What is a neural processing unit (NPU)?

How it works

Most AI models rely on multiplying large grids of numbers, known as matrices. IEEE Spectrum describes an NPU as a specialized chip designed for those calculations. It says NPUs generally handle low-precision arithmetic better than laptop graphics chips do. Microsoft's developer guide says many NPUs support only integer math in compact formats such as INT8, so models must be converted, or quantized, before they run on one.

An NPU works alongside the central processor (CPU) and graphics processor (GPU) rather than replacing them. Microsoft's support site says an NPU uses less power on AI tasks than either and frees them for other work. The developer guide says Windows ML software checks which accelerators a PC has, selects the best performer and falls back to the GPU or CPU if that one fails.

How NPUs are measured

Chipmakers rate NPUs in TOPS, or trillions of operations per second. Apple quoted a similar measure for a phone chip on September 12, 2017. Its iPhone X announcement said the A11 Bionic's dual-core neural engine performed up to 600 billion operations per second and enabled Face ID.

On May 20, 2024, Microsoft introduced Copilot+ PCs, a class of Windows PCs with NPUs rated at 40 TOPS or more. It said the first models would launch with Qualcomm's Snapdragon X Elite and X Plus chips.

Graphics chips can post much higher figures at a cost in power. IEEE Spectrum notes that Nvidia's GeForce RTX 5090 desktop graphics card is specified at up to 3,352 TOPS and designed to draw up to 575 watts.

Where things stand in 2026

IEEE Spectrum reported in November 2025 that AMD and Intel NPUs delivered 40 to 50 TOPS. It put the 2023 figure at about 10 TOPS on the few AMD chips that had NPUs. Qualcomm's Snapdragon X2, unveiled in September 2025, was quoted at 80 TOPS against 45 for the first generation. As of October 2026, Qualcomm's product page lists 80 TOPS for some Snapdragon X2 Elite chips and up to 85 for others. These ratings are the chipmakers' own figures.

NPUs do not handle every AI task. Nvidia announced its RTX Spark chip for Windows PCs at the Computex 2026 trade show. IEEE Spectrum reported that its NPU would run some background features such as Windows Recall. The GPU would handle large language models and image generation, it said.