CUDA Profiler Tools Interface (CUPTI) 13.1 is now available

Developer Tools CUPTI – CUDA Profiler Tools Interface
1

CUDA Profiler Tools Interface (CUPTI) 13.1 is now available for download in the NVIDIA Registered Developer Program.

The NVIDIA CUDA Profiler Tools Interface (CUPTI), distributed as part of the CUDA Toolkit, is a library that enables the creation of profiling and tracing tools that target CUDA applications. CUPTI provides a set of APIs targeted at ISVs creating profilers and other performance optimization tools:
  • the Activity API,
  • the Callback API,
  • the Host Profiling API,
  • the Range Profiling API,
  • the PC Sampling API,
  • the SASS Metric API,
  • the PM Sampling API,
  • the Checkpoint API,
  • the Profiling API,
  • the Python API (available separately)
Using these CUPTI APIs, independent software developers can create profiling tools that provide low and deterministic profiling overhead on the target system, while giving insight into the CPU and GPU behavior of CUDA applications.

Updates in CUDA Toolkit 13.1

    New Features
    • Added support for Compute Engine context switch events. For more details, refer to the section Compute Engine Context Switch.
    • Added tracing for host execution nodes in CUDA Graphs i.e. nodes of type CU_GRAPH_NODE_TYPE_HOST. Enable with the activity kind CUPTI_ACTIVITY_KIND_GRAPH_HOST_NODE; records are reported as CUpti_ActivityGraphHostNode.
    • Added tracing for host launches done by CUDA through cudaLaunchHostFunc() API. This is important in understanding the device bubbles in the stream timeline. Enable with the activity kind CUPTI_ACTIVITY_KIND_HOST_LAUNCH; records are reported as CUpti_ActivityHostLaunch.
    • Users can query the collection scope for any metric. The CUpti_MetricCollectionScope enum lists the possible scopes: context or device. A parameter, metricCollectionScope, is added to CUpti_Profiler_Host_GetMetricProperties_Params to return the collection scope for a metric.
    • By default, the counter availability image stores only context-level metrics in the binary blob. A new parameter, bAllowDeviceLevelCounters , has been added to the counter availability API struct CUpti_Profiler_GetCounterAvailability_Params to include device-level metrics in the image.
    • Added a new parameter sku in CUpti_Profiler_DeviceSupported_Params for checking profiling support for a GPU SKU.
    • Added a new parameter priority in the kernel record to provide the launch priority of the kernel. The CUpti_ActivityKernel10 activity record is deprecated and replaced by the new CUpti_ActivityKernel11 activity record.
    Resolved Issues
    • Fixed an issue that could cause invalid (zero) kernel timestamps when using the function cuptiActivityRegisterTimestampCallback to register a timestamp callback. This issue was introduced in the CUDA 12.6 Update 2 release.

Getting Started

Latest PRODUCT INFO