NVIDIA CUDA Profiling Tools Interface (CUPTI) - CUDA Toolkit

The NVIDIA CUDA Profiling Tools Interface (CUPTI) 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. 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 13.3. You may obtain the latest version of CUPTI by Downloading the CUDA Toolkit 13.3.0

Download the CUDA Toolkit 13.3 Now
Download the CUPTI Python API 13.0.0 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 13.3

New Features

The User-Defined Activity Records feature is transitioning from Beta to production in this release. For more details, refer to CUPTI User-Defined Activity Records and CUDA tracing with User-Defined Activity Records.
The Multiple Subscribers for Activity Tracing feature is transitioning from Beta to production in this release. For more details, refer to Multiple Subscribers.
Added support for profiling workloads on Green Contexts with the new APIs cuptiRangeProfilerGetDevicePartitionInfo and cuptiProfilerHostSetDevicePartitionInfo. These APIs enable retrieval and configuration of device partition information, ensuring accurate metric evaluation by accounting for partitioned device resources. The cuptiRangeProfilerGetDevicePartitionInfo API retrieves partition information from the range profiler, while cuptiProfilerHostSetDevicePartitionInfo applies this information to the profiler host for metric scaling.
Enhanced the cuptiActivityEnableHWTrace API to allow invocation before CUDA driver initialization. After driver initialization, CUPTI may switch to semaphore-based tracing if HES based tracing is unsupported, with errors reported via callbacks in the CUPTI_CB_DOMAIN_STATE domain.
Added profiling support for previously unsupported SKUs, with reduced metric support, including GPUs such as RTX 6000D BSE (GB202-891), RTX 5090DD (GB202-240), and RTX 5090D (GB202-250). For more details, refer to ERR_NVGPU. To indicate this limited support, CUPTI has added the CUPTI_PROFILER_CONFIGURATION_LIMITED_SUPPORT value to the CUpti_Profiler_Support_Level enum to indicate limited support for these SKUs. When using the counter availability image, CUPTI will list all supported metrics during a query, and report an error for unsupported metrics during config image creation. If no counter availability image is provided, CUPTI will instead return NaN values for unsupported metrics when evaluating the counter data image.

Resolved Issues

Fixed an issue where repeatedly calling a CUPTI API could result in an overall performance drop. This issue was introduced in the CUDA 13.2 release.
Fixed a memory leak that occurred when tracing a CUDA graph containing memcpy or memset nodes, even when these activities were not enabled.