AMD XDNA driver adds 30 ms quantum
AMD’s Linux XDNA accelerator driver for Ryzen AI NPUs is gaining hardware scheduler time-quantum support, a fairness mechanism meant to stop one workload or user context from monopolizing the NPU. The current proposal sets a default 30 ms slice and exposes a module parameter, `amdxdna.time_quantum_ms=`, so administrators can tune behavior for their workload mix.
This is the kind of unglamorous kernel plumbing that matters once NPUs start being shared in real workloads. AMD is clearly responding to the same resource-starvation problem Intel already addressed in its own NPU driver, which is a sign that Linux AI accelerator support is maturing from “it runs” to “it behaves well under contention.”
- –The hardware scheduler can enforce fixed per-context time slices, which is the right lever for multi-user fairness.
- –The 30 ms default is conservative enough to prevent runaway monopolization without making the device feel overly preempted.
- –The tunable module parameter gives distro maintainers and power users a way to balance latency against throughput.
- –If merged, this is another marker that NPU drivers are becoming first-class shared resources, not single-app appliances.
DISCOVERED
2h ago
2026-04-28
PUBLISHED
6h ago
2026-04-28
RELEVANCE
AUTHOR
Fcking_Chuck