NPU跑本地LLM：骁龙X、酷睿Ultra实测
每台新笔记本都宣传NPU算力，但NPU跑LLM还很折腾。
Local LLM on NPU 2026: Snapdragon X, Intel Lunar Lake and What Actually Works
NPUs (neural processing units) are in every new laptop, promising efficient local AI. For LLMs specifically, the story in 2026 is mixed: NPUs excel at small, fixed models but remain awkward for general LLM inference. Here is what actually works.
The Three Platforms
1. Qualcomm Snapdragon X Elite/Plus. The most mature LLM story. 45 TOPS NPU, plus strong CPU. Works with:
- Ollama via the official Windows ARM build
- Llama.cpp with QNN backend (community, improving)
- Windows Copilot+ AI features (Recall, live captions)
Real-world: 7B Q4 models run at 10-20 tokens/s using the NPU; the CPU alone gets similar speeds, so gains vary.
2. Intel Core Ultra (Meteor Lake / Lunar Lake). OpenVINO is the official path. Intel's AI Playground app runs small LLMs and image models. Lunar Lake's NPU (up to 48 TOPS) handles 3B-8B models well; larger models spill to CPU.
3. AMD Ryzen AI. XDNA NPU up to 50 TOPS. Support improved through 2025-2026: Ryzen AI Software supports Llama models via ONNX Runtime. Compatibility is narrower than Qualcomm's.
How to Actually Run One
# Snapdragon X: Ollama works out of the box
curl -fsSL https://ollama.com/install.sh | sh
ollama pull llama3.2:3b
ollama run llama3.2:3b
# Intel: use AI Playground or OpenVINO
pip install openvino-genai
# then load a model with the NPU device
Honest Expectations
| Task | NPU result |
|---|---|
| 1-8B chat models | Good, efficient |
| 8B+ models | NPU is not enough; CPU/GPU takes over |
| Coding agents | Not yet practical on NPU |
| Embeddings/vision | Excellent - NPU's sweet spot |
| Continuous 24/7 inference | Excellent - low power |
FAQ
Is NPU faster than the laptop's CPU for LLMs? For small models, roughly comparable; the NPU's real win is power efficiency and freeing the CPU. For large models, neither is enough - use a GPU.
Why does my NPU show 0% usage? Most apps still default to CPU/GPU. Only software built for the NPU (OpenVINO, QNN, ONNX Runtime with NPU EP) uses it.
Should I buy an NPU laptop for local LLMs? Buy it for battery life and AI features; treat local LLM speed as a bonus, not the reason.
