US 
What Is an NPU? Neural Processing Units for Edge AI
An NPU is a dedicated AI accelerator built into a processor to run machine-learning inference efficiently. It is the hardware behind on-device AI — now across Intel Core Ultra, AMD Ryzen AI and NVIDIA Jetson.
An NPU — Neural Processing Unit — is a dedicated AI accelerator built into a processor to run machine-learning inference efficiently. It is the hardware behind “on-device AI”, and it is now appearing across embedded platforms: Intel Core Ultra, AMD Ryzen AI and NVIDIA Jetson. This guide explains what an NPU does, how it differs from a CPU and a GPU, and why it matters at the edge.
What an NPU Does
An NPU is optimized for the maths behind neural networks — large numbers of parallel, low-precision multiply-accumulate operations. It runs trained AI models (inference) far more efficiently than a general-purpose CPU, and at much lower power than a GPU. Its performance is usually quoted in TOPS. The NPU does not replace the CPU; it sits alongside it, taking the AI workload while the CPU runs everything else.
NPU vs CPU vs GPU
All three can run AI, but they are good at different things — and at the edge the difference is power.
CPU
General-purpose and flexible; handles control, logic and light inference, but inefficient for large AI models.
GPU
Massively parallel; excellent for AI training and heavy inference, but power-hungry and often overkill at the edge.
NPU
Purpose-built for inference: high AI throughput per watt, ideal for always-on, on-device AI in power- and thermally-constrained systems.
Why NPUs Matter at the Edge
Edge devices often cannot afford a discrete GPU — the power, heat and space are not there. An NPU brings useful AI inference (vision, detection, classification, voice) to a compact, fanless or battery-powered system, and keeps the data on the device rather than sending it to the cloud. For industrial edge AI, an NPU can be the difference between adding AI to an existing platform and needing a separate GPU computer.
Where You Find NPUs
NPUs are now across the embedded landscape: Intel Core Ultra integrates one on the x86 platform; AMD Ryzen AI adds one to Ryzen Embedded; and NVIDIA Jetson pairs an Arm CPU with a GPU and dedicated accelerators for heavier edge AI. Which suits depends on the model size and the throughput you need — a light model runs comfortably on an NPU, while a demanding vision pipeline may still want a Jetson GPU. See our guide to CPU generations and embedded processors.
Adding AI at the Edge?
Tell us the AI model and the throughput you need, and we will recommend the right platform — NPU, GPU or both — for on-device inference.