Cambricon MLU370-X8 (Siyuan 370)
Product Overview
The Cambricon MLU370-X8 is Cambricon's unified training and inference AI accelerator card based on MLUarch03 (third-generation MLU architecture), in a dual-chip Siyuan 370 configuration with a 7nm process. Official peak performance: FP16 96 TFLOPS / BF16 96 TFLOPS / INT8 256 TOPS / FP32 24 TFLOPS, equipped with 48GB LPDDR5 (614.4 GB/s), full-height full-length dual-slot 250W, and aggregated MLU-Link 200 GB/s (bidirectional). It is paired with the NeuWare software stack + MagicMind. In Cambricon's product line it sits above the MLU290 and below the MLU590, and was one of Cambricon's mainstream shipping models before the Siyuan 590 (official commercial availability in 2026-07).
📌 Data correction (2026-09 cross-validation): This page previously misrecorded "INT8 96 TOPS / BF16 48 TFLOPS / TDP 35W / 48GB HBM2". Cambricon's official product page confirms the MLU370-X8: INT8 256 TOPS, INT16 128 TOPS, FP16/BF16 96 TFLOPS each, FP32 24 TFLOPS, 48GB LPDDR5, 614.4 GB/s, 250W; the Siyuan 370 series uses LPDDR5 rather than HBM.
Key lineage:
- MLU 270 (2019): 16nm — early training/inference
- MLU 290 (2020): 7nm, MLU-Link — first 7nm training generation
- MLU370-X8 (released 2021 / mass-produced 2022): 7nm dual-chip, 48GB LPDDR5, INT8 256 TOPS, 250W — this page
- MLU590 (official commercial availability 2026-07): 7nm, 96GB HBM2e — current mainstream
- MLU690 (mass production in early 2026): dual-Die Chiplet, 196GB HBM3 — flagship
Core Specifications
| Item | Parameter |
|---|
| Architecture | Cambricon MLUarch03 (third generation, dual-chip Siyuan 370) |
| Process | TSMC 7nm |
| FP32 | 24 TFLOPS |
| FP16 | 96 TFLOPS |
| BF16 | 96 TFLOPS |
| INT8 | 256 TOPS |
| INT16 | 128 TOPS |
| Memory | 48GB LPDDR5 |
| Memory Bandwidth | 614.4 GB/s |
| TDP | 250 W |
| Form Factor | PCIe Gen4 ×16, full-height full-length dual-slot (passive cooling) |
| Interconnect | Aggregated MLU-Link 200 GB/s (bidirectional, 3.1x PCIe 4.0) |
| Video Codec | 264-channel HEVC full-HD decode / 48-channel encode, up to 8K |
| Release | 2021 (Siyuan 370 launch) |
| Mass Production | 2022 |
| Unit Price | ~¥38,000-42,000 (e-commerce channel reference) |
vs MLU 290 (2020)
| Metric | MLU370-X8 | MLU 290 (2020) | Change |
|---|
| Process | 7nm | 7nm | Same generation |
| Memory | 48GB LPDDR5 | 32GB HBM2 | +50% capacity, switched to LPDDR |
| Bandwidth | 614.4 GB/s | 307 GB/s | 2x |
| INT8 | 256 TOPS | 64 TOPS | 4x |
| FP16/BF16 | 96 TFLOPS | 32 TFLOPS | 3x |
| TDP | 250 W | 50 W | Full-size data center card |
| Interconnect | 200 GB/s | 100 GB/s | 2x |
| Software | NeuWare | NeuWare 0.5 | New generation |
vs NVIDIA T4
| Metric | MLU370-X8 | NVIDIA T4 | Difference |
|---|
| Process | 7nm | 12nm | MLU370 newer generation |
| INT8 | 256 TOPS | 130 TOPS | MLU370 doubles it |
| FP16 | 96 TFLOPS | 65 TFLOPS | MLU370 +48% |
| BF16 | 96 TFLOPS | Not supported | Exclusive to MLU370 |
| TDP | 250W | 70W | T4 is more power-efficient |
| Memory | 48GB LPDDR5 | 16GB GDDR6 | MLU370 3x |
| Bandwidth | 614.4 GB/s | 320 GB/s | MLU370 1.9x |
| Software | NeuWare + MagicMind | CUDA | T4 is mature |
MLU370-X8 positioning: In tests published on its basic software platform SDK, Cambricon officially claims that on a single card the performance of four common AI models is on par with a mainstream 350W RTX GPU; in multi-card scenarios, MLU-Link (200GB/s) delivers better parallel speedup.
Use Cases
- ✅ Unified training and inference (official positioning, full FP32/FP16/BF16/INT8/INT4 coverage)
- ✅ High-density inference + multimodal concurrency (264-channel HEVC full-HD decode)
- ✅ Government/SOE AI projects (domestic substitution policy)
- ✅ Single-server 8-card training / distributed inference (MLU-Link 200GB/s)
- ❌ FP8 (not supported; requires MLU590/690)
- ❌ Cutting-edge large-model training (move to Siyuan 590/690 recommended)
- ❌ International market (no CUDA compatibility)
| 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 |
| ChatGLM-6B | FP16 | ~30 tok/s | Chinese |
| Stable Diffusion 1.5 | FP16 | 2x vs MLU 290 | Image generation |
48GB LPDDR5 advantage: Compared to the contemporaneous NVIDIA T4 with 16GB, it can fully load 13B-class LLM FP16 inference; combined with 264-channel video decoding, it was the mainstream domestic multimodal/vision + mid-scale LLM inference workhorse of 2022-2024.
Software Stack NeuWare
| Layer | Tool | Description |
|---|
| AI Frameworks | NeuWare | Unified programming platform |
| PyTorch (NeuWare backend) | Automatic 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 |
| Inference Engine | MagicMind | Inference acceleration engine |
| Quantization | NeuQuant | INT8 automatic |
| Model Zoo | ModelZoo | CV/NLP/LLM |
| Item | Details |
|---|
| Company | Cambricon Technologies |
| Founders | Chen Tianshi and Chen Yunji brothers (CAS Institute of Computing Technology) |
| Founded | 2016-03 |
| IPO | 2020-07-20 STAR Market (688256) |
| Siyuan 370 Launch | 2021-Q4 (MLU370-X8 mass-produced in 2022) |
| Key Customers | China Mobile, Inspur, Sugon, ByteDance, Zhipu AI |
| National Projects | Recommended chip for the "East Data West Compute" project |
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 (first 7nm generation) |
| 2021-Q4 | Siyuan 370 released (this page's product) |
| 2022 | MLU370-X8 mass production + customer deployment |
| First unveiled 2024 | MLU 590 (official commercial availability 2026-07) |
| Early 2026 | MLU 690 mass production (196GB HBM3) |
Cambricon Product Line
| Product | Launch | Process | Memory | INT8 | TDP | Status |
|---|
| MLU370-X8 | 2021-Q4 / mass-produced 2022 | 7nm dual-chip | 48GB LPDDR5 | 256 TOPS | 250W | In mass production and on sale |
| MLU590 | Unveiled 2024 / commercial 2026-07 | 7nm | 96GB HBM2e | 512 TOPS | 350W | Current mainstream |
| MLU690 | Mass production in early 2026 | Dual-Die Chiplet | 196GB HBM3 | 2400 TOPS (media-reported figures) | 500W | Flagship |
Key Features
- Unified training and inference: official positioning, full FP32/FP16/BF16/INT16/INT8/INT4 precision coverage
- 48GB LPDDR5: large memory among domestic cards of its 2022 generation (vs NVIDIA T4 16GB)
- MLU-Link 200GB/s: 3.1x PCIe 4.0, 8-card full interconnect in a single server
- 264-channel video decoding: a powerful tool for multimodal/vision scenarios
- Weaknesses: no FP8, LPDDR5 bandwidth weaker than HBM, cutting-edge large-model training requires moving to 590/690
vs Contemporary Domestic AI Chips (2021-2022)
| Metric | Cambricon MLU370-X8 | Huawei Ascend 310 | Alibaba Hanguang 800 (2021) |
|---|
| Process | 7nm | 12nm | 12nm |
| INT8 | 256 TOPS | 22 TOPS | 820 TOPS |
| TDP | 250W | 8W | 168W |
| Memory | 48GB LPDDR5 | 8GB LPDDR4 | 32GB HBM2 |
| Bandwidth | 614.4 GB/s | 25 GB/s | 700 GB/s |
| Target | Training + Inference | Edge | Data center inference |
2021-2022 domestic AI top three: Hanguang 800 has the strongest compute (820 TOPS), MLU370-X8 has the most complete ecosystem (unified training/inference + MLU-Link), Ascend 310 has the best efficiency (8W).