Arm China resets its AI push with the Zhouyi X3 [Edge] NPU
Staff reporter, Taipei; Levi Li, DIGITIMES Asia
Thursday 20 November 2025
Arm China, long under industry scrutiny following the leadership dispute involving former CEO Allen Wu, is signaling a return to stability with a symbolically significant AI product launch. The move marks its strongest effort to exit years of internal...
Copilot:
Arm China's Zhouyi X3 NPU is built on a DSP+DSA architecture, but its exact semiconductor process node has not been publicly disclosed. Based on performance claims and deployment targets, it is likely fabricated on a mature or advanced node—possibly 12nm to 7nm class, but not cutting-edge like 5nm or 3nm.
What We Know About Zhouyi X3
- Architecture: DSP (Digital Signal Processor) + DSA (Domain-Specific Architecture), optimized for large model inference at the edge.
- Performance:
- Each cluster delivers 8–80 FP8 TFLOPS.
- Bandwidth per core reaches 256 GB/s, with support for W4A8/W4A16 acceleration modes.
- Integrates hardware decompression (WDC) for compressed weights.
- Software stack: Compass AI SDK supports PyTorch, TensorFlow, ONNX, Hugging Face, and GPTQ quantization.
Node Inference (Speculative)
- Not disclosed: Arm China has not officially stated the process node.
- Edge inference focus: Targets infrastructure, automotive, mobile, and IoT—suggesting power-efficient nodes, likely 12nm, 10nm, or 7nm.
- Not 5nm or below: No mention of EUV, FinFET scaling, or advanced packaging that would imply cutting-edge nodes.
Strategic Context
- Arm China’s “All in AI” strategy aims to anchor domestic GenAI on locally controlled IP, which may favor domestic foundries like SMIC or HuaHong.
- If fabricated domestically, 28nm or 14nm FinFET is more likely due to export controls and tooling limitations.
Let me know if you want to compare Zhouyi X3 against Huawei Ascend NPUs or track SMIC’s node roadmap for domestic AI silicon. |
