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 13.5. You may obtain the latest version of CUPTI by Downloading the CUDA Toolkit 13.5.0

Download the CUDA Toolkit 13.5 Now
Download the CUPTI Python API 13.5.0 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 13.5

New Features

Added an activity attribute CUPTI_ACTIVITY_ATTR_HES_BUFFER_SIZE to control the buffer size for storing timestamps for HES. This attribute can be set using cuptiActivitySetAttribute and cuptiActivitySetAttribute_v2 APIs.
Added Range Profiling support for the metrics sm__inst_executed_op_pmtrig_<N>_pred_on_any and smsp__inst_executed_op_pmtrig_<N>_pred_on_any, where N ranges from 0 through 7. Each metric counts PMTRIG instructions executed by a warp with at least one active, predicated-on thread. These metrics expose the legacy prof_trigger counters at the SM and SM subpartition levels.
Added CUDA graph correlation fields to CUDA event activity records. The fields graphId and graphNodeId identify the executable graph and event-record node, and sourceGraphId and sourceGraphNodeId identify the source graph and node. CUpti_ActivityCudaEvent2 is deprecated and replaced by CUpti_ActivityCudaEvent3.
Added tracing support for batched memcpy nodes in CUDA graphs, including graph-executable and node-parameter updates, converted nodes, device-launched graphs, and both HES and software tracing paths.
On Windows, the CUPTI DLL name now uses the CUDA major version, for example cupti64_13.dll. The stable major-version name allows an application to use later CUPTI releases with a compatible ABI without relinking for every CUDA minor release.
Added support for CUDA coredump collection while CUPTI tracing is active.
On GB20B/C, C2C peak-utilization metrics using pct_of_peak_sustained_elapsed are not available and report NaN. Use the C2C user-data byte counters and per-second metrics for traffic and bandwidth values.

Resolved Issues

Reduced host CPU overhead for CUDA Graph node-level tracing in multi-GPU and multi-host workloads by eliminating lock contention.
Enforced the ordering and session-sticky behavior of CUPTI_ACTIVITY_ATTR_ENABLE_MULTI_SUBSCRIBER_GRAPH_TRACE. In multi-subscriber mode, the attribute must be set before any activity kind is enabled, and graph trace policy and mode now remain active until cuptiFinalize.
Fixed zero-valued kernel timestamps delivered to a second subscriber when HES tracing is active.
Fixed missing channel IDs in activity records for CUDA graph kernel, memcpy, and memset nodes collected using HES tracing.
Fixed crashes and incorrect records when a CUDA graph node changes implementation after a graph executable or node-parameter update, including batched memcpy and memset nodes.
Fixed incorrect byte counts and source or destination memory kinds for batched memcpy activities produced from converted CUDA graph nodes.
Fixed a null-pointer crash when activity kinds were disabled and re-enabled concurrently with memory operations.
Fixed a hang that could occur when CUPTI attached late to an application while multiple application threads were launching kernels concurrently.
Fixed software tracing that reported a conditional-graph memset operation as a kernel activity.
Fixed a deadlock when PC sampling start, stop, and data collection ran concurrently with kernel launches in another thread on the same CUDA context.
Fixed several failures, crashes, and hangs when tracing CUDA graph workloads on Blackwell GPUs.