Skip to main content

Cambricon MLU220 (Siyuan 220)

Product Overview​

The Cambricon MLU220 (Siyuan 220) is an AI inference SoC from Cambricon for edge and terminal devices, officially released in November 2019 (the 21st China Hi-Tech Fair), built on the MLUv02 architecture and TSMC 16nm process. Shaped as a USB-drive-sized M.2 accelerator card (MLU220-M.2), it integrates 8 TOPS (INT8) of compute with power of just 8.25W, and is Cambricon's representative edge product in its "cloud-edge-terminal" layout, filling the gap in edge acceleration solutions at the time.

The MLU220 marked Cambricon's full coverage of cloud, edge, and terminal, mainly targeting edge computing scenarios such as smart grid, smart manufacturing, smart rail transit, smart finance, drones, and autonomous driving in-vehicle computing, enabling AI for terminal devices quickly through a standard M.2 interface.

Core Specifications​

ParameterValue
ArchitectureCambricon MLUv02
Process NodeTSMC 16nm
INT8 Compute8 TOPS
INT4 Compute16 TOPS
FP16 / BF16 ComputeNot disclosed (configurable fixed-point computation supported)
FP32 ComputeNot disclosed
Memory Capacity~4 GB (onboard LPDDR4x, estimated from 4096 MiB reported by cnmon)
Memory TypeLPDDR4x (64 bit)
Memory BandwidthNot disclosed
TDP8.25 W (passive cooling)
InterconnectM.2 (PCIe 3.0 ×2)
InterfaceM.2 2280 (B+M key)
CodecH.264 / HEVC(H.265) / VP8 / VP9; JPEG up to 8192×8192
Launch2019-11
Mass Production/AvailabilityIn mass production

⚠️ Specification notes: INT8 8 TOPS / INT4 16 TOPS / 8.25W / M.2 2280 are consistent figures from Cambricon's official website and press releases. The onboard memory capacity was not officially given as a number; about 4GB LPDDR4x is estimated from the 4096 MiB shown by the driver tool cnmon and is marked as an estimate. FP16/FP32 peaks are not disclosed.

Key Features​

  • Ultra-low power: 8.25W with passive cooling, suited to fanless/harsh environment deployments
  • Ultra-compact M.2 form factor: finger-sized, quickly integrated into existing systems via a standard interface
  • Unified edge-cloud software: supports Cambricon NeuWare, compatible with TensorFlow, Caffe, MXNet, and PyTorch
  • Configurable fixed-point computation: 16/8/4-bit configurable for flexible, excellent inference performance
  • Rich I/O: integrated hardware video/image codec units

Vendor Information​

ItemDetails
CompanyCambricon Technologies Corporation Limited
HeadquartersBeijing
Founded2016
IPOSTAR Market 688256

Use Cases​

  • ✅ Edge AI inference (smart grid, smart manufacturing, smart rail transit, smart finance)
  • ✅ Drone / in-vehicle edge computing
  • ✅ AI enablement of low-power terminal devices
  • ✅ Real-time video/image analytics
  • ❌ Large-scale model training (positioned as an edge inference SoC)
  • ❌ High-compute data center training (compute and power constrained)

References​