Arm-China Zhouyi NPU

Zhouyi is Arm-China’s family of neural processing units, also called AIPUs, licensed to SoC vendors for on-device AI inference. The NPU does not depend on the host CPU architecture, so it is found in both Arm and RISC-V SoCs. It scales from low-power edge designs up to multi-core clusters for high-performance SoCs.

Key Features

  • Scalability: From single-core edge configurations up to multi-core clusters.
  • CPU Independent: Integrates next to Arm or RISC-V host CPUs.
  • Transformer Support: Zhouyi X2 and later accelerate Transformer-based models, including large language models.
  • Toolchain: The Zhouyi Compass software stack compiles, quantizes and deploys models from TensorFlow, ONNX and PyTorch.
  • Linux Support: Open-source kernel driver, proposed for upstreaming into the Linux accel subsystem in 2024.

Generations

GenerationYearNotes
Zhouyi Z12018Arm-China’s first in-house IP, unveiled at the World Internet Conference in Wuzhen
Zhouyi Z22020Mid-to-high-end applications: security, smart cockpits, ADAS and edge servers. Up to 128 TOPS per SoC
Zhouyi X22023Multi-core clusters and Transformer support. Up to 320 TOPS
Zhouyi X32025DSP + DSA design for large model inference. 8 to 80 FP8 TFLOPS per cluster

SoCs with Zhouyi NPU

SoCNPU Performance
Cix CD8180 / P130 TOPS
SiEngine SE10003.34 TOPS + 3.5 TOPS (two NPU cores, Zhouyi Z2)
Allwinner R329

Software

The Zhouyi Compass software stack is open source and available on Arm-China’s GitHub:

Resources