CUDA Profiler Tools Interface (CUPTI) 13.2 is now available

Developer Tools CUPTI – CUDA Profiler Tools Interface
1

CUDA Profiler Tools Interface (CUPTI) 13.2 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 13.2 Update 1

    New Features

    • Added three new fields to the green context record: workqueueResourceId, workqueueConcurrencyLimit AND workqueueSharingScope, which provide the resource ID, concurrency limit and sharing scope of the work queue, respectively. The activity record CUpti_ActivityGreenContext has been deprecated and replaced by CUpti_ActivityGreenContext2

    Resolved Issues

    • Removed usage of C++ features in the CUPTI public interface, which caused build issues on some platforms.

Updates in CUDA Toolkit 13.2

    New Features

    • User-Defined Activity Records: CUPTI now supports user-defined activity records, allowing users to select specific fields for activity records instead of collecting complete predefined records. This feature addresses the limitations of fixed-field records by providing significant memory efficiency through custom field selection tailored to application-specific profiling needs. Key benefits include optimized memory usage by eliminating unused fields and padding, improved performance through compact data structures and faster data access, and improved backward compatibility as new fields can be added in future CUPTI versions without impacting existing user code. The feature is enabled using the CUPTI_ACTIVITY_ATTR_USER_DEFINED_RECORDS attribute, and new APIs have been added to support this functionality. For detailed information, see CUPTI User-Defined Activity Records and CUDA tracing with User-Defined Activity Records.

      Note

      This feature is currently in beta. APIs and behavior may change in future releases.

    • Added tracing support for Memory Locality Optimized Partition (MLOPart) devices.

    • Added numTpcs parameter to device record to report the total number of Thread Processing Clusters (TPCs) in the device. The activity record CUpti_ActivityDevice5 is deprecated and replaced by CUpti_ActivityDevice6.

    • Added activity kind CUPTI_ACTIVITY_KIND_GREEN_CONTEXT and structure CUpti_ActivityGreenContext to trace green context allocations, which enable GPU resource partitioning by assigning dedicated subsets of SMs and TPCs to specific contexts for fine-grained resource management and isolation. Starting from CUDA 13.2, CUPTI only emits CUpti_ActivityGreenContext for green contexts and no longer emits CUpti_ActivityContext for them.

    • The portDev0 and portDev1 fields in NVLink records are now dynamically allocated arrays, via malloc(), sized to physicalNvLinkCount ports. Clients must release this memory using free() when no longer needed. The record CUpti_ActivityNvLink4 has been deprecated and replaced by CUpti_ActivityNvLink5. This change is not backward compatible. Clients using CUDA 13.2 or later must update their code to use the new record structure.

    • Introduced CUPTI_MAX_DEVICES macro in cupti_common.h to represent the theoretical maximum number of devices supported by CUPTI.

    Resolved Issues

    • Report CUPTI_CBID_RESOURCE_GRAPH_NODE_SET_PARAMS callback for CUDA Graph nodes like memcpy, memset, host and event.

Getting Started

Latest PRODUCT INFO

3

Thanks for the heads-up on the CUpti_ActivityNvLink5 breaking change. Requiring clients to manually free() the dynamically allocated ports array is definitely something we need to patch in our custom profiler right away. Quick question: if we mix older CUDA runtimes with this CUPTI version on a multi-GPU setup, will it gracefully fall back or just SegFault if the memory isn’t handled correctly?

4

This is strictly a CUPTI-level change, the behavior is completely independent of the CUDA runtime version you are using. There is no graceful fallback tied to older runtimes. If you are using CUPTI 13.2+, your profiler must handle the memory allocations for the NvLink record.