This repo houses the code the UMich ARM Lab uses to benchmark their GPUs.
These benchmarks require a CUDA-capable GPU, meaning a somewhat recent NVIDIA GPU and driver.
The only software dependencies for running these benchmarks is Python 3.7+ and pipenv. Python 3 should be on your path as python3. Your system installation of Python is sufficient as the benchmarking functions handle virtual environment creation for you. Additionally, pipenv should be installed and on your path.
If pipenv is not installed, install it with python3 -m pip install pipenv.
Initialize the repository with:
./init_repo.shThe script clones the necessary repositories and creates a pipenv environment for running the benchmarks. Note that this will ask you which CUDA version you would like for PyTorch to be installed with.
- Ensure that you have initialized the repository by following instructions in the Installation section.
cdto where this repository is located.- Activate the
pipenvenvironment withpipenv shell. - Run the benchmarks with
python run_benchmarks.py.
To compare the results of your machine with others, run python compare_results.py.
We include the profiling results of a few of our older computers for reference.
However, the comparison script isn't all-encompassing so if you want fine-grained control of plotting, you may desire to make your own script or notebook.