Skip to main content

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)

ItemSpec
ArchitectureCambricon MLUv07 (7th generation)
ProcessTSMC 5nm / SMIC 5nm localization (speculative)
Chiplets2x chiplet
HBM192GB HBM3E
Memory Bandwidth5 TB/s
FP8 dense2 PFLOPS
FP16 / BF16 dense1 PFLOPS
INT84 POPS
TDP~500W
Form FactorOAM / PCIe Gen5 x16
InterconnectMLU-Link 1.2 TB/s (NVLink 5-class)
Mass Production2025 H2 - 2026 H1
Unit Price (OAM)~$8,000-12,000 (speculative)

vs MLU 590

MetricMLU 690 (Speculative)MLU 590Improvement
Process5nm7nmNew gen
HBM192GB HBM3E96GB HBM22x
Bandwidth5 TB/s600 GB/s8x
FP8 dense2 PFN/A (FP16 125 TF)New
FP16 / BF161 PF125 TF8x
INT84 POPS256 TOPS15x
InterconnectMLU-Link 1.2 TB/sMLU-Link 600 GB/s2x
TDP500W250W2x
Price (speculative)~$10K~$5K2x
SoftwareNeuWare 2.0 + MindSpore 2.0NeuWare 1.0New gen

vs NVIDIA B200

MetricCambricon MLU 690 (Speculative)NVIDIA B200Difference
Process5nmTSMC 4NComparable
Memory192GB HBM3E192GB HBM3ESame
Bandwidth5 TB/s8 TB/sB200 +60%
FP8 dense2 PF4.5 PF sparseB200 2.25x
BF16 dense1 PF2.25 PF sparseB200 2.25x
FP4N/A9 PF sparseB200 exclusive
InterconnectMLU-Link 1.2 TB/sNVLink 5 1.8 TB/sB200 1.5x
TDP500W1000WMLU 690 -50%
SoftwareNeuWare + MindSporeCUDAB200 advantage
Price (speculative)~$10K~$30-40KMLU 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

ProductLaunchProcessMemoryFP16 denseStatus
MLU 100201816nm8GB16 TFEOL
MLU 270201916nm16GB128 TFEOL
MLU 29020207nm32GB256 TFEOL
MLU 37020217nm48GB HBM296 TFIn production
MLU 5902023-Q47nm96GB HBM2125 TFCurrent flagship
MLU 6902025 H2 - 2026 H15nm192GB HBM3E1 PF (FP8 2 PF)Roadmap
MLU 790 (speculative)20273nm384GB HBM42.5 PFLong-term

NeuWare 2.0 + MindSpore 2.0 Software Stack

LayerToolNotes
AI FrameworksMindSpore 2.0Huawei/CAICT-led, PyTorch compatible
PyTorch (NeuWare backend)MLU device mapping
TensorFlow (NeuWare backend)Compatible
CompilerBANG C/C++Cambricon proprietary language
NeuWare Graph CompilerGraph compilation optimization (MLU 690 new)
Operator LibraryCNML 2.0CUDA cuDNN-like, ~80% operator coverage (vs MLU 590 70%)
QuantizationNeuQuant 2.0INT8/FP8 automatic
Model ZooModelZoo (1000+ models)CV/NLP/Multimodal
ClusterNeuWare Cluster1024 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.

Vendor Information

ItemDetails
CompanyCambricon Technologies
FoundersChen Tianshi and Chen Yunji brothers (CAS ICT)
Founded2016-03
IPO2020-07-20 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%)
HeadquartersHaidian District, Beijing
Websitehttps://www.cambricon.com
Key CustomersChina 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

DimensionCambricon MLU 690 (Speculative)Huawei Ascend 920Moore Threads MTT S5000
Compute2 PF FP8900 BF16 TF50 BF16 TF
Memory192GB HBM3E96GB HBM2E48GB GDDR6
FP8YesNoNo
EcosystemMindSpore 2.0MindSpore + CANNMUSA
MarketGeneral + intelligent computing centersData center + gov/enterprise cloudGeneral + graphics
2025 RevenueCNY 7.2BIncluded in Huawei CloudCNY 2.2B
LocalizationHBM Samsung + domestic CPU/packagingDomestic (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)