Is there a way in NVIDIA Nsight Systems to limit threads displayed?
I have a project which has thousands of threads, but I want to use the Nsight System to profile the CUDA code. However, loading the report takes a while which I believe is due to the high number of thread information, in addition to all the visual clutter of those threads which I don't currently care about information on.
Is there a way to toggle collecting thread information or limit it before loading a report in the Nsight System GUI?
Is there a way to toggle collecting thread information?
If profiling through the CLI
Check the -s/--sample and --cpuctxsw options, for the profile or start commands, link to documentation. You can set both to none, to minimize the amount of information collected from the CPU side.
If profiling a Linux target: check also the -t/--trace option for the profile or launch commands. Essentially you would like to exclude osrt from the trace options, it is enabled by default.
If you want to collect only CUDA events, then you can use nsys profile -t cuda -s none --cpuctxsw=none <app>.
If profiling through the GUI
You can deselect the "Collect CPU IP/backtrace samples" and "Collect CPU context switch trace" boxes.
If profiling a Linux target: you can additionally deselect the "Collect OS runtime libraries trace" box.
Is there a way to limit it before loading a report in the Nsight System GUI?
If the data is collected, it is not possible to exclude it from rendering on the GUI. You can minimize threads, or hide them by right clicking on "Threads" -> "Show less".
- CUDA incompatible with gcc version
- Does thrust::device_ptr take over the lifetime of the object it points to?
- CUDA code runs when compiled with sm_35, but fails with sm_30
- what's cga in cuda programming model
- Proper way to cast 'threadIdx.x's into higher type in CUDA kernel (%lu format in printf malfunctions in the CUDA kernel?)
- CUDA thread mapping
- cublasSgemm row-major multiplication
- How do I select which GPU to run a job on?
- how many processors can I get in a block on cuda GPU?
- How can I debug code 700 "illegal memory access" aka `CUDA_EXCEPTION_14, Warp Illegal Address`?
- Do I need to to mark a global variable used in a kernel as __device__?
- Are global device-side variables in CUDA bad practice?
- how to set the cuda path in the conda environment?
- Kernel accessing device-allocated struct does not print
- nvcc error: cuda/std/variant: No such file or directory
- initialize subclass of cuda::std::variant
- PyCharm 2025.1 Jupyter kernel fails on import torch with ImportError: libnccl.so.2, but same virtual-env works in terminal
- ImportError: libcuda.so.1: cannot open shared object file
- What is the easiest thing i can do to compile C++ CUDA code on Arch linux to Windows .exe?
- How does cudaMallocPitch help avoid bank conflict?
- CUDA driver version is insufficient for CUDA runtime version
- Why is my call of the CUDA math library sqrt() function failing?
- Using cublas<T>gemmStridedBatched with row major matrices
- How do I use atomicMax on floating-point values in CUDA?
- Calculating FLOPS (Floating-point Operations per Seconds)
- How does CUDA exactly synchronize threads in a warp at barriers and conditional expressions?
- When should I prefer write-combined CUDA-allocated mapped host memory?
- Does CUDA really not have a calloc()-like API call?
- Do two atomicAdds in different threads in the same warp always happen simultaneously?
- What is the correct way to perform 4D FFT in Cuda by implementing 1D FFT in each dimension using cufftPlanMany API