On November 17, Arm Technology officially launched its next-generation NPU IP ‘Zhouyi’ X3 at a product launch event in Shanghai on the 13th of this month. It adopts a hybrid architecture of DSP + DSA and can be widely used in four core areas: infrastructure, smart vehicles, mobile terminals, and smart IoT.
The X3 is designed to balance CNN and Transformer, with a single cluster FP8 performance ranging from 8 to 80 TFLOPS and can be flexibly configured, with a single-core bandwidth of up to 256GB/s, supporting the necessary W4A8 / W4A16 computation acceleration modes for running large models on the edge. Additionally, it integrates Arm’s self-developed WDC decompression hardware, further enhancing the effective bandwidth by 15%.
In terms of software ecosystem, the X3 is equipped with the Compass AI software platform, compatible with mainstream AI frameworks such as TensorFlow, ONNX, PyTorch, supporting mainstream quantization schemes for large models like GPTQ, the Hugging Face model library, and LLM, VLM, VLA, and MoE models.
According to Arm Technology, the FP16 TFLOPS of the X3 is 16 times that of the previous generation, and the computational core bandwidth is 4 times that of the previous generation. The performance of CNN models has improved by 30%-50% compared to the previous generation, and under the same performance specifications, the capabilities of AIGC large models have increased by 10 times.