Skip to main content

Huawei Ascend 920: China's Highest Bandwidth at 4 Tbps + 3× H20 Compute for Domestic Substitution

· 5 min read
Industry Research Team

Huawei Ascend 920 (昇腾 920) entered large-scale mass production in 2025 H2, representing a major breakthrough for Chinese domestic AI chips. This article analyzes its specifications, comparison with NVIDIA H20, the CloudMatrix 384 Ultra system, and its significance for China's AI industry.

Core Specifications

ItemAscend 910CAscend 920Improvement
ArchitectureDa Vinci v3Da Vinci v4New generation
Process7nm6nm (SMIC domestic)More advanced
Chiplets2× (dual die)same
HBM capacity~128 GB~96 GBslight decrease
HBM bandwidth3.2 Tbps4 Tbps1.25×
BF16 compute780 TFLOPS900+ TFLOPS1.15×
FP16 compute1,560 TFLOPS1,800 TFLOPS1.15×
INT8 compute3,120 TOPS3,600 TOPS1.15×
TDP~310 W~400 W+29%
Release date2025-042025 H2

4 Tbps bandwidth = China's highest domestic HBM bandwidth, a 25% improvement over Ascend 910C. The 900+ BF16 TFLOPS compute also surpasses 910C.

Ascend 920 vs NVIDIA H20 (Target Comparison)

NVIDIA H20 is the "compliance" AI chip specifically designed for the Chinese market under U.S. export controls:

MetricAscend 920NVIDIA H20
PositioningDomestic substitutionChina-compliant AI chip
Process6nm (SMIC)TSMC 4N (partially domestic after restrictions)
Memory~96 GB96 GB HBM3
Memory bandwidth4 Tbps4.0 Tbps
BF16 compute900 TFLOPS296 TFLOPS
BF16 compute ratio1× (baseline)
InterconnectHCCS 1.2 TbpsNVLink 900 GB/s
SoftwareCANN + MindSporeCUDA (restricted)
Import compliance✅ Domestic⚠️ U.S. export controls

💡 Ascend 920 significantly leads H20 in BF16 compute (), with 4 Tbps bandwidth on par with H20. This is a key victory for domestic substitution.

CloudMatrix 384 Ultra System

Ascend 920 will be used in the CloudMatrix 384 Ultra supernode system:

ItemConfiguration
Chip count384 Ascend 920 chips
Rack count16 (12 compute + 4 network)
Total HBM~36 TB (96GB × 384)
InterconnectFully optical mesh, 8,000+ LPO optical modules
BF16 compute (system)~345 PFLOPS (estimated 900 × 384)
TDP (system)~150 kW

CloudMatrix 384 Ultra system-level BF16 compute of ~345 PFLOPS2.4× NVIDIA GB200 NVL72 cluster (~144 PF FP8 dense).

Why Ascend 920 Is the Key Victory for Domestic Substitution?

1. First Time Surpassing H20 by 3× in Compute

PeriodDomesticNVIDIA China EditionMultiple
2023910B = 320 TFLOPSH20 = 296 TFLOPS1.08×
2024910B = 320 TFLOPSH20 = 296 TFLOPS1.08×
2025 H1910C = 780 TFLOPSH20 = 296 TFLOPS2.6×
2025 H2920 = 900 TFLOPSH20 = 296 TFLOPS3.0×

Starting from 2025 H2, Chinese domestic AI chip compute stably surpasses H20 by three times.

2. SMIC 6nm Domestic Process

Ascend 920 uses SMIC N+1 / N+2 6nm process:

  • ✅ Fully indigenous and controllable
  • ✅ Not subject to U.S. export controls
  • ⚠️ Yield and cost still lag behind TSMC 4N

3. 4 Tbps — China's Highest Domestic HBM

Ascend 920's 4 Tbps HBM bandwidth:

  • First domestic chip to reach 4 Tbps level (previous max 3.2 Tbps)
  • On par with H20
  • Presumed to use CXMT (ChangXin Memory Technologies) HBM3 or indigenous HBM

4. CANN + MindSpore Software Stack

  • CANN 8.x (Compute Architecture for Neural Networks): analog to CUDA
  • MindSpore 2.4+: Huawei's indigenous AI framework
  • PyTorch 2.3+ MindSpore backend: PyTorch compatible
  • vLLM 0.7+ Ascend backend: low-latency inference
  • ONNX-Runtime Ascend backend: cross-framework inference
  • Atlas 900/950 series servers: OEM complete systems

China Market Deployment Status

Scaled-Up Customers

CustomerApplication
China MobileLarge model training (990M customers)
China TelecomIntelligent customer service + business insights
China UnicomGovernment + industry AI
State GridPower grid scheduling + fault prediction
CNPCExploration + logistics optimization
Major banksRisk control + anti-fraud
Internet companies (Baidu, Alibaba, Tencent)LLM inference

Industry Layout

  • Government: 100% domestic requirement
  • Finance: policy-driven domestic requirement
  • Telecom: fast HBM domestication progress
  • Energy: fast HBM domestication progress
  • Internet: sensitive workloads shifting to domestic
  • Education / Healthcare: gradual domestication

Limitations and Challenges

LimitationImpact
FP8/FP4 supportAscend 920 still BF16/FP16-primary, FP8 optimization in progress
HBM capacity96 GB is below NVIDIA Rubin R200 288 GB / AMD MI400 432 GB
CUDA compatibilityCANN 8 still requires migration; direct CUDA app execution is limited
SMIC 6nm yield10-20% lower yield than TSMC 4N
HBM sourceCXMT HBM production capacity limited
Interconnect bandwidthHCCS 1.2 Tbps far below NVLink 6 (3.5 TB/s)

Comparison with Contemporaneous Domestic Chips

VendorChipBF16 ComputeHBM BandwidthMass Production
HuaweiAscend 920900 TFLOPS4 Tbps2025 H2
HuaweiAscend 910C780 TFLOPS3.2 Tbps2025-04
CambriconSiyuan 590~480 TFLOPS2.4 Tbps2024
Moore ThreadsMTT S5000~250 TFLOPS1.6 Tbps2024
BirenBR104~300 TFLOPS1.6 Tbps2024
IluvatarCoreX Bi-150~200 TFLOPS1.2 Tbps2024

Huawei Ascend 920 maintains a clear lead among Chinese domestic AI chips.

Detailed Product Pages

Summary

Huawei Ascend 920 is a key victory for Chinese AI chips in 2025 H2:

  1. 900+ BF16 TFLOPS = 3× H20 — first time stably surpassing H20 by three times
  2. SMIC 6nm domestic — indigenous and controllable
  3. 4 Tbps — China's highest domestic HBM bandwidth — HBM domestication breakthrough
  4. CloudMatrix 384 Ultra system — single system surpasses GB200 NVL72
  5. CANN + MindSpore — maturing software ecosystem

Starting from 2025 H2, China's AI industry enters a new phase where "domestic chips can independently support large-scale AI applications."