Overview
Cambricon MLU 370 (Siyuan 370) is Cambricon's fourth-generation AI training/inference chip, released in 2021-Q4, 7nm process, 96 INT8 TOPS compute, 48GB HBM2 memory, 614 GB/s bandwidth, TDP 35W (one of the most energy-efficient 7nm data center AI chips in the industry). Paired with NeuWare 1.0 software stack + MindSpore. It is the predecessor to MLU 590 and was Cambricon's flagship product before the MLU 590 (2023).
Key lineage:
- MLU 100 (2018): 16nm, 8GB, 16 TFLOPS — 1st generation
- MLU 270 (2019): 16nm, 16GB, 128 TFLOPS — early training
- MLU 290 (2020): 7nm, 32GB, 256 TFLOPS — 1st 7nm gen
- MLU 370 (2021): 7nm, 48GB HBM2, 96 INT8 TOPS, 35W — this page
- MLU 590 (2023): 7nm, 96GB HBM2, 256 INT8 TOPS, 250W — existing page
- MLU 690 (2025-2026 speculative): 5nm, 192GB HBM3E, 2 PF FP8 — existing page
Core Specifications
| Item | Spec |
|---|
| Architecture | Cambricon MLUv04 (4th generation) |
| Process | TSMC 7nm |
| Compute Cores | 64x Siyuan 4 cores (proprietary ISA) |
| HBM | 48GB HBM2 |
| Memory Bandwidth | 614 GB/s |
| INT8 | 96 TOPS |
| BF16 | 48 TFLOPS |
| FP32 | 24 TFLOPS |
| TDP | 35W (industry's most efficient 7nm data center AI) |
| Form Factor | PCIe Gen4 x16 |
| Interconnect | MLU-Link 200 GB/s |
| Mass Production | 2021-Q4 |
| Unit Price | ~$1,500-2,500 |
vs MLU 290 (2020)
| Metric | MLU 370 (2021) | MLU 290 (2020) | Improvement |
|---|
| Process | 7nm | 7nm | Same |
| HBM | 48GB HBM2 | 32GB HBM2 | +50% |
| Bandwidth | 614 GB/s | 307 GB/s | 2x |
| INT8 | 96 TOPS | 64 TOPS | +50% |
| BF16 | 48 TFLOPS | 32 TFLOPS | +50% |
| TDP | 35W | 50W | -30% |
| Interconnect | 200 GB/s | 100 GB/s | 2x |
| Software | NeuWare 1.0 | NeuWare 0.5 | New gen |
vs Contemporary NVIDIA T4 (2021)
| Metric | Cambricon MLU 370 | NVIDIA T4 | Difference |
|---|
| Process | 7nm | 12nm | MLU 370 newer gen |
| INT8 | 96 TOPS | 130 TOPS | T4 +35% |
| BF16 | 48 TFLOPS | N/A | MLU 370 exclusive |
| TDP | 35W | 70W | MLU 370 -50% |
| Efficiency | 2.74 TOPS/W | 1.86 TOPS/W | MLU 370 +47% |
| Memory | 48GB HBM2 | 16GB GDDR6 | MLU 370 3x |
| Bandwidth | 614 GB/s | 320 GB/s | MLU 370 1.9x |
| Software | NeuWare + MindSpore | CUDA | T4 mature |
MLU 370 killer features: TDP only 35W (50% of T4) + 48GB HBM2 (3x T4) + BF16 support (T4 has no BF16), domestic + energy-efficient + large memory.
Use Cases
- ✅ Domestic AI inference (energy-efficient + localized)
- ✅ Domestic AI training (48GB HBM2 fits moderately large models)
- ✅ Government/SOE AI projects (localization policy mandate)
- ✅ Intelligent computing centers (35W efficient, high rack density)
- ✅ LLaMA 1 13B FP16 inference (48GB HBM2 sufficient)
- ❌ Cutting-edge AI training (FP8 missing)
- ❌ International market (no CUDA compatibility)
- ❌ Very large LLMs (48GB limited)
| Model | Quantization | Performance (tok/s) | Notes |
|---|
| LLaMA 1 7B | FP16 | ~25 tok/s | Mainstream |
| LLaMA 1 13B | FP16 | ~12 tok/s | Full FP16 |
| LLaMA 1 30B | Q4_K_M | ~5 tok/s | Quantized |
| LLaMA 1 65B | Q4_K_M | ~3 tok/s | 70GB slightly exceeds |
| ChatGLM-6B | FP16 | ~30 tok/s | Chinese |
| Stable Diffusion 1.5 | FP16 | 2x vs MLU 290 | Image generation |
48GB HBM2 advantage: Compared to contemporaneous NVIDIA T4 16GB, can fit 13B LLM in full FP16 (26GB slightly small), was the mainstream domestic LLM inference workhorse in 2021-2022.
Software Stack NeuWare 1.0
| Layer | Tool | Notes |
|---|
| AI Frameworks | NeuWare 1.0 | Unified programming platform |
| PyTorch (NeuWare backend) | Auto MLU mapping |
| TensorFlow (NeuWare backend) | Compatible |
| MindSpore | Huawei/CAICT-led, PyTorch compatible |
| Compiler | BANG C/C++ | Cambricon proprietary language |
| Operator Library | CNML | CUDA cuDNN-like (~70% coverage) |
| Quantization | NeuQuant | INT8 automatic |
| Model Zoo | ModelZoo | CV/NLP/LLM |
MLU 370 software maturity: Operator coverage ~70% (vs CUDA 99%+), mainstream LLMs runnable but require manual optimization.
| Item | Details |
|---|
| Company | Cambricon Technologies |
| Founders | Chen Tianshi and Chen Yunji brothers (CAS ICT) |
| Founded | 2016-03 |
| IPO | 2020-07-20 STAR Market (688256) |
| MLU 370 Launch | 2021-Q4 |
| Key Customers | China Mobile, Inspur, Sugon, ByteDance, Zhipu AI |
| National Projects | "East Data West Compute" recommended chip |
Key Timeline
| Date | Event |
|---|
| 2016-03 | Cambricon founded (CAS ICT spinout) |
| 2018-05 | First chip MLU 100 released (16nm) |
| 2020-07-20 | STAR Market IPO (688256) |
| 2020 | MLU 290 (7nm 1st gen) |
| 2021-Q4 | MLU 370 released (this page) |
| 2022 | MLU 370 mass production + customer deployment |
| 2023-Q4 | MLU 590 released (replaces 370) |
| 2025-2026 speculative | MLU 690 released (replaces 590) |
Cambricon Product Line
| Product | Launch | Process | Memory | INT8 | TDP | Status |
|---|
| MLU 370 | 2021-Q4 | 7nm | 48GB HBM2 | 96 TOPS | 35W | In production -> EOL 2023 |
| MLU 590 | 2023-Q4 | 7nm | 96GB HBM2 | 256 TOPS | 250W | Current flagship |
| MLU 690 | 2025-2026 speculative | 5nm | 192GB HBM3E | 4 POPS | 500W | Roadmap |
| MLU 790 (speculative) | 2027 | 3nm | 384GB HBM4 | 8 POPS | 800W | Long-term |
Key Features
- 48GB HBM2: 2021 domestic AI large memory (vs contemporary NVIDIA T4 16GB)
- TDP 35W: Industry's most efficient 7nm data center AI
- Efficiency 2.74 TOPS/W: 1.5x NVIDIA T4
- BF16 support: T4 has no BF16, MLU 370 exclusive
- MindSpore ecosystem: Deep Huawei collaboration
- Weaknesses: Compute below T4, ecosystem ~70% coverage, already EOL
vs Contemporary Domestic AI Chips (2021-2022)
| Metric | Cambricon MLU 370 | Huawei Ascend 310 | Alibaba Hanguang 800 (2021) |
|---|
| Process | 7nm | 12nm | 12nm |
| INT8 | 96 TOPS | 22 TOPS | 820 TOPS |
| TDP | 35W | 8W | 168W |
| Memory | 48GB HBM2 | 8GB LPDDR4 | 32GB HBM2 |
| Bandwidth | 614 GB/s | 25 GB/s | 700 GB/s |
| Target | Training + Inference | Edge | Data center inference |
2021-2022 domestic AI top three: Hanguang 800 strongest compute (820 TOPS), MLU 370 largest memory (48GB), Ascend 310 best efficiency (8W).