NVIDIA CUDA Profiling Tools Interface

The NVIDIA CUDA Profiling Tools Interface (CUPTI) provides a set of APIs for profiling execution of CUDA C++ , PTX , and SASS code on GPUs. Critically, it synchronizes timestamps across the CPU host and the GPU device.

CUPTI's interfaces are consumed by, for example, the NSight Profiler and the PyTorch Profiler .

You can find its documentation here .

For details on using profiling tools for GPU applications running on Modal, see this example from our documentation .