CUDA 13.1 brings GPU resource partitioning to Runtime API

NVIDIA says CUDA 13.1 makes green contexts accessible through its Runtime API, allowing applications to assign concurrent workloads to selected GPU execution resources. The feature has been available through the Driver API since CUDA 12.4. Applications can divide GPU compute units between workloads and provision workqueues to reduce unintended serialization. This gives developers a way to reserve resources for latency-sensitive tasks alongside background work, while retaining the existing stream-based programming model. Dedicated resources can let critical work avoid waiting for bulk tasks to release compute units. The tradeoff is that fewer units remain available to the bulk workload; applications can adopt the feature incrementally.