Cambricon MLU 690 (2025-2026 Speculative)
:::warning Speculative Content
Specs on this page are based on Cambricon 2024 public statements + Chen Tianshi 2025-Q1 roadmap + industry analyst projections. Cambricon has not officially released complete MLU 690 specs. Official data subject to actual 2025 H2 / 2026 H1 announcement.
:::
Overview
Cambricon MLU 690 is Cambricon's seventh-generation AI training/inference chip, expected to launch 2025 H2 to 2026 H1 (Siyuan 690). Based on 5nm process (TSMC or SMIC localization), 2 PFLOPS FP8 dense compute (2x MLU 590), 192GB HBM3E memory, 5 TB/s bandwidth. Paired with MindSpore 2.0 + Cambricon NeuWare 2.0 software stack.
Strategic significance: Amid competition from NVIDIA B200 / AMD MI355X / Huawei Ascend 920, MLU 690 is Cambricon's key product to reclaim domestic AI leadership. 2025 Cambricon revenue CNY 7.2B, MLU 690 targets CNY 15-20B (2026).
Core Specifications (Speculative)
| Item | Spec |
|---|
| Architecture | Cambricon MLUv07 (7th generation) |
| Process | TSMC 5nm / SMIC 5nm localization (speculative) |
| Chiplets | 2x chiplet |
| HBM | 192GB HBM3E |
| Memory Bandwidth | 5 TB/s |
| FP8 dense | 2 PFLOPS |
| FP16 / BF16 dense | 1 PFLOPS |
| INT8 | 4 POPS |
| TDP | ~500W |
| Form Factor | OAM / PCIe Gen5 x16 |
| Interconnect | MLU-Link 1.2 TB/s (NVLink 5-class) |
| Mass Production | 2025 H2 - 2026 H1 |
| Unit Price (OAM) | ~$8,000-12,000 (speculative) |
vs MLU 590
| Metric | MLU 690 (Speculative) | MLU 590 | Improvement |
|---|
| Process | 5nm | 7nm | New gen |
| HBM | 192GB HBM3E | 96GB HBM2 | 2x |
| Bandwidth | 5 TB/s | 600 GB/s | 8x |
| FP8 dense | 2 PF | N/A (FP16 125 TF) | New |
| FP16 / BF16 | 1 PF | 125 TF | 8x |
| INT8 | 4 POPS | 256 TOPS | 15x |
| Interconnect | MLU-Link 1.2 TB/s | MLU-Link 600 GB/s | 2x |
| TDP | 500W | 250W | 2x |
| Price (speculative) | ~$10K | ~$5K | 2x |
| Software | NeuWare 2.0 + MindSpore 2.0 | NeuWare 1.0 | New gen |
vs NVIDIA B200
| Metric | Cambricon MLU 690 (Speculative) | NVIDIA B200 | Difference |
|---|
| Process | 5nm | TSMC 4N | Comparable |
| Memory | 192GB HBM3E | 192GB HBM3E | Same |
| Bandwidth | 5 TB/s | 8 TB/s | B200 +60% |
| FP8 dense | 2 PF | 4.5 PF sparse | B200 2.25x |
| BF16 dense | 1 PF | 2.25 PF sparse | B200 2.25x |
| FP4 | N/A | 9 PF sparse | B200 exclusive |
| Interconnect | MLU-Link 1.2 TB/s | NVLink 5 1.8 TB/s | B200 1.5x |
| TDP | 500W | 1000W | MLU 690 -50% |
| Software | NeuWare + MindSpore | CUDA | B200 advantage |
| Price (speculative) | ~$10K | ~$30-40K | MLU 690 -75% |
MLU 690 advantages: TDP only 500W (50% of B200) + 25% the price, FP8 and HBM3E same generation as B200, the best B200 alternative under export controls in China.
Cambricon Product Line
| Product | Launch | Process | Memory | FP16 dense | Status |
|---|
| MLU 100 | 2018 | 16nm | 8GB | 16 TF | EOL |
| MLU 270 | 2019 | 16nm | 16GB | 128 TF | EOL |
| MLU 290 | 2020 | 7nm | 32GB | 256 TF | EOL |
| MLU 370 | 2021 | 7nm | 48GB HBM2 | 96 TF | In production |
| MLU 590 | 2023-Q4 | 7nm | 96GB HBM2 | 125 TF | Current flagship |
| MLU 690 | 2025 H2 - 2026 H1 | 5nm | 192GB HBM3E | 1 PF (FP8 2 PF) | Roadmap |
| MLU 790 (speculative) | 2027 | 3nm | 384GB HBM4 | 2.5 PF | Long-term |
NeuWare 2.0 + MindSpore 2.0 Software Stack
| Layer | Tool | Notes |
|---|
| AI Frameworks | MindSpore 2.0 | Huawei/CAICT-led, PyTorch compatible |
| PyTorch (NeuWare backend) | MLU device mapping |
| TensorFlow (NeuWare backend) | Compatible |
| Compiler | BANG C/C++ | Cambricon proprietary language |
| NeuWare Graph Compiler | Graph compilation optimization (MLU 690 new) |
| Operator Library | CNML 2.0 | CUDA cuDNN-like, ~80% operator coverage (vs MLU 590 70%) |
| Quantization | NeuQuant 2.0 | INT8/FP8 automatic |
| Model Zoo | ModelZoo (1000+ models) | CV/NLP/Multimodal |
| Cluster | NeuWare Cluster | 1024 nodes = 8K MLU 690 cards |
MLU 690 software improvements: Operator coverage from 70% -> 80% vs MLU 590, FP8 support, improved PyTorch compatibility, reduced deep learning model migration cost.
| 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 speculative) | ~CNY 500B (+50% YoY) |
| 2025 Revenue | ~CNY 7.2B (+340% YoY) |
| 2026 Revenue Target | ~CNY 15-20B (MLU 690 contributes 50%) |
| Headquarters | Haidian District, Beijing |
| Website | https://www.cambricon.com |
| Key Customers | China Mobile, Inspur, Sugon, ByteDance, Zhipu AI, Alibaba Tongyi, Baidu Wenxin |
| National Policy | "East Data West Compute" recommended chip |
Cambricon vs Huawei Ascend vs Moore Threads
| Dimension | Cambricon MLU 690 (Speculative) | Huawei Ascend 920 | Moore Threads MTT S5000 |
|---|
| Compute | 2 PF FP8 | 900 BF16 TF | 50 BF16 TF |
| Memory | 192GB HBM3E | 96GB HBM2E | 48GB GDDR6 |
| FP8 | Yes | No | No |
| Ecosystem | MindSpore 2.0 | MindSpore + CANN | MUSA |
| Market | General + intelligent computing centers | Data center + gov/enterprise cloud | General + graphics |
| 2025 Revenue | CNY 7.2B | Included in Huawei Cloud | CNY 2.2B |
| Localization | HBM Samsung + domestic CPU/packaging | Domestic (partial) | 60% localized |
MLU 690 advantages: Only domestic chip supporting FP8 + HBM3E same generation as B200 + Cambricon STAR Market listed (strong fundraising capacity).
Key Features
- FP8 2 PF: First domestic chip supporting FP8, catching up to NVIDIA Blackwell
- HBM3E 192GB: Same capacity as B200
- 5nm localization: Possibly using SMIC 5nm (domestic milestone)
- MLU-Link 1.2 TB/s: Competing with NVLink 5
- NeuWare 2.0: 80% operator coverage, better PyTorch compatibility
- Price advantage: ~25% of B200 price
Use Cases
- ✅ Domestic AI training (B200 export control alternative)
- ✅ LLM training (HBM3E 192GB + FP8)
- ✅ FP8 model training (MLU 690 exclusive domestic capability)
- ✅ Government/SOE AI projects (STAR Market listed)
- ✅ Intelligent computing centers ("East Data West Compute" hubs)
- ✅ Internet companies (ByteDance, Zhipu, Alibaba)
- ❌ International market (no CUDA compatibility)
- ❌ FP4 model training (FP8 only)
- ❌ CUDA proprietary workloads
Key Risks
- 5nm localization progress: SMIC 5nm yield unverified, may depend on TSMC
- HBM3E supply: HBM3E from SK Hynix (Korea) / Samsung, may face US sanctions
- Software migration cost: PyTorch model migration requires manual optimization (80% coverage, 20% still needs custom work)
- NVIDIA countermeasures: NVIDIA launches H20 improved versions (already H20 -> H30 -> ongoing)