Skip to main content

Cambricon MLU 370 (2021)

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, 35Wthis 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

ItemSpec
ArchitectureCambricon MLUv04 (4th generation)
ProcessTSMC 7nm
Compute Cores64x Siyuan 4 cores (proprietary ISA)
HBM48GB HBM2
Memory Bandwidth614 GB/s
INT896 TOPS
BF1648 TFLOPS
FP3224 TFLOPS
TDP35W (industry's most efficient 7nm data center AI)
Form FactorPCIe Gen4 x16
InterconnectMLU-Link 200 GB/s
Mass Production2021-Q4
Unit Price~$1,500-2,500

vs MLU 290 (2020)

MetricMLU 370 (2021)MLU 290 (2020)Improvement
Process7nm7nmSame
HBM48GB HBM232GB HBM2+50%
Bandwidth614 GB/s307 GB/s2x
INT896 TOPS64 TOPS+50%
BF1648 TFLOPS32 TFLOPS+50%
TDP35W50W-30%
Interconnect200 GB/s100 GB/s2x
SoftwareNeuWare 1.0NeuWare 0.5New gen

vs Contemporary NVIDIA T4 (2021)

MetricCambricon MLU 370NVIDIA T4Difference
Process7nm12nmMLU 370 newer gen
INT896 TOPS130 TOPST4 +35%
BF1648 TFLOPSN/AMLU 370 exclusive
TDP35W70WMLU 370 -50%
Efficiency2.74 TOPS/W1.86 TOPS/WMLU 370 +47%
Memory48GB HBM216GB GDDR6MLU 370 3x
Bandwidth614 GB/s320 GB/sMLU 370 1.9x
SoftwareNeuWare + MindSporeCUDAT4 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)

LLM Inference Performance (48GB)

ModelQuantizationPerformance (tok/s)Notes
LLaMA 1 7BFP16~25 tok/sMainstream
LLaMA 1 13BFP16~12 tok/sFull FP16
LLaMA 1 30BQ4_K_M~5 tok/sQuantized
LLaMA 1 65BQ4_K_M~3 tok/s70GB slightly exceeds
ChatGLM-6BFP16~30 tok/sChinese
Stable Diffusion 1.5FP162x vs MLU 290Image 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

LayerToolNotes
AI FrameworksNeuWare 1.0Unified programming platform
PyTorch (NeuWare backend)Auto MLU mapping
TensorFlow (NeuWare backend)Compatible
MindSporeHuawei/CAICT-led, PyTorch compatible
CompilerBANG C/C++Cambricon proprietary language
Operator LibraryCNMLCUDA cuDNN-like (~70% coverage)
QuantizationNeuQuantINT8 automatic
Model ZooModelZooCV/NLP/LLM

MLU 370 software maturity: Operator coverage ~70% (vs CUDA 99%+), mainstream LLMs runnable but require manual optimization.

Vendor Information

ItemDetails
CompanyCambricon Technologies
FoundersChen Tianshi and Chen Yunji brothers (CAS ICT)
Founded2016-03
IPO2020-07-20 STAR Market (688256)
MLU 370 Launch2021-Q4
Key CustomersChina Mobile, Inspur, Sugon, ByteDance, Zhipu AI
National Projects"East Data West Compute" recommended chip

Key Timeline

DateEvent
2016-03Cambricon founded (CAS ICT spinout)
2018-05First chip MLU 100 released (16nm)
2020-07-20STAR Market IPO (688256)
2020MLU 290 (7nm 1st gen)
2021-Q4MLU 370 released (this page)
2022MLU 370 mass production + customer deployment
2023-Q4MLU 590 released (replaces 370)
2025-2026 speculativeMLU 690 released (replaces 590)

Cambricon Product Line

ProductLaunchProcessMemoryINT8TDPStatus
MLU 3702021-Q47nm48GB HBM296 TOPS35WIn production -> EOL 2023
MLU 5902023-Q47nm96GB HBM2256 TOPS250WCurrent flagship
MLU 6902025-2026 speculative5nm192GB HBM3E4 POPS500WRoadmap
MLU 790 (speculative)20273nm384GB HBM48 POPS800WLong-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)

MetricCambricon MLU 370Huawei Ascend 310Alibaba Hanguang 800 (2021)
Process7nm12nm12nm
INT896 TOPS22 TOPS820 TOPS
TDP35W8W168W
Memory48GB HBM28GB LPDDR432GB HBM2
Bandwidth614 GB/s25 GB/s700 GB/s
TargetTraining + InferenceEdgeData center inference

2021-2022 domestic AI top three: Hanguang 800 strongest compute (820 TOPS), MLU 370 largest memory (48GB), Ascend 310 best efficiency (8W).