What Is Intel Core Ultra? Series 1, 2 and 3 Explained

Intel Core Ultra is Intel’s processor brand for its higher-performance Core parts since late 2023, built from separate tiles and the first Intel line to put an NPU, a dedicated AI engine, on every chip. It replaced the familiar Core i5, i7 and i9 names with Core Ultra 5, 7 and 9, and it counts generations as Series 1, 2 and 3 rather than 12th, 13th or 14th Gen.

For industrial computing, Core Ultra brings hardware AI acceleration to the x86 platform that most industrial software already runs on. A Core Ultra system can run vision, inference or analytics on the NPU while the CPU cores handle control and the integrated Arc graphics drive the displays, without a separate graphics card.

This guide explains what each Core Ultra series is, how the naming works, how Core Ultra differs from Core i7 and from the non-Ultra Core range, what the NPU does and how much AI it delivers, and where Core Ultra fits in fanless, edge AI and long-life industrial systems.

Each Core Ultra series covers several codenames for different segments. The table sets out the parts that matter for industrial and edge systems. NPU figures are Intel’s peak INT8 ratings, rounded, and vary by model.

The Series number is the generation. Within a series, the suffix tells you the segment, and the model number tells you the tier and position, so a Core Ultra 7 255H is a Series 2 Arrow Lake mobile part and a Core Ultra 5 325 is a Series 3 Panther Lake part. Because the name no longer carries a simple generation digit, check the exact model on Intel’s specification pages before matching it to a board. Our guide to CPU generations and embedded processors explains how Core Ultra sits among Intel’s other families.

Core Ultra Series 1 followed the 13th and 14th Gen Core processors, so it is sometimes described as the 14th or 15th generation, but Intel does not use those numbers for it. Series 2 followed in 2024, with Lunar Lake for thin laptops and Arrow Lake for desktops and larger laptops, and Series 3, Panther Lake, arrived in 2026 as the first Intel processors on the Intel 18A process.

The latest series is not automatically the right one for an industrial build. Series 1 and Series 2 parts are well established in fanless box PCs, NUC boards and panel PCs, with boards and BIOS that have been in production for some time, while Series 3 boards are newer. Lunar Lake, with its memory on the package and thin-laptop focus, appears rarely in industrial systems. Choose the series whose boards fit the enclosure, power budget and lifecycle, then the model within it.

Core Ultra is more than a rename. The change from Core i7 to Core Ultra 7 came with a new way of building the processor and a new AI engine, though the hybrid core layout carried over from 12th Gen Core.

In practice a Core Ultra 7 is not always faster than a Core i7 of the previous generation at the CPU’s own work; the gains are in efficiency, graphics and AI. Our explainer on Core Ultra vs Core i7 compares the two directly, and our guide to P-cores and E-cores explains the hybrid layout both share.

Intel also sells plain Core 3, Core 5 and Core 7 processors. They are a separate range, not a cut-down Core Ultra, and they matter in industrial computing because many low-power and long-life boards use them.

Our Intel pages cover every family we supply. The Bartlett Lake and Wildcat Lake pages show the non-Ultra parts in our range, and the Intel Core Ultra page gathers every Core Ultra system we supply.

Earlier Core processors were cut from one piece of silicon. Core Ultra is assembled from tiles, each made on the process that suits it and stacked on a base with Intel’s Foveros packaging. Meteor Lake, for example, has a compute tile with the CPU cores, a graphics tile, an SoC tile with the NPU, media engine and low-power E-cores, and an I/O tile.

The low-power E-cores in the SoC tile are a Core Ultra feature that matters at the edge: background tasks, network services and video playback can run there while the main cores sleep, which cuts idle power in fanless systems. Intel describes the tiled approach as letting it mix process nodes and scale the design; for a system builder it mainly means more models, with different graphics and NPU sizes, on a common platform. Our guide to the system on chip explains how this compares with a monolithic SoC.

The NPU, which Intel calls Intel AI Boost, is a block of hardware built for the matrix maths in neural networks. It runs inference, such as object detection, classification, noise suppression or text recognition, at a fraction of the power the CPU or GPU would need, and it keeps working in the background while the CPU runs the application.

Software reaches it through Intel’s OpenVINO toolkit or through Windows ML, DirectML and ONNX Runtime, and the same model can be moved between CPU, GPU and NPU to find the best balance of speed and power. In a typical edge build the NPU takes steady, always-on models, the GPU takes heavier or batched inference, and the CPU handles control and the user interface. Our guide to what an NPU is explains the hardware in more detail, and our comparison of NPU vs GPU covers when each is the better home for a model.

Intel quotes two AI figures for Core Ultra: the NPU’s own rating and a platform figure that adds the CPU and GPU. The platform figure is much larger, which is why marketing numbers differ so widely. The table gives Intel’s approximate peak INT8 figures for representative parts.

Read these as peaks under ideal conditions. Real throughput depends on the model, its precision and the software, and the GPU and NPU share the same memory bandwidth and power budget, so they rarely reach their peaks together. Our explainers on what TOPS means and how many TOPS an NPU needs help translate the figures into camera streams and models.

For light and medium edge AI, yes. A Core Ultra system can run several detection models on camera streams, speech and text models, or anomaly detection on sensor data, while still running the main application, which removes the need for a graphics card in many projects. For many camera streams, large vision models or robotics, a discrete GPU or an NVIDIA Jetson module is still the stronger choice. If the project is still choosing between architectures, our comparison of x86 and Arm sets out the wider trade.

Our edge AI computers include Core Ultra, Jetson and GPU systems, so the platform can follow the model rather than the other way round. Our explainer on Intel Arc covers the graphics side of the platform.

The platform around the cores changes between series as much as the cores do, and it decides which boards and expansion a system can have.

Check the board as well as the processor: a processor may support PCIe 5.0 or four displays, but the board decides which lanes and outputs are brought out. Our comparison of DDR4 and DDR5 covers the memory types, and our guide to PCIe generations covers lanes and bandwidth.

AMD’s answer to Core Ultra is the Ryzen AI range, which also pairs x86 cores and integrated graphics with an NPU. Both run Windows and Linux and both suit compact edge AI systems, so the choice usually comes down to the board, the software toolchain and the graphics the project needs.

Our explainer on AMD Ryzen AI covers the AMD side, and the AMD Ryzen AI 300 page lists the boards and systems we supply.

Core Ultra runs the same operating systems as earlier Intel processors, with one caveat: the NPU and the newest graphics need current drivers. Windows 11 IoT Enterprise LTSC 2024 supports the NPU through Windows ML and DirectML and is the usual choice for long-life Windows systems; our guide to Windows 11 IoT Enterprise LTSC covers the editions and support dates. On Linux, recent Ubuntu LTS kernels support Core Ultra graphics and the NPU driver, and OpenVINO runs on both.

Older operating systems are the risk. Windows 10 and older Linux kernels may boot but leave the NPU, graphics or power management without full support, which is why we confirm driver support for every interface before we lock an image for a fleet.

Core Ultra suits modern industrial PCs and edge AI nodes that need AI acceleration alongside x86 software compatibility. The power class sets where it can go: U and H parts run fanless in compact box PCs and on NUC and Mini-ITX boards, while desktop S and T parts suit larger fanless or actively cooled systems with PCIe cards.

Base power is not the whole story: turbo limits can be set higher or lower in the BIOS, and we tune them to the enclosure so performance holds under continuous load. Our explainer on TDP covers how those limits work.

Intel offers edge versions of Core Ultra with long-life availability, and boards built for industrial use add the features embedded designs need. In our range, Series 3 NUC and Mini-ITX boards support Intel In-Band ECC, desktop W880 and Q870 boards add ECC or vPro remote management, and many boards run from wide DC inputs and are rated from −20 to 70 °C. Our NUC and 4X4 motherboards and industrial Mini-ITX motherboards are the usual starting points for a custom build, and our guide to choosing an embedded motherboard covers the board side.

These systems and boards cover all three series. We fit memory and storage, set the power limits to the enclosure, image Windows 11 IoT Enterprise or Linux with the application and AI runtime, and test each system before it ships.