Skip to main content

Huawei Ascend 920 (Ascend 920)

Overview​

Huawei Ascend 920 (Ascend 920) is the next-generation Ascend chip after the Ascend 910C, entering mass production in H2 2025. Built on a 6nm process (SMIC N+1 / N+2 domestic), it delivers 900+ BF16 TFLOPS compute and 4 Tbps (4,000 GB/s) HBM bandwidth, making it the highest-bandwidth HBM implementation among domestic AI chips.

Ascend 920 is the core of Huawei CloudMatrix 384 / CloudMatrix 384 Ultra and is the flagship of China's domestic AI cluster for 2025-2026.

Core Specifications​

ItemSpecification
ArchitectureDa Vinci v4
Process Node6nm (SMIC N+1 / N+2 domestic)
Chiplet Count2× (dual-die packaging, similar to 910C)
HBM8× HBM3 modules
HBM Capacity~96 GB (estimated)
HBM Bandwidth4 Tbps = 4,000 GB/s
BF16 Compute900+ TFLOPS
FP16 Compute1,800 TFLOPS (estimated)
INT8 Compute3,600 TOPS (estimated)
TDP~400 W
InterconnectHCCS (Huawei Cache Coherence System)
LaunchH2 2025 mass production (roadmap)

📌 4 Tbps bandwidth = highest domestic HBM bandwidth, 25% higher than Ascend 910C's 3.2 Tbps.

Ascend 910C vs Ascend 920 Upgrade Comparison​

MetricAscend 910CAscend 920Improvement
ArchitectureDa Vinci v3Da Vinci v4New generation
Process Node7nm6nmMore advanced
HBM Bandwidth3.2 Tbps4 Tbps1.25×
BF16 Compute780 TFLOPS900+ TFLOPS1.15×
TDP~310 W~400 W+29%
Launch2025-04H2 2025—

Ascend 920 vs NVIDIA H20 (Comparison)​

MetricAscend 920NVIDIA H20
PositioningDomestic substitutionChina-compliant AI chip
Process Node6nmTSMC 4N (partially SMIC domestic after restrictions)
Memory~96 GB96 GB HBM3
Memory Bandwidth4 Tbps4.0 Tbps
BF16 Compute900 TFLOPS296 TFLOPS
BF16 Compute Ratio3×1× (baseline)
InterconnectHCCS 1.2 TbpsNVLink 900 GB/s
SoftwareCANN + MindSporeCUDA (restricted)
Import Compliance✅ Domestic⚠️ US export controls

💡 Ascend 920 significantly leads H20 in BF16 compute (3×), with matching 4 Tbps bandwidth. This is a critical victory for domestic substitution.

CloudMatrix 384 Ultra System (Estimated)​

ItemConfiguration
Chip Count384 Ascend 920
Rack Count16 (12 compute + 4 network)
Total HBM~36 TB (96GB × 384)
InterconnectAll-optical mesh, 8,000+ LPO optical modules
BF16 Compute (system)~345 PFLOPS (estimated 900 × 384)
TDP (system)~150 kW

CloudMatrix 384 Ultra total compute > NVIDIA GB300 NVL72 cluster (GB300 NVL72 single rack ~144 PFLOPS FP8 dense; CloudMatrix 384 Ultra single rack ~22 PFLOPS BF16 dense). CloudMatrix 384 Ultra system-level = 1.5-2× GB200 NVL72.

CANN + MindSpore Software Ecosystem​

  • CANN 8.x (Compute Architecture for Neural Networks): CUDA equivalent
  • MindSpore 2.4+: Huawei's self-developed 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
ScenarioRecommended Configuration
China domestic AI clusterCloudMatrix 384 Ultra (384 × Ascend 920)
Government/finance domestic substitutionAtlas 950 + Ascend 920
Large-model trainingAscend 920 + MindSpore
Inference servingAscend 920 + vLLM Ascend
Scientific computingAscend 920 (FP64 advantage)

Use Cases​

  • ✅ China domestic substitution (government, finance, telecom, energy)
  • ✅ Large-model training (900 BF16 TFLOPS / 4 Tbps)
  • ✅ Ultra-large-scale clusters (CloudMatrix 384 Ultra)
  • ✅ Scientific computing (FP64 advantage)
  • ❌ Export markets (US export controls)
  • ❌ Non-domestic software ecosystem (high migration cost)

Vendor Information​

ItemDetails
VendorHuawei Technologies
LaunchH2 2025 mass production
Process FoundrySMIC (Semiconductor Manufacturing International Corporation, 6nm N+1/N+2)
Product Pagehttps://www.huawei.com/en/products/computing/ascend
Complete SystemAtlas 950 / CloudMatrix 384 Ultra
Software StackCANN + MindSpore