blog-triton-profiling

Repo containing blog demo materials for profiling Triton GPU kernels on

• CUDA
• ROCm

=======

Triton Profiling

Triton profiling introduction demo materials and runtime container images.

Requirements

• NVIDIA GPU for the NVIDIA blog
• AMD GPU for the AMD blog
• Podman or Docker
• make

NOTE: ROCm's Compute profiler only runs on AMD CDNA GPUs, i.e. MI300X (02-12-2026).

Images

Container images that provide a runtime environment for the target GPU

• cuda - Requires an NVIDIA GPU
• nsight - Does not require an NVIDIA GPU
• rocm - Requires an AMD GPU (i.e. MI300X)

NOTE: The nsight target only provides an environment to run the Nsight tools.

Instructions

Build

With an NVIDIA GPU

make cuda-image

Without an NVIDIA GPU

make nsight-image

With an AMD GPU

make rocm-image

Runtime

All targets will start a new container and remove it when you exit.

CUDA/Nsight

Console

Runs the target image and leaves the user inside it at a bash prompt.

make [cuda | nsight]-console

Jupyter Notebook

The nsight image can only be used to view the Jupyter notebook, an NVIDIA GPU and the cuda image are required to run it.

make [cuda | nsight]-jupyter

Nsight Systems

Runs the Nsight Systems UI.

make [cuda | nsight]-systems

Nsight Compute

Runs the Nsight Compute UI.

make [cuda | nsight]-compute

AMD ROCm

Console

Runs the target image and leaves the user inside it at a bash prompt.

make rocm-console

Jupyter Notebook

make rocm-jupyter