NVIDIA CUDA Profiling Tools Interface (CUPTI) - CUDA Toolkit

The NVIDIA CUDA Profiling Tools Interface (CUPTI) is a dynamic 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 Event API,
  • the Metric API,
  • the Profiler API,
  • the PC Sampling API,
  • the SASS Metric API, and
  • the Checkpoint 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. Normally packaged with the CUDA Toolkit, NVIDIA occasionally uses this page to provide CUPTI improvements and bug fixes between toolkit releases.



There is currently no CUPTI update to the CUDA Toolkit 12.6 Update 2. You may obtain the latest version of CUPTI by Downloading the CUDA Toolkit 12.6.2

Download the CUDA Toolkit 12.6 Update 2 Now Download the CUDA Toolkit 12.6 Update 1 Now Download the CUDA Toolkit 12.6 Now

Download CUDA Python 12.6.0 Now
Revision History

Key Features

  • Trace CUDA API by registering callbacks for API calls of interest
    • Full support for entry and exit points in the CUDA C Runtime (CUDART) and CUDA Driver
  • GPU workload trace for the activities happening on the GPU, which includes kernel executions, memory operations (e.g., Host-to-Device memory copies) and memset operations.
  • CUDA Unified Memory trace for transfers from host to device, device to host, device to device and page faults on CPU and GPU etc.
  • Normalized timestamps for CPU and GPU trace
  • Profile hardware and software event counters, including:
    • Utilization metrics for various hardware units
    • Instruction count and throughput
    • Memory load/store events and throughput
    • Cache hits/misses
    • Branches and divergent branches
    • Many more
  • Enables automated bottleneck identification based on metrics such as instruction throughput, memory throughput, and more
  • Range profiling to enable metric collection over concurrent kernel launches within a range
  • Metrics attribution at the high-level source code and the executed assembly instructions.
  • Device-wide sampling of the program counter (PC). The PC Sampling gives the number of samples for each source and assembly line with various stall reasons.
See the CUPTI User Guide for a complete listing of hardware and software event counters available for performance analysis tools.

Updates in CUDA Toolkit 12.6 Update 2

    New Features
    • In CUDA 12.6 GA release, CUPTI added a new set of APIs around Perfworks host APIs to simplify usage, shield users from low-level concepts, and ease adaptation to changes in Perfworks APIs. These host APIs were provided in the header cupti_profiler_host.h. Complementing these, a new set of target APIs are added in CUDA 12.6 Update 2 to simplify profiling for new users and align the call structure with other profiling APIs for faster learning and better adaptability. These target APIs are provided in the header cupti_range_profiler.h, and will be referred to as Range Profiling APIs. For range profiling, it is strongly recommended that users, especially those new to the field, utilize the new host and target APIs introduced in CUDA 12.6. The range_profiling sample shows how to use the host and target range profiling APIs.
        NOTE: It is important to note that existing Profiling API may be deprecated and potentially removed in future releases, so transitioning to new host and target APIs will ensure continued support and compatibility.
    • Added support for tracing and profiling on Microsoft Compute Driver Model (MCDM). This requires an NVIDIA display driver version of 565 or higher.
    Resolved Issues
    • Improved the conversion accuracy of timestamps which are captured on the GPU and converted to CPU timestamps.
    • Timestamps reported in the activity records were not always correct for Windows Subsystem for Linux (WSL) systems. This issue has been fixed.
    • Fixed an issue that resulted in failure to capture the tracing information for kernels inside the body of the conditional node of CUDA Graph. This requires an NVIDIA display driver version of 565 or higher.
    • PM Sampling trigger mode CUPTI_PM_SAMPLING_TRIGGER_MODE_GPU_TIME_INTERVAL is supported for all GA10x (GA102+) chips.
    • Fixed issues with Enhanced compatibility (aka minor version compatibility) of a few CUPTI activities with the driver versions shipped with the CUDA 12.2 or prior releases.
    • Made flushing of activity buffers thread-safe for per-thread activity buffer feature.

Updates in CUDA Toolkit 12.6 Update 1

    Resolved Issues
    • Disabling all callbacks in a domain using the cuptiEnableDomain API could unintentionally disable other activities and callbacks. This issue has been addressed and resolved.
    • Fixed issues with Enhanced compatibility (aka minor version compatibility) of a few CUPTI activities with the driver versions shipped with the CUDA 12.2 or prior releases.

Updates in CUDA Toolkit 12.6

    New Features
    • CUPTI introduces PM Sampling APIs for collecting many hardware metrics by sampling the GPU's performance monitors (PM) periodically at fixed intervals for a CUDA workload. These APIs are provided in the header file cupti_pmsampling.h. These are supported on Turing and later GPU architectures, i.e. devices with compute capability 7.5 and higher. Refer to the section PM Sampling API for more details.
    • The Profiling is split into four main phases - Enumeration, Configuration, Collection and Evaluation. APIs for three phases enumeration, configuration and evaluation are provided directly by the low level Perfworks APIs. CUPTI is adding a new set of APIs around these Perfworks host APIs to simplify usage, shield users from low-level concepts, and ease adaptation to changes in Perfworks APIs. These APIs promote consistency, reduce confusion, and enhance integration with CUPTI APIs. New host APIs are provided in the header cupti_profiler_host.h. Additionally, new target APIs will be introduced in a future release to simplify profiling for new users and align the call structure with other profiling APIs for faster learning and better adaptability.
    • Added support for identifying the source of the memory allocation. This is useful when memory allocation request is made from a library or shared object which is called from the application. A new field source is added in the memory activity record. The activity record CUpti_ActivityMemory3 is deprecated and it is replaced by a new activity record CUpti_ActivityMemory4. Refer to the section Device Memory Allocation Source Tracking for more details.
    • Added a new overhead kind CUPTI_ACTIVITY_OVERHEAD_ACTIVITY_BUFFER_REQUEST to report the overhead incurred while requesting activity buffers from the CUPTI client.
    • Added static library libcupti_static.a for ARM Server (arm64 SBSA).
    Deprecated and dropped features
    • Source/SASS level metrics from the header cupti_activity.h are deprecated on Volta and later GPU architectures and these will be removed in a future release. It includes metrics CUPTI_ACTIVITY_KIND_GLOBAL_ACCESS, CUPTI_ACTIVITY_KIND_SHARED_ACCESS, CUPTI_ACTIVITY_KIND_BRANCH and CUPTI_ACTIVITY_KIND_INSTRUCTION_EXECUTION and related correlation kinds CUPTI_ACTIVITY_KIND_SOURCE_LOCATOR and CUPTI_ACTIVITY_KIND_INSTRUCTION_CORRELATION. It is recommended to move to the SASS Metric API from the header cupti_sass_metrics.h which is supported on Volta and later GPU architectures.
    Resolved Issues
    • Exported the profiling API cuptiProfilerIsPassCollected.
    • Reduced the collection overhead for memcpy tracing.
    • For some activities, duration for the activity buffer request was getting included in the activity duration. This issue is resolved.

Initial Release of CUPTI Python Profiling APIs

  • CUPTI Python provides Python APIs for creation of profiling and tracing tools that target CUDA Python applications.
  • This release of CUPTI Python supports a subset of CUPTI C Activity and Callback APIs for linux x86_64, including samples.
  • Available separately from the CUDA Toolkit only.
    Please refer to the CUPTI Python 12.6.0 overview, release notes, and user guide for more details.

Requirements

    Supported platforms
    • Linux x86_64[1]
    • Windows x86_64[1]
    • Linux aarch64 SBSA[1]
    • DRIVE OS QNX aarch64[2]
    • DRIVE OS Linux aarch64[2]
    [1] available in the CUDA Desktop Toolkit only
    [2] available in the Embedded or Drive toolkits only
    Supported NVIDIA GPU architectures
    • Activity and Callback APIs
      • All architectures supported by CUDA Toolkit
    • Event and Metric APIs
      • Volta
      • Pascal
      • Maxwell
    • Profiling and PC Sampling APIs
      • Hopper: GH100
      • Ada: AD10x
      • Ampere: A100 with Multi-Instance GPU, GA10x
      • Turing
      • Volta: GV100, GV10B
    CUDA Toolkit
    Drivers
      Please use the following drivers
      • 560.94 (Windows)
      • 560.35.03 (Linux)
      provided with CUDA Toolkit 12.6 Update 2 production release or a more recent version.

Documentation

Support

To provide feedback, request additional features, or report issues, please use the Developer Forums.

Installation Overview

When installing CUDA Toolkit 12.6 Update 2 and specifying options, be sure to select CUDA > Development > Tools > CUPTI.