Product Overview
Huawei Ascend 310B is a low-cost / low-power derivative of the Ascend 310 family, positioned for embedded and lightweight edge scenarios. Note that Huawei reused the "310B" name in 2025 for a new-generation edge AI chip, causing two-generation name reuse confusion in market materials; this card clarifies both.
Core Specifications (two-generation comparison)
| Parameter | Early 310B (c. 2021) | 2025 New-Generation 310B |
|---|
| Release | c. 2021 | 2025 (Huawei name reuse) |
| Process | 12nm (presumed) | 7nm (SMIC) |
| Positioning | Embedded / ultra-low power | Medium-load edge inference |
| INT8 Compute | ~16 TOPS | ~120-200 TOPS (third-party sources) |
| FP16 Compute | ~4-8 TFLOPS | Not disclosed |
| TDP | ≤5 W (fanless design) | ~5-15 W (varies by board) |
| Package | Compact SoC / POP | More integrated module |
| Typical Scenarios | IPC, edge boxes, on-device devices | Smart gateways, device-cloud synergy |
⚠️ Naming note: The early 310B and the 2025 new-generation 310B are not the same product; when purchasing or cross-checking materials, be sure to distinguish them by process and compute.
Positioning within the 310 Family
| Model | Positioning | INT8 | Power |
|---|
| Ascend 310 (1st gen) | On-device/edge inference | 16 TOPS | 8W |
| Ascend 310B (early) | Embedded ultra-low power | ~16 TOPS | ≤5W |
| Ascend 310B (2025) | Medium-load edge | ~120-200 TOPS | 5-15W |
| Ascend 310P | Upgraded edge inference | 140 TOPS | 72W |
Key Features
- Extreme low power: the early 310B is ≤5W and supports fanless passive cooling, fitting space-constrained devices
- High integration: SoC / POP packaging, embeddable directly in IPC and edge boxes
- CANN ecosystem: compatible with PyTorch / TensorFlow, AscendCL development
- Two coexisting generations: distinguish the early and 2025 versions by process and compute
Use Cases
- ✅ Smart cameras / IPC
- ✅ Industrial gateways, edge boxes
- ✅ Low-latency device-cloud collaborative inference
- ✅ Battery-powered / fanless devices
External Links