | Huawei | Cambricon |
| Architecture | Da Vinci | MLUarch |
| Process | TSMC 7nm+ | 5nm -class |
| Release Date | 2022-11-08 | 2025 |
| FP8 Compute | — | — |
| FP16 Compute | 256 TFLOPS | 700+ TFLOPS |
| FP32 Compute | 51.2 TFLOPS | — |
| INT8 Compute | 512 TOPS | 2,800+ TOPS |
| Memory Type | — | — |
| Memory Capacity | 64 GB HBM2e/ HBM3e | 196 GB HBM3 |
| Memory Bandwidth | 600 GB/s/ 1,200 GB/s | 3.35 TB/s |
| TDP Power | 310 W | ~500 W |
Key Differences
- Power: Huawei Ascend 910B has a TDP of 310 W, lower than 寒武纪 思元690(国产 AI 训练/推理芯片)'s ~500 W, friendlier to datacenter PUE and cooling.
- Memory capacity: 寒武纪 思元690(国产 AI 训练/推理芯片) packs 196 GB HBM3, more than Huawei Ascend 910B's 64 GB HBM2e/ HBM3e, more comfortable for single-card hosting of very large models.
Selection Advice
- When chasing extreme single-card compute and a mature toolchain, prioritize Huawei Ascend 910B; if budget, power wall, or local support are hard constraints, 寒武纪 思元690(国产 AI 训练/推理芯片) often fits better. Use this site's AI Compute Card Comparison Tool to validate multiple chips side-by-side before deciding.
FAQ
What are the main differences between Huawei Ascend 910B and 寒武纪 思元690(国产 AI 训练/推理芯片)?
The core difference is architecture and compute density: Huawei Ascend 910B uses Da Vinci, FP8 ~No public FP8 data, memory 64 GB HBM2e/ HBM3e; 寒武纪 思元690(国产 AI 训练/推理芯片) uses MLUarch, FP8 ~No public FP8 data, memory 196 GB HBM3. See the comparison table above.
What is the TDP (power) of Huawei Ascend 910B?
Huawei Ascend 910B has a TDP of 310 W; actual whole-system power also includes board, fans, and PUE.
Which is better for large-model training / inference?
Training values memory capacity, bandwidth, and multi-card interconnect; inference values single-card throughput and power efficiency. Combine the "Key Differences" and "Selection Advice" above with your batch size, model size, and SLA.
How much do Huawei Ascend 910B and 寒武纪 思元690(国产 AI 训练/推理芯片) differ in memory capacity?
Huawei Ascend 910B is 64 GB HBM2e/ HBM3e, 寒武纪 思元690(国产 AI 训练/推理芯片) is 196 GB HBM3; the gap directly affects loadable model size and context length.
Related Pages
- Huawei Ascend 910B
- 寒武纪 思元690(国产 AI 训练/推理芯片)
- AI 算力卡对比工具 — Compare 2–4 chips side-by-side online