NPU

How it works

Most of the computation an AI model does is repeating huge numbers of multiplications and additions. An NPU is built with many of these multiply-add units packed onto the chip, running them all at once. It also often computes with lower-precision numbers (quantized numbers) to increase speed and save power. The key point is that, unlike the general-purpose CPU or the GPU, which started out as a graphics chip, its circuits were designed from the start with only neural network computation in mind.

Why it matters

AI models are now used beyond data centers, in smartphones, laptops, cars, cameras and more. Because battery life and heat make it hard to fit power-hungry chips into these devices, NPUs, which run AI on little power, are becoming a core component of on-device AI. In data centers as well, more companies are building NPUs dedicated to inference to reduce dependence on GPUs and lower costs. This also ties in with the sovereign AI trend, in which countries try to build up their own domestic AI chips.

Things to know

NPUs go by different names at different companies (TPU, AI accelerator, etc.) and their architectures vary, so the term does not refer to a single standard product. Because they are optimized for specific computations, when a new type of model appears, software support may lag behind and the chip may not reach its full performance. That is why development tools and the software ecosystem matter as much as the chip's own performance. Also, training large models from scratch is still mostly done on GPUs, while NPUs are used more for inference, which runs already-trained models.

Example

Recent smartphones and laptops marketed as 'AI PCs' include an NPU alongside the CPU and GPU, handling features such as photo enhancement, speech recognition and real-time translation on the device without an internet connection. One news item on this site reports that an NPU designed in-house by the Korean company Mobilint is going to space aboard a secondary payload satellite on the fifth launch of Nuri, Korea's domestically built space rocket, an attempt to run AI computation even on a satellite with limited power and weight. Setups that mix several kinds of chips are also increasing, as in the report that AMD is pursuing heterogeneous infrastructure combining its CPUs and GPUs with Korean-made NPUs.

Where it came up on this site

• briefing Mobilint NPU goes to space on a secondary payload satellite on Nuri's fifth launch: first satellite to carry an NPU designed in Korea 2026-10-08
• briefing AMD CEO Lisa Su visits Korea: plans to set up an 'AI CoE' in Korea and hire researchers, and to pursue heterogeneous infrastructure combining AMD CPUs and GPUs with Korean-made NPUs (report) 2026-10-07

Related terms

GPU On-device AI Inference Quantization HBM Sovereign AI