Yesterday the GitHub-hosted AMD XDNA driver code saw a new tagged release as version 202610.2.21.17. That itself wasn’t too interesting but while diving into there is new yet-to-be-merged code for a new “NPU3A” revision to their NPU3 IP in Ryzen AI.
The AMD XDNA 202610.2.21.17 driver release simply notes user pointer buffer object allocation support as the new feature of this tagged release. User pointer support is beneficial albeit no other details on this new tagged version. Going through the GitHub activity shows a variety of recent commits to this AMD Ryzen AI NPU support code.
One potentially interesting bit is apparent work on making the AMDXDNA user-space driver code suitable for upstreaming into AMD XRT proper. AMD XRT is the Xilinx Runtime for AIE PCIe accelerator cards and FPGA platforms. The AMDXDNA / Ryzen AI driver code has been downstream of XRT but this recent commit notes “minor tweaks to build xdna-driver UMD for upstreaming with XRT.” I haven’t seen any official word but we’ll see if the AMDXDNA user-space driver code is ultimately moved upstream into AMD XRT run-time itself as opposed to being an independent downstream.
The AMDXDNA 202610.2.21.17 driver can be downloaded from GitHub.
Not yet merged / not in that release but spotted when digging through the latest AMD XDNA GitHub activity is this pull request for adding new device IDs for a new NPU3A variant to NPU3. But no description on that pull for outlining what may be new/different with this new “A” variant to their neural processing unit IP. From the driver side just two new device IDs are added for AMD NPU3A.
Here’s to hoping for a more compelling AMD Ryzen AI NPU story for Linux users in 2026.
