Skip to main content

Huawei Ascend 310P

Product Overview​

Huawei Ascend 310P is a low-power AI processor from HiSilicon built on the Da Vinci architecture for edge computing and on-device inference, an upgraded version of the Ascend 310. It integrates 10 Da Vinci AI Cores, TaiShan CPU cores, and a video codec unit, with peak INT8 compute of 140-176 TOPS and ~70 TFLOPS FP16; typical chip power is only 8W (whole card such as the Atlas 300I Pro: ~70W).

Strategic position: The 310P is Huawei's mainstay edge AI chip, widely deployed in Atlas 300I series inference cards and AI computing boxes (e.g. OrangePi AI Station, running Llama-2-7B locally), and the first choice for domestic edge inference.

Core Specifications​

ParameterValue
ArchitectureDa Vinci
Process12nm
AI Cores10 Da Vinci AI Cores (some versions: 8)
FP16 Compute70 TFLOPS
INT8 Compute140 TOPS (Atlas 300I Pro) / 176 TOPS (peak)
MemoryLPDDR4X (24GB / 48GB / 96GB, varies by board)
Memory TypeLPDDR4X
Memory Bandwidth204.8 GB/s
Video Decoding128-channel 1080P@30fps
TDP70 W (whole card; chip ~8W)
Release2022
Priceapprox. ¥15,000 (Atlas 300I Pro card)

Key Features​

  • Low power, high efficiency: typical 8W chip power, suited to battery-powered devices
  • Large memory bandwidth: 384-bit LPDDR4X, 204.8 GB/s theoretical
  • Video codec: 128-channel 1080P hardware decoding
  • CANN ecosystem: compatible with PyTorch / TensorFlow, AscendCL development

Vendor Information​

ParameterValue
CompanyHuawei Technologies Co., Ltd. (HiSilicon)
Websitehttps://www.hiascend.com
Release2022
ArchitectureDa Vinci

Use Cases​

  • ✅ Smart video analytics (security, industrial inspection)
  • ✅ Edge AI boxes (local LLM inference)
  • ✅ Robots / unmanned devices
  • ✅ Private lightweight inference