QOCO GPU Benchmarks

This repository was used to generated the numerical results in the paper titled QOCO-GPU: A Quadratic Objective Conic Optimizer with GPU Acceleration.

To run the benchmarks follow the steps

1. Create a python 3.13 virtual environment

2. Run pip install -r requirements.txt

3. Install CuClarabel: https://www.cvxpy.org/install/index.html?h=cuclarabel. You may have to install it in the PyCall julia environment. Since I use anaconda this required running the following commands:

   First, pin juliacall to a version whose bundled PythonCall matches what the CuClarabel branch requires (PythonCall = "=0.9.31"). Newer juliacall releases pin PythonCall to a different version and make the Julia environment unsolvable, which causes the PythonExt extension (and cupy_to_cucsrmat) to be missing at runtime:

pip install 'juliacall==0.9.31'
python -c 'import juliacall'   # re-resolves the Julia env with PythonCall 0.9.31

julia --project=/home/govind/anaconda3/envs/test/julia_env -e 'import Pkg; Pkg.add(Pkg.PackageSpec(
    url="https://github.com/oxfordcontrol/Clarabel.jl",
    rev="CuClarabel"
))'

julia --project=/home/govind/anaconda3/envs/test/julia_env -e 'import Pkg; Pkg.add("CUDA")'

4. Run ./run_all.sh and the figures will be created in the figures/ directory

Citing

@article{chari2026qocogpu,
  title = {{QOCO}-{GPU}: A Quadratic Objective Conic Optimizer with GPU Acceleration},
  author = {Chari, Govind M and A{\c{c}}{\i}kme{\c{s}}e, Beh{\c{c}}et},
  journal = {arXiv preprint arXiv:2603.29197},
  year = {2026},
}