NVIDIA CUDA Profiling Tools InterfaceCUPTI
A set of NVIDIA APIs for profiling CUDA C++, PTX and SASS code running on GPUs, with CPU–GPU timestamp synchronization.
The NVIDIA CUDA Profiling Tools Interface (CUPTI) is a set of APIs for profiling the execution of CUDA C++, PTX and SASS code on GPUs.
Key idea CUPTI synchronizes timestamps across the CPU host and the GPU device, so host and device activity can be placed on one timeline.
Profilers build on CUPTI rather than using it directly. Examples include the Nsight Systems profiler and the PyTorch Profiler.
See the CUPTI documentation. For profiling GPU applications on Modal, see this example from Modal's documentation.
Related#
- Nsight Systems: a GUI profiler built on CUPTI
- Performance: why profiling matters for GPU programs