Name

VK_NV_compute_occupancy_priority - device extension

VK_NV_compute_occupancy_priority

Name String

VK_NV_compute_occupancy_priority

Extension Type

Device extension

Registered Extension Number

646

Revision

1

Ratification Status

Not ratified

Extension and Version Dependencies

VK_KHR_get_physical_device_properties2 or Vulkan Version 1.1

Contact

• Chris Lentini @clentini%0A*Here describe the issue or question you have about the VK_NV_compute_occupancy_priority extension*

Extension Proposal

VK_NV_compute_occupancy_priority

Other Extension Metadata

Last Modified Date

2025-12-01

Contributors

• Chris Lentini, NVIDIA
• Eric Werness, NVIDIA
• Lionel Duc, NVIDIA
• Peter Deayton, NVIDIA

Description

This extension provides applications with control over how their compute workloads utilize GPU compute resources, specifically allowing prioritization relative to other simultaneously executing workloads. Applications can specify the priority with which compute workloads should occupy GPU compute resources, allowing for a fine-grained distinction between workloads that may want to execute at a background priority over a long period of time versus workloads with harder latency requirements.

The extension introduces a new command cmdSetComputeOccupancyPriorityNV that allows applications to set the occupancy priority for subsequent compute dispatches. The occupancy priority affects how compute workloads utilize GPU compute resources relative to other simultaneously executing workloads.

The occupancy priority is stateful on a command buffer. All commands listed in the Dispatching Commands chapter issued subsequent to a cmdSetComputeOccupancyPriorityNV call will be executed with the specified priority parameters until another cmdSetComputeOccupancyPriorityNV call is made.

For convenience, three named occupancy priority values are defined:

• VK_COMPUTE_OCCUPANCY_PRIORITY_LOW_NV - a constant value that can be used for ComputeOccupancyPriorityParametersNV::occupancyPriority to specify a low priority level.
• VK_COMPUTE_OCCUPANCY_PRIORITY_NORMAL_NV - a constant value that can be used for ComputeOccupancyPriorityParametersNV::occupancyPriority to specify a normal priority level. This represents the default priority level.
• VK_COMPUTE_OCCUPANCY_PRIORITY_HIGH_NV - a constant value that can be used for ComputeOccupancyPriorityParametersNV::occupancyPriority to specify a high priority level.

All command buffers (primary and secondary) start with a priority level equal to the VK_COMPUTE_OCCUPANCY_PRIORITY_NORMAL_NV value. The priority state is not inherited by secondary command buffers - each command buffer maintains its own independent priority state.

New Commands

• cmdSetComputeOccupancyPriorityNV

New Structures

• ComputeOccupancyPriorityParametersNV

• Extending PhysicalDeviceFeatures2, DeviceCreateInfo:

  • PhysicalDeviceComputeOccupancyPriorityFeaturesNV

New Enum Constants

• COMPUTE_OCCUPANCY_PRIORITY_HIGH_NV

• COMPUTE_OCCUPANCY_PRIORITY_LOW_NV
• COMPUTE_OCCUPANCY_PRIORITY_NORMAL_NV
• NV_COMPUTE_OCCUPANCY_PRIORITY_EXTENSION_NAME
• NV_COMPUTE_OCCUPANCY_PRIORITY_SPEC_VERSION

• Extending StructureType:

  • STRUCTURE_TYPE_COMPUTE_OCCUPANCY_PRIORITY_PARAMETERS_NV
  • STRUCTURE_TYPE_PHYSICAL_DEVICE_COMPUTE_OCCUPANCY_PRIORITY_FEATURES_NV

The extension only allows specification of occupancy priority for compute workloads, however, the priorities will also impact the prioritization of compute workloads relative to simultaneously executing graphics workloads. In such a scenario, the graphics workload may be thought of as executing at VK_COMPUTE_OCCUPANCY_PRIORITY_NORMAL_NV priority, and so a simultaneously executing compute workload with VK_COMPUTE_OCCUPANCY_PRIORITY_HIGH_NV occupancy priority will preferentially utilize available compute resources.

Workloads specified with a higher priority may begin execution after workloads specified with a lower priority, at which point they may find GPU compute resources already occupied. So, while they will from that point forward preferentially occupy available compute resources, they may not ramp up to full occupancy until the already present lower priority work has reached a point where it can relinquish compute resources.

Issues

None.

Version History

• Revision 1, 2025-08-06 (Chris Lentini)

  • Initial revision

See Also

No cross-references are available

Document Notes

For more information, see the Vulkan Specification.

This page is a generated document. Fixes and changes should be made to the generator scripts, not directly.