nvprobe

NVIDIA GPU benchmark suite — automate CUDA workloads, run HPL & HPCG, profile custom kernels, compare hardware, and generate rich reports.

Python 3.10+ CUDA 12 / 13 HPC
View on PyPI View on GitHub

Why nvProbe?

📊

Pre-built Benchmarks

Memory bandwidth, latency, matrix multiply, attention — run standard CUDA workloads with zero setup.

🏎️

HPL & HPCG

Self-contained Linpack and CG benchmarks via NVIDIA HPC Benchmarks binaries — downloads and runs automatically.

🧠

MLPerf Inference

Run MLPerf inference benchmarks with configurable mode, batch size, and precision.

📈

HTML Reports

Generate interactive HTML reports with charts comparing GPUs, precisions, and problem sizes.

📦

Bundled CUDA Runtime

No system CUDA toolkit required — use cupy-cuda13x[ctk] to bundle CUDA libs with pip.

🖥️

Slurm Integration

Submit, monitor, and collect results from HPC clusters — all from a single CLI.

Quick Start

Install

pip install --user nvprobe

Initialize config

nvprobe init

Run benchmarks

nvprobe run --local

Generate report

nvprobe report --open

Benchmarks

Bandwidth

Memory read/write, copy, stride, pointer-chase latency

CUDA C

Custom Kernels

MatMul, tiled MatMul, fused attention, reduction

CuPy

HPL

High Performance Linpack (FP64)

MPI Datacenter GPUs

HPCG

High Performance Conjugate Gradients

MPI Datacenter GPUs

MLPerf

MLPerf Inference (test / find_performance / full)

TFLite