Skip to main content

Huawei Ascend 910C

Overview

Huawei Ascend 910C entered mass production in April 2025, featuring a dual-chiplet design packaging two Ascend 910B compute chips together. It delivers 780 BF16 TFLOPS compute (~60% of H100). In the CloudMatrix 384 system, 384 Ascend 910C chips form a 16-rack super system with total BF16 compute exceeding NVIDIA GB200 NVL72.

Core Specifications

ItemSpecification
ArchitectureDa Vinci (dual-chiplet)
Process Node7nm-class (includes SMIC domestic production)
Chiplet Count2× (2× Ascend 910B compute chips)
HBM8× HBM2E modules
BF16 Compute780 TFLOPS (~60% of H100)
Total Memory Capacity~128 GB (dual-chip)
TDP~310 W (dual-chip, ~45% of H100)

CloudMatrix 384 System

ItemConfiguration
Chip Count384 Ascend 910C
Rack Count16 (12 compute + 4 network)
Total HBM~49 TB (128GB × 384)
InterconnectAll-optical mesh, 6,912 LPO optical modules
Total BF16 Compute~300 PFLOPS (system-level)

Vendor Information

ItemDetails
ManufacturerHuawei Technologies Co., Ltd.
Official Websitehttps://www.hiascend.com
CANNhttps://www.hiascend.com/en/software/cann
MindSporehttps://www.mindspore.cn
2025 Plan800,000 units shipped

Key Features

  • Dual-chiplet design: 7nm process stacking
  • No central I/O die: 2 compute chips directly packaged
  • HBM2E × 8: Supplied by Samsung
  • Export control circumvention through CoAsia, Faraday, and other intermediaries

DeepSeek Performance Reference

  • DeepSeek tested Ascend 910C inference speed at 60% of H100 (BF16)
  • Does not support FP8 (NVIDIA Blackwell has a 2× advantage)
  • Primarily applicable to the China market

Use Cases

  • China market LLM training and inference
  • Large-model inference deployment
  • Government and state-owned enterprise AI projects
  • Domestic substitution