AMD XDNA Neural Processing Units (Ryzen AI NPUs)
The on-die AI accelerator (neural processing unit) found in AMD's Ryzen AI laptop and embedded processors, branded XDNA. It offloads machine-learning inference workloads from the CPU and GPU and is exposed to userspace through the kernel's accelerator framework as a PCI device.
recommendation
It should stay in the kernel because this supports the NPU (neural processing unit) built into AMD's Ryzen AI client and embedded CPUs, which is current shipping silicon. AMD launched Ryzen AI 300 laptops with XDNA 2 in early 2025 and announced Ryzen AI Embedded parts using the same NPU in early 2026, and the driver itself is actively maintained upstream with regular bug fixes and no signs of retirement.
repository signals
sources
- docs.kernel.org
Upstream kernel documentation has a dedicated amdxdna driver section and states it supports the AMD NPU.
- amd.com
AMD said Ryzen AI 300 / PRO 300 systems with AMD XDNA 2 NPUs were expected to ship starting in Q1 2025, showing fresh commercial deployment in the target window.
- amd.com
AMD announced Ryzen AI Embedded processors in 2026 with XDNA 2 NPUs and an open software stack, indicating ongoing new embedded deployments beyond laptops.
codex reasoning notes (technical)
Keep. Local inspection via shell `sed` of Kconfig and `amdxdna_pci_drv.c` showed this is a real PCI DRM accel driver for AMD NPUs integrated into Ryzen AI client CPUs, with firmware entries and supported device IDs. Local git history via `git -c safe.directory=... log` showed active maintenance through 2026-03-10 with many bugfixes and no deprecation/removal discussion hits for the directory; the only 'remove' subjects were scoped internal cleanups such as dropping specific support paths, not driver retirement. URLs were obtained with web search plus `open`/`find`: kernel docs page for official upstream scope, AMD's 2025 CES press release for new-in-2025 Ryzen AI 300 availability, and AMD's 2026 embedded press release for continued current-market XDNA deployment. `lei`/lore MCP was unavailable in this environment, so upstream-attention checking used the local kernel history fallback instead of live lore queries.