What Is AMD Ryzen AI?

AMD Ryzen AI is AMD’s name for the AI capability built into its Ryzen processors: a neural processing unit (NPU) on the processor die, plus the software that runs AI models on it. The NPU sits beside the Zen CPU cores and the Radeon graphics, so one chip carries three engines that can each run inference.

The NPU uses AMD’s XDNA architecture, which came from the AI Engine technology AMD gained when it acquired Xilinx. It is an array of small compute tiles with local memory, built for the matrix maths in neural networks. That lets it run AI workloads such as vision, speech and background effects at a fraction of the power the CPU or GPU would need, leaving them free for other work. Because the NPU sits on the same die, it shares the system memory and adds no extra board space, cabling or cooling. Our guide to NPUs explains how this class of processor works.

AMD first put an NPU into Ryzen mobile processors in 2023, and each generation since has raised its performance. The table lists the main families and AMD’s published NPU figures.

The NPU figures are INT8 peaks for the NPU alone. AMD also quotes whole-processor totals that add the CPU and GPU: the Ryzen 7 8845HS, for example, lists a 16 TOPS NPU and 38 TOPS overall. When comparing processors, make sure both figures are on the same basis.

Hardware is only useful if models can reach it. AMD’s Ryzen AI Software is the main route, and it follows a standard pattern.

The NPU is at its best on sustained, steady inference: a detection model watching one or two cameras, speech recognition on a kiosk, or noise suppression on a call, running all day at low power. It frees the CPU for the application and the Radeon graphics for displays and video decode. Our NPU vs CPU comparison covers that split in more detail.

It is less suited to very large models, many high-resolution streams or anything that needs training, where the integrated Radeon graphics, a discrete GPU or a dedicated accelerator does better. Not every model layer runs on the NPU either, so check that the model’s operators are supported before relying on the headline TOPS.

Ryzen AI brings on-chip inference to compact AMD-based industrial computers. Our Ryzen AI 300 computers include an industrial AI PC built on the Ryzen AI 9 HX 370, with options for Hailo accelerators alongside the 50 TOPS NPU, and NUC PCs and 4×4 boards built on the Ryzen AI 7 350 and Ryzen AI 5 340. Our AMD Ryzen 8000 computers pair a 16 TOPS NPU with Radeon graphics in fanless and embedded designs.

Embedded versions of Ryzen processors also carry longer availability than consumer parts, which matters for systems that must be supplied unchanged for years. Our guide to AMD Ryzen Embedded boards and industrial PCs covers the embedded families.

The two families suit different jobs. Ryzen 8000 processors are a mature choice with Zen 4 cores and strong Radeon graphics, and their 16 TOPS NPU handles light, always-on models such as a single detection stream or speech commands. Ryzen AI 300 processors move to Zen 5 cores, newer RDNA 3.5 graphics and an NPU with roughly three times the throughput, which leaves headroom for several models at once or for small language models running locally.

In practice, start from the workload. If the AI task is modest and the build needs the widest choice of boards and enclosures, Ryzen 8000 is often enough. If the NPU will carry the main AI load, or the system must last through several model updates, Ryzen AI 300 gives more room to grow.

We build AMD Ryzen AI computers with the NPU drivers and AI software installed as standard, and can validate your model on the target processor before the system ships.