The AMDXDNA accelerator driver for Ryzen AI NPUs is preparing a new feature called hardware scheduler time quantum for ensuring fairness between multiple users/contexts wanting to leverage this neural processing unit for AI workloads.

The Intel IVPU accelerator driver for Core Ultra NPUs recently enacted [1]a change to prevent single programs from exhausting all NPU resources . That change was picked up for Linux 7.1 and prevents a single non-root user/program from occupying all contexts and doorbells of the Intel NPU. AMD is also working on preventing single programs from hogging all NPU resources and for their approach they are rolling out hardware scheduler time quantum support.

The hardware scheduler time quantum is a hardware capability of the Ryzen AI NPU for improving fairness between concurrent contexts. The NPU hardware scheduler can enforce a fixed time slice per context to prevent long-running workloads from using all resources and not allowing access by other users/workloads. The default time quantum is 30 ms and via the new amdxdna.time_quantum_ms= module parameter that threshold can be configured.

This hardware scheduler time quantum support is currently out for review on [2]the LKML . Given the timing it will hopefully be ready for the Linux v7.2 cycle. Great seeing both Intel and AMD having given thought to NPU resource sharing while I'm sure most Linux desktop users would be glad to see their NPUs actually just rather used in the first place. At least the number of Linux software packages natively making use of Intel/AMD NPUs is slowly expanding.



[1] https://www.phoronix.com/news/Intel-NPU-Restrictions-Linux-71

[2] https://lore.kernel.org/all/20260415171139.904947-1-lizhi.hou@amd.com/