bandwidth-test

CUDA microbenchmarks for measuring bandwidth between global memory and shared memory, plus a Python plotting utility for generating PNG/PDF figures.

What is included

• global2shared.cu: benchmarks global -> shared copies (float, float4).
• shared2global.cu: benchmarks shared -> global copies (float, float4).
• plot.py: merges CSV outputs and generates plots.
• GPU result folders (3090/, 4090/, 5090/, Titan/, a100/) with sample outputs.

Requirements

• NVIDIA GPU with CUDA support
• CUDA toolkit + nvcc
• CMake >= 3.18
• Python >= 3.9

Build

cmake -S . -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build -j

Run benchmark binaries

The binaries write fixed filenames:

• global_to_shared_async_constexpr.csv
• shared_to_global.csv

Run them from a result directory to keep outputs organized:

mkdir -p results
cd results
../build/global2shared
../build/shared2global
cd ..

Generate plots

Install Python dependencies (choose one):

uv sync

or

python -m pip install -e .

Then render plots from benchmark CSVs:

python plot.py --input-dir results --output-dir results

This creates:

• results/png/ (PNG plots)
• results/pdf/ (PDF plots)
• results/csv/plot_data.csv (combined data)

Existing result plots

Merged bandwidth plots

GPU	Plot
RTX 3090
RTX 4090
RTX 5090
Titan
A100

Notes

• plot.py defaults to reading global_to_shared_async_constexpr.csv and shared_to_global.csv from --input-dir.
• CUDA architecture selection is handled in CMakeLists.txt (native when supported by CMake).