Nordic Semiconductor has launched the nRF71 series Wi-Fi SoCs, which integrate an Axon NPU to run edge AI inference such as keyword spotting, anomaly detection and image classification directly on-chip without a separate AI processor. The lineup includes the dual-band Wi-Fi 6 nRF7120 and tri-band Wi-Fi 6E nRF7120E, both with Bluetooth LE, Thread and Matter support, and cuts standby power by up to 20x versus the nRF70.
NPU
神经处理单元
Processor unit specialized for accelerating neural network matrix operations, common in on-device inference.
Curated Intelligence
Qualcomm argues that on-device AI is moving from single-purpose features such as imaging and speech transcription toward agentic experiences capable of intent understanding and task orchestration. Its approach pairs an 'on-device first' cloud-edge continuum with Mixture-of-Experts sparsity to cut per-inference compute, backed by NPU hardware and a full-stack software ecosystem.
Synaptics has launched the Astra SR200 series MCUs, integrating Google's Coral NPU and high-bandwidth custom DRAM to enable multimodal edge AI on microcontroller-class hardware. With up to 400 GOPS of compute power, the chip targets next-gen wearables, industrial vision, and battery-powered sensors requiring larger on-device models.
Qualcomm released an early developer preview for Snapdragon X2 Linux, aiming to push support into the mainline kernel to resolve long-standing software fragmentation on ARM laptops. Built on Debian 13 with standard UEFI and systemd-boot, it integrates graphics via Freedreno, Turnip, and Rusticl in Mesa. A key breakthrough is exposing the Hexagon NPU to user-space frameworks through upstream FastRPC drivers, eliminating proprietary daemon dependencies and enabling standardized local AI inference.
Qualcomm unveiled the gFiber Wi-Fi 8 Premium Gateway at Network X 2026, built on the Dragonwing NPro A8 Elite platform. Featuring a Hexagon NPU, it supports quad-band Wi-Fi 8 and 20Gig peer-to-peer wired connectivity, enabling local agentic AI execution to improve privacy and response times.
Huawei officially launched the Kirin 9050 series, featuring its first application of 'Tao's Law' and logic-folding technology to boost performance and efficiency. As the first native AI Phone platform, the Kirin 9050 Pro integrates a new Da Vinci NPU architecture, enabling the industry's first on-device deployment of a 30-billion-parameter Mixture-of-Experts (MoE) large model.
Huawei executive defines the Kirin 9030 series as a "peak" in China's semiconductor industry, citing it as a culmination of collaborative efforts. Launched with the Mate 80 in late 2025, the chip features official performance gains of 25% in CPU, 40% in GPU, and 70% in NPU, highlighting significant advancements in on-device AI compute.
Microsoft's Phi-2 model successfully runs locally on laptops equipped with Intel Meteor Lake chips, demonstrating the capability of lightweight LLMs for edge deployment. This case validates the feasibility of using new-architecture NPUs to accelerate small language model inference, offering a reference for offline AI applications.