/jetson-mem-bench

CPU and GPU memory bandwidth benchmark for NVIDIA Jetson and CUDA platforms. STREAM-like Copy and Triad kernels, solo vs contended runs, automatic sizing, CSV/JSON/Markdown output, and theoretical vs sustained performance comparisons.

Primary LanguagePythonMIT LicenseMIT

Jetson Memory Bandwidth Benchmark

STREAM-style CPU/GPU memory bandwidth benchmark for Jetson and CUDA devices.

Measure sustained DRAM bandwidth using STREAM-like Triad and Copy kernels for both CPU (OpenMP) and GPU (CUDA).
Reports solo vs contended performance, GiB/s vs GB/s, theoretical vs sustained, and exports CSV/JSON/Markdown results.

Quick Start

# Clone and build
git clone https://github.com/jetsonhacks/jetson-mem-bench.git
cd jetson-mem-bench
make -j

On Jetson before running tests:

sudo nvpmodel -m 0     # performance mode
# Turn off Dynamic Voltage and Frequency Scaling; Clocks full speed
# to restore and turn DVFS back on : sudo jetson_clocks --restore
sudo jetson_clocks     
export OMP_NUM_THREADS=$(nproc)
export OMP_PROC_BIND=close
export OMP_PLACES=cores

Run tests:

# Run default Triad benchmark
python3 tools/bench.py

# Run best-case Copy benchmark
python3 tools/bench.py --op copy

Platforms

  • Jetson AGX Thor (SM=110, theoretical 273 GB/s)

  • Jetson AGX Orin (SM=87, theoretical 204.8 GB/s)

  • Jetson Orin Nano (SM=87, theoretical 68 GB/s)

Also runs on x86 + CUDA-capable GPUs

More Information

See the User Guide for details on:

  • How the benchmarks work

  • Units (GiB/s vs GB/s)

  • Solo vs contended interpretation

  • CLI parameters and tuning knobs

  • Example results and common pitfalls