What Is an NPU, and Why Is It in Every New CPU?
Cet article n'est disponible qu'en anglais pour le moment.
NPUs went from a niche feature to a standard bullet point on every new CPU's spec sheet in a short span — here's what they actually do.
What an NPU is built specifically for
An NPU (Neural Processing Unit) is a dedicated chip component built specifically to run neural network inference — the "using an already-trained AI model to get a result" step — far more power-efficiently than a CPU or even a GPU doing the same job.
Why not just use the CPU or GPU for this?
- CPUs are general-purpose and flexible but relatively power-inefficient for the specific, repetitive math (matrix multiplication at scale) that neural network inference relies on heavily.
- GPUs are much better suited to this kind of parallel math and are what's used for training AI models and heavy inference workloads, but they draw significantly more power than a purpose-built NPU for smaller, everyday inference tasks.
- NPUs trade general flexibility for efficiency at exactly this one job — running inference on already-trained models — which is why they can do it using a fraction of the power a GPU would need for the same specific task.
Where NPUs actually get used today
- OS-level AI features — some real-time features like background blur/noise suppression in video calls, live captioning, and certain image processing features increasingly run on the NPU rather than the CPU/GPU, specifically to save battery on laptops.
- Creative and productivity software — some photo/video editing tools use NPU acceleration for AI-assisted features (object removal, upscaling) when available.
- On-device AI assistants — running smaller language or vision models locally rather than sending data to a cloud service, partly for privacy and partly for responsiveness.
Why efficiency, not raw speed, is the actual selling point
An NPU generally isn't about doing something a GPU couldn't already do — it's about doing it using dramatically less power, which matters most for battery life on laptops and for enabling "always-on" AI features that would drain a battery quickly if run on the GPU instead.
A simple comparison
| CPU | GPU | NPU | |
|---|---|---|---|
| Flexibility | Highest | High | Narrow (inference-specific) |
| Efficiency for AI inference | Low | Moderate | High |
| Typical use here | General logic | Training, heavy inference, graphics | Everyday, always-on inference tasks |
The practical takeaway
An NPU isn't something you need to actively use or configure — it's a background component that specific software increasingly takes advantage of automatically when present. Its absence isn't a dealbreaker for most current use cases, but its presence is becoming a standard, low-cost inclusion in new CPUs as more everyday software adds NPU-accelerated features.
Questions fréquentes
Is an NPU the same as a GPU?
No — a GPU is a general-purpose parallel processor originally built for graphics, capable of running AI workloads reasonably well. An NPU is purpose-built specifically for the narrower category of neural network inference, trading general flexibility for much better efficiency at that specific job.
Do I actually need an NPU right now?
It depends on whether the software you use has NPU-accelerated features (some OS-level AI features, some creative and productivity apps) — if it doesn't yet, the NPU in your current CPU sits mostly idle, which is common, and not a reason to avoid a CPU that happens to include one.
Does an NPU replace the need for a discrete GPU?
No — they serve different purposes. An NPU handles specific AI inference tasks efficiently; a GPU is still what you need for gaming, video rendering, and other general parallel compute workloads.