Skip to main content

Canaan Kendryte K510

Product Overview​

The Kendryte K510 is the second-generation edge AI chip released by Canaan Inc. in July 2021 (at the World Artificial Intelligence Conference), positioned for the mid-to-high-end edge inference market. Compared with the K210, its bus architecture, IP cores, and video subsystem were comprehensively redesigned, and officially it delivers 3× the K210's performance.

The K510 carries an upgraded KPU2.0 (General Neural Network Engine, GNNE) that concentrates on breaking through the "memory wall" and "performance wall", providing about 3 TOPS of compute; it also integrates a new ISP (2D/3D noise reduction, WDR, fisheye correction), H.264 encoding, and Gigabit Ethernet, targeting drone aerial photography, HD video conferencing, robotics, assisted driving, and industrial/professional cameras.

Core Specifications​

ParameterValue
ArchitectureTri-core 64-bit RISC-V (2×@800MHz + 1×DSP @800MHz) + KPU (GNNE) + FFT/VAD
Process Node28nm
FP16 / BF16 ComputeNot disclosed
INT8 ComputeAbout 3 TOPS (KPU/GNNE peak)
FP32 ComputeNot disclosed
Memory Capacity1.5MB on-chip SRAM (1MB + 512KB) + external LPDDR3/LPDDR4
Memory TypeLPDDR3 / LPDDR4
Memory BandwidthNot disclosed
TDPAbout 2 W (typical)
InterconnectMIPI CSI / DSI, Gigabit Ethernet, USB OTG, I2S/I2C/UART/GPIO
InterfaceSoC onboard / dev board
Launch2021-07
Mass Production/AvailabilityReleased 2021 (dev board ecosystem followed)

Key Features​

  • KPU2.0 at about 3 TOPS: 3× the K210, breaking through the memory/performance walls.
  • New ISP: 2D/3D noise reduction, WDR wide dynamic range, fisheye correction, hardware 3A.
  • Multimedia: H.264 encoding at 1080p60, JPEG codec, MIPI CSI/DSI.
  • Multi-framework: supports TensorFlow, PyTorch, ONNX, TVM, Caffe, and more.
  • Flexible OS: deployable on bare metal, embedded RTOS, or Linux.
  • About 2W typical power: balancing performance against edge power constraints.

Vendor Information​

ItemDetails
CompanyCanaan Inc. (Nasdaq: CAN)
HeadquartersBeijing, China
Founded2013

Use Cases​

  • ✅ Drone HD aerial photography, HD panoramic video conferencing, robotics, assisted driving, industrial/professional cameras, STEAM education
  • ❌ LLM inference, ultra-low-power always-on operation (weaker than the K210's 0.3W)

References​