Last month [1]NXP posted open-source Linux kernel driver patches for their Neutron NPU accelerator . The NXP Neutron NPU aims to help with edge AI applications and this neural processing unit is found in their different SoCs. Unfortunately, their GitHub repository for the user-space software ends up containing a binary-only blob that will end up delaying plans on getting this driver into the mainline Linux kernel.

NXP hosts a GitHub repository with a user-space library and custom LiteRT (TensorFlow-Lite) delegate workloads on this Neutron NPU found with NXP SoCs like the i.MX95. Unfortunately, it's now been discovered by upstream Linux kernel developers that this open-source repository ends up depending upon a binary-only library, libNeutronDriver.so .

NXP engineers commented on the Linux kernel mailing list that they plan to publish the source code but it could still be a few months away.

Thus the latest on [2]the kernel mailing list thread is that NXP engineers are raising the matter internally. But without an open-source user-space implementation being met, the kernel driver won't be mainlined if there are no open-source user-space usage -- a situation that has delayed or prevented other kernel drivers from being upstreamed in the past if they are only useful with binary blobs in user-space.



[1] https://www.phoronix.com/news/NXP-Neutron-Linux-Accel-Driver

[2] https://lore.kernel.org/lkml/PA4PR04MB9413D4380B1F71071401EBB69456A@PA4PR04MB9413.eurprd04.prod.outlook.com/