CUDA 13.5 expands control over shared GPUs

NVIDIA’s release moves Vera Rubin support from early access to general availability. It also expands controls for applications sharing GPU resources. Python developers gain a way for cooperating GPU functions to overlap work within one execution stream. This requires compute capability 9.0 or newer and functions implementing the required synchronization protocol; enabling a launch option alone is not sufficient.

The MPS guide explains the scheduling boundary. Version 3 is optional; legacy version 2 remains the default. With driver r620 or newer, weighted timeslicing changes how long a server’s computing turn lasts, without guaranteeing its share of GPU utilization. Set the common scheduling policy before creating any server entries. Even configured servers that have not started prevent changes. A server’s individual percentage can change only when none of its namespaces has attached clients.

The Python reference documents that the overlap option is off by default. When enabled, it permits overlap with a preceding function in the same stream that signals completion programmatically. Another configuration detail matters when reading launch settings: the grid counts blocks normally, but counts clusters when a cluster is specified. The block setting describes threads within each block.