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
| Item | Spec |
|---|
| Architecture | Cambricon MLU 5th Gen (MLUv05) |
| Process | TSMC 7nm (with some SMIC localization) |
| HBM | 96 GB HBM2 |
| Memory Bandwidth | 600 GB/s |
| INT8 Compute | 256 TOPS |
| BF16 Compute | 125 TFLOPS |
| FP32 Compute | 62.5 TFLOPS |
| TDP | ~250 W |
| PCIe | PCIe 4.0 x16 |
| Interconnect | MLU-Link (proprietary, NVLink-like) |
| Form Factor | PCIe / OAM module |
| Mass Production | 2023-Q4 |
| Unit Price (OAM) | ~$3,500-5,000 |
vs Previous MLU 370
| Metric | MLU 590 | MLU 370 | Improvement |
|---|
| Process | 7nm | 7nm | Same |
| HBM | 96GB HBM2 | 48GB HBM2 | 2x |
| Bandwidth | 600 GB/s | 307 GB/s | 1.95x |
| INT8 | 256 TOPS | 128 TOPS | 2x |
| BF16 | 125 TFLOPS | 64 TFLOPS | 1.95x |
| Interconnect Bandwidth | MLU-Link 600 GB/s | 200 GB/s | 3x |
| TDP | 250W | 150W | +67% |
| Perf/W | 1.0 TOPS/W | 0.85 TOPS/W | +18% |
Siyuan 590 Training Cluster
| Item | Config |
|---|
| Board | 8x Siyuan 590 OAM |
| Node | 2x Siyuan 590 servers |
| Cluster | 1024 nodes = 8192 cards |
| Total Compute | 1.05 EFLOPS BF16 |
| Total HBM | 786 TB |
| Interconnect | MLU-Link fully connected |
Software Stack
| Layer | Framework/Tool | Notes |
|---|
| AI Frameworks | MindSpore (Huawei/CAICT-led) | PyTorch compatible |
| PyTorch (Cambricon backend) | MLU device mapping |
| TensorFlow (Cambricon backend) | Legacy ecosystem |
| Compiler | BANG C/C++ | Cambricon proprietary language |
| Operator Library | CNML | CUDA cuDNN-like |
| Model Zoo | ModelZoo | CV/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.
| Item | Details |
|---|
| Company | Cambricon Technologies |
| Founders | Chen Tianshi and Chen Yunji brothers (CAS ICT) |
| Founded | 2016-03 |
| IPO | 2020-07-20 STAR Market (688256) |
| Market Cap (2026-05) | ~CNY 320B |
| 2025 Revenue | ~CNY 7.2B (+340% YoY) |
| Headquarters | Haidian District, Beijing |
| Website | https://www.cambricon.com |
| Key Customers | China 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 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
| Dimension | Cambricon MLU 590 | Huawei Ascend 910C |
|---|
| Compute | 125 BF16 TFLOPS | 780 BF16 TFLOPS |
| Memory | 96GB HBM2 | 128GB HBM2E |
| Ecosystem | MindSpore (PyTorch-compatible) | MindSpore + CANN |
| National Support | STAR Market listed | Huawei in-house |
| Market Position | General + intelligent computing centers | Data center + gov/enterprise cloud |
| 2025 Revenue | ~CNY 7.2B | Included in Huawei Cloud |