Skip to main content

Cambricon MLU 590 (China AI Training/Inference)

Overview

Cambricon Technologies is a leading Chinese AI chip company, founded in 2016 (spun out from the Institute of Computing Technology, Chinese Academy of Sciences), with its STAR Market IPO on 2020-07-20 (ticker 688256). The MLU 590 is its latest-generation dual-purpose training and inference AI accelerator: 7nm process, 256 TOPS INT8 compute, 96GB HBM2 memory, 600 GB/s bandwidth. Paired with the MindSpore full-stack AI framework (led by CAICT), key customers include government, state-owned enterprises, and Chinese internet companies.

Strategic position: Under NVIDIA H100/H200 export controls, Cambricon is one of China's national team mainstays for AI domestic replacement (alongside Huawei Ascend and Hygon DCU).

Core Specifications

ItemSpec
ArchitectureCambricon MLU 5th Gen (MLUv05)
ProcessTSMC 7nm (with some SMIC localization)
HBM96 GB HBM2
Memory Bandwidth600 GB/s
INT8 Compute256 TOPS
BF16 Compute125 TFLOPS
FP32 Compute62.5 TFLOPS
TDP~250 W
PCIePCIe 4.0 x16
InterconnectMLU-Link (proprietary, NVLink-like)
Form FactorPCIe / OAM module
Mass Production2023-Q4
Unit Price (OAM)~$3,500-5,000

vs Previous MLU 370

MetricMLU 590MLU 370Improvement
Process7nm7nmSame
HBM96GB HBM248GB HBM22x
Bandwidth600 GB/s307 GB/s1.95x
INT8256 TOPS128 TOPS2x
BF16125 TFLOPS64 TFLOPS1.95x
Interconnect BandwidthMLU-Link 600 GB/s200 GB/s3x
TDP250W150W+67%
Perf/W1.0 TOPS/W0.85 TOPS/W+18%

Siyuan 590 Training Cluster

ItemConfig
Board8x Siyuan 590 OAM
Node2x Siyuan 590 servers
Cluster1024 nodes = 8192 cards
Total Compute1.05 EFLOPS BF16
Total HBM786 TB
InterconnectMLU-Link fully connected

Software Stack

LayerFramework/ToolNotes
AI FrameworksMindSpore (Huawei/CAICT-led)PyTorch compatible
PyTorch (Cambricon backend)MLU device mapping
TensorFlow (Cambricon backend)Legacy ecosystem
CompilerBANG C/C++Cambricon proprietary language
Operator LibraryCNMLCUDA cuDNN-like
Model ZooModelZooCV/NLP/Multimodal

⚠️ Ecosystem limitations: Compared to NVIDIA CUDA + 10 years of software, Cambricon's ecosystem is only 3-4 years old. PyTorch models need conversion, BANG C has a steep learning curve, and model migration cost is relatively high.

Vendor Information

ItemDetails
CompanyCambricon Technologies
FoundersChen Tianshi and Chen Yunji brothers (CAS ICT)
Founded2016-03
IPO2020-07-20 STAR Market (688256)
Market Cap (2026-05)~CNY 320B
2025 Revenue~CNY 7.2B (+340% YoY)
HeadquartersHaidian District, Beijing
Websitehttps://www.cambricon.com
Key CustomersChina Mobile, Inspur, Sugon, ByteDance, Zhipu AI
National Policy"East Data West Compute" recommended chip

Key Features

  • High localization: HBM from Samsung/SK Hynix, domestic packaging (JCET)
  • Siyuan architecture evolution: MLU 100 (2018) -> 270 (2019) -> 290 (2020) -> 370 (2021) -> 590 (2023) -> 690 (2025 speculative)
  • Unified training + inference: Same hardware supports both
  • MindSpore ecosystem binding: Deep collaboration with Huawei (Ascend also uses MindSpore)
  • Multimodal support: CV / NLP / Speech / Multimodal LLM
  • Weakness: No FP8 support (NVIDIA Hopper/Blackwell 2-4x advantage), ecosystem weaker than CUDA

DeepSeek / Zhipu Performance Reference

  • DeepSeek V3 training: Siyuan 590 cluster performance approximately 50-60% of H100 cluster
  • Zhipu GLM-4 inference: Siyuan 590 single card 256 GB/s x 4 = 1 TB/s total bandwidth, 50 tok/s inference speed (FP16 70B)
  • Stable Diffusion XL training: Siyuan 590 approx 80% A100 speed (BF16)

Use Cases

  • ✅ China market LLM training and inference
  • ✅ Government, SOE AI projects (policy-mandated)
  • ✅ Large model inference deployment
  • ✅ Domestic replacement projects
  • ✅ Intelligent computing center construction ("East Data West Compute" hubs)
  • ❌ International market (CUDA ecosystem lock-in)
  • ❌ Cutting-edge frontier model training (FP8 missing)

Cambricon vs Huawei Ascend

DimensionCambricon MLU 590Huawei Ascend 910C
Compute125 BF16 TFLOPS780 BF16 TFLOPS
Memory96GB HBM2128GB HBM2E
EcosystemMindSpore (PyTorch-compatible)MindSpore + CANN
National SupportSTAR Market listedHuawei in-house
Market PositionGeneral + intelligent computing centersData center + gov/enterprise cloud
2025 Revenue~CNY 7.2BIncluded in Huawei Cloud