On November 13, 2025, Arm Technology (China) Co., Ltd. (hereinafter referred to as “Arm Technology”) held a product launch conference in Shanghai, officially introducing the next-generation NPU IP – “Zhouyi” X3. This product features the latest DSP+DSA architecture specifically designed for large models, balancing CNN and Transformer architectures, and is complemented by the user-friendly “Zhouyi” NPU Compass AI software platform. It aims to provide AI computing cores for four major fields: infrastructure, smart vehicles, mobile terminals, and smart IoT, setting a new benchmark for edge AI computing efficiency and accelerating the large-scale deployment of edge and endpoint AI.
Highlights of the “Zhouyi” X3 NPU IP
Hardware Architecture
The latest DSP+DSA architecture designed specifically for large models
The “Zhouyi” X3 is based on the latest DSP+DSA architecture designed for large models, transitioning from fixed-point to floating-point computation, with a single cluster offering 8-80 FP8 TFLOPS computing power that can be flexibly configured, and a single core bandwidth of up to 256GB/s. It supports essential W4A8/W4A16 computation acceleration modes for running large models on the edge, integrating Arm Technology’s self-developed decompression hardware WDC, which provides an additional approximately 15% equivalent bandwidth after lossless compression of large model weights through hardware decompression, significantly enhancing computing efficiency and density.
At the same time, the “Zhouyi” X3 integrates an AI-specific hardware engine AIFF, paired with a dedicated hardened scheduler, achieving an ultra-low CPU load of just 0.5% and low scheduling latency, allowing the NPU to efficiently and smoothly handle multiple AI tasks in parallel.
Software Ecosystem
Optimizing end-to-end performance for large models
Supporting customer-defined AI development
The “Zhouyi” X3 is equipped with the comprehensive and user-friendly Compass AI software platform, which fully optimizes end-to-end performance for large models. The platform is widely compatible with mainstream AI frameworks such as TensorFlow, ONNX, and PyTorch, featuring industry-leading support for dynamic shapes in large models, and supports mainstream quantization schemes for large models like GPTQ, as well as models from the Hugging Face model library, including LLM, VLM, VLA, and MoE.
The “Zhouyi” NPU Compass AI software platform
At the same time, Arm Technology is actively building an open ecosystem. The Compass AI software platform has successively open-sourced core components such as the Parser, Optimizer, Linux Driver, TVM, and internal IR formats, and offers a rich set of debugging tools to meet developers’ white-box deployment needs. It also supports a more user-friendly DSL operator programming language. In deep development mode, customers can use the Compiler, Debugger, and DSL language within the platform to develop their own custom operators, or create their own model compilers using tools like Parser and Optimizer, greatly enhancing development efficiency. At launch, the “Zhouyi” Compass AI software platform supports over 160 operators and more than 270 models, significantly improving development and deployment efficiency, helping customers quickly realize AI application innovation and implementation.
With deep collaboration between software and hardware, the performance of the “Zhouyi” X3 has significantly improved, with CNN model performance enhanced by 30%-50% compared to the previous generation. Under the same computing power specifications, the AIGC large model capability has increased tenfold, thanks to 16 times the FP16 TFLOPS, four times the computing core bandwidth, and over ten times the performance improvement in Softmax and LayerNorm. Measured multi-core computing power linearity reached 70%-80%. In terms of large model performance, the “Zhouyi” X3 achieved a computing power utilization rate of 72% during the Prefill phase, and an effective bandwidth utilization rate exceeding 100% during the Decode phase with WDC enabled, fully unleashing the potential of hardware computing power.
Application Scenarios
Covering infrastructure, smart vehicles
Mobile terminals, smart IoT across four major fields
In terms of application scenarios, the “Zhouyi” X3 can be widely used in AI devices such as accelerator cards, smart cockpits, ADAS, embodied intelligence, AI PCs, AI smartphones, smart gateways, and smart IPCs, deeply empowering four core fields, enhancing AI computing efficiency, and optimizing AI application experiences.
Infrastructure: In accelerator card scenarios, it supports acceleration for diverse AI model structures such as CNN and large models;
Smart Vehicles: In ADAS systems, it can provide AI computing power support for assisted driving features like automatic parking; in IVI systems, it supports intelligent interaction based on voice and video image inputs from inside and outside the vehicle;
Mobile Terminals: On AI PCs and AI smartphones, it can achieve super-resolution rendering, enhancing resolution, and supports AI Agent applications based on large models;
Smart IoT: In smart gateways, smart IPCs, and other devices, it provides local AI inference capabilities, enhancing response speed and optimizing smart experiences.
The launch of the “Zhouyi” X3 further enhances Arm Technology’s NPU IP product family layout in the edge AI field. Currently, the “Zhouyi” NPU product line comprehensively covers core fields and can meet multi-scenario AI computing needs. In the future, Arm Technology will actively respond to the “AI+” action plan, deepen technological innovation, connect with global cutting-edge technologies, and work with partners to build an ecosystem, laying a solid foundation for the domestic “AI+” industry upgrade.
END
Source | Arm Technology
Editor | Qiao Jun
Recommended Reading
Weekly News | New requirements for integrated vehicle-road-cloud; the industry’s first wireless AI scheduling technology field test; Blaize goes public; Inspur releases new business strategy I3C
Insights | Edge computing defines a new business engine: insights into the data value of typical edge scenarios, with PDF download
Frontier | Academician Cao Jiannong: The future of edge computing: tending towards distributed intelligence
Report | Generative AI application development report (2024), with download
White Paper | Guide to building computing power factories white paper, with PDF download
Essentials | Edge computing blue paper, with PDF download
Report | 212-page PPT! 2024 Global AI Status Panorama Report, with download
Academician Talks | Academician Wu Hequan: Four paths for the implementation of large models
Recommended Reading| Beijing University of Posts and Telecommunications: Computing power network architecture and key technologies for industrial large models
Insights | Top 10 trends in the information and communication industry (ICT) for 2025, with full PPT download
White Paper | China Mobile leads the release of the white paper “Network Computing Enables Edge Network Collaboration”, with download
AI | Annual development trend report of large AI models, with PPT download
Academician Talks | Academician Gao Wen: Seizing the historic opportunity for AI development
Recommended Reading | Edge computing: The forefront and challenges of future technology
Artificial Intelligence | DeepSeek opens the era of AI algorithm revolution, with download
Deepseek Predictions | Analysis of the development trends of the edge computing industry in 2025
