What Is Intel AI Boost?

Intel AI Boost is Intel’s brand name for the neural processing unit (NPU) built into its Core Ultra processors. It first appeared in Core Ultra Series 1, codenamed Meteor Lake, in late 2023, and every Core Ultra generation since has carried one. In Windows Task Manager and in Linux it simply appears as the Intel NPU.

AI Boost sits beside the CPU cores and Intel graphics on the same processor, so a computer gains a dedicated AI engine without an extra card, slot or power budget. It is designed for sustained, low-power inference: the kind of model that runs continuously on a camera feed, a microphone or a sensor stream. Short, heavy bursts of AI work usually run faster on the integrated GPU, so the two complement each other.

AI Boost takes on AI work that would otherwise load the CPU cores or graphics.

AI Boost ratings have risen sharply with each generation. The figures below are Intel’s peak INT8 ratings for the NPU alone; whole-chip platform figures are higher because they add the CPU and GPU.

Arrow Lake reused an NPU close to Meteor Lake’s, so the large jump came with Lunar Lake and Panther Lake. Check the exact processor, because ratings vary within a generation. Our Intel Core Ultra white paper covers the rest of each generation’s design.

Applications do not program the NPU directly. They go through a runtime that compiles the model for it. Intel’s OpenVINO toolkit is the main route on both Windows and Linux, and it can load models from PyTorch, TensorFlow and ONNX. ONNX Runtime reaches the NPU through its OpenVINO execution provider, and on Windows 11, Windows ML and DirectML can target it too.

On Linux, the NPU needs Intel’s open-source NPU driver, which is in the mainline kernel as the intel_vpu driver, plus the matching user-space firmware and libraries. Not every model layer runs on the NPU; unsupported layers fall back to the CPU or GPU, so test the full model on the target processor.

On Windows 11, Task Manager shows the NPU as its own entry on the Performance tab, with a utilization graph that rises when an application sends work to it. If the entry is missing, the Intel NPU driver is not installed or the NPU is disabled in the BIOS. On Linux, the driver creates an accel device under /dev/accel, and OpenVINO lists NPU among its available devices once the user-space libraries are installed.

The simplest functional test is to run the same model through OpenVINO on the CPU and then on the NPU, and compare latency, throughput and power draw. A model that runs much slower on the NPU usually has layers falling back to the CPU, which the OpenVINO logs will show.

For edge AI, AI Boost is one of the main reasons to choose Core Ultra. A fanless box PC can run a vision model all day on the NPU while the CPU handles communications, logging and the user interface, with no GPU card to cool. Our Intel Core Ultra computers list the NPU rating for each processor, from compact fanless systems to panel PCs.

If the NPU is not enough for the workload, the integrated Arc graphics can share the load, or a Hailo accelerator or discrete GPU can be added. Our white paper on what an NPU is and how it works explains how to split work between them.

We configure Intel Core Ultra computers with the NPU driver and OpenVINO runtime installed and validated as standard, on Windows or Linux. Tell us about your model and workload, and we will recommend the Core Ultra system that suits it.