nvprof

nvprof is a command-line profiler for CUDA applications, included with the NVIDIA CUDA Toolkit. It collects performance data from the GPU and the host during program execution, focusing on kernel execution times, memory transfer activity, and CUDA API calls. The tool can produce both a high-level summary and a detailed kernel-by-kernel report, and it can generate a timeline-style trace for post-processing.

Usage is typically nvprof [options] executable [args]. Options allow you to specify what to measure, commonly using --metrics or --events, and to enable API tracing or kernel tracing. Output can be written to stdout or to a log file via --log-file. It can profile multiple processes and multi-GPU workloads and can profile child processes with appropriate options.

Output and interpretation: The result includes per-kernel execution times, memory copy timings, and a breakdown of time spent in CUDA API calls. The generated report helps identify bottlenecks, while the timeline trace enables correlation of events over time. Developers can use it to guide optimizations at the kernel or memory subsystem level.

Status and limitations: nvprof adds runtime overhead and provides a more coarse-grained view compared to modern tools. It is designed for CUDA workloads and does not cover non-CUDA GPU activities. In recent CUDA releases nvprof has been deprecated in favor of NVIDIA Nsight Compute and Nsight Systems, with NVIDIA recommending migration for new projects.

See also: CUDA Toolkit, Nsight Compute, Nsight Systems.