NPU

NPU stands for neural processing unit. It identifies a class of hardware specialized for neural-network computation, rather than one universal processor design. An NPU may be integrated into a processor alongside other computing resources. (Intel: AI accelerators)

Inference means using a trained model to produce a result from new input. Integrated NPUs are one way to accelerate this work on end-user devices. The NPU does not replace the general-purpose software and processing needed by the rest of the application. (Intel: CPU vs. GPU) (Intel: AI accelerators)

A GPU also accelerates AI, but has a broader parallel-processing role that includes graphics and other computing tasks. An NPU focuses on neural-network operations. Neither label alone establishes which device is best for a particular application. (Intel: CPU vs. GPU) (Intel: AI accelerators)

Imagine a device that uses a trained model to interpret a voice command. As a simplified example, the model computation could be assigned to an NPU while the application handles the resulting action. Intel lists voice-command responses among AI-accelerator use cases. (Intel: AI accelerators)

For evaluation, ask which part of the application actually runs on the accelerator. Model execution time and the time a user waits for a response are different measurements; a useful test should state its scope.

Use these questions as a reading checklist, not as a device-selection recommendation. Exact manufacturer documentation takes priority over a general definition.

• Confirm that the model and software runtime support the device.
• Check supported operations and numerical formats in device documentation.
• Specify the workload and operating conditions for any efficiency comparison.
• Measure the full application, not only an isolated accelerator operation.

For the broader explanation, compare npu, gpu and cpu roles. This page defines the term; the companion guide connects it to the rest of the system.