ParEval Leaderboard: Evaluating the Ability of Large Language Models to Generate Parallel Code

We introduced the ParEval benchmark in “Can Large Language Models Write Parallel Code?” to evaluate the capability of LLMs at parallel code generation. We found a significant gap between their ability to generate sequential vs parallel code for a large array of computational problems and parallel programming models. On this page we keep an up-to-date table tracking the progress of state-of-the-art LLMs on ParEval.

ParEval Results

Model	No. Parameters	HumanEval pass@1	ParEval Serial pass@1	ParEval Parallel pass@1
StarCoder2-3B	3B	31.7	42.7	9.6
StarCoder2-7B	7B	35.4	59.4	15.9
CodeLlama-7B	7B	29.9	48.4	15.3
CodeLlama-13B	13B	35.0	52.8	17.4
StarCoder2-15B	15B	46.3	61.6	23.1
StarCoderBase	15.5B	30.3	51.7	18.6
CodeLlama-34B	34B	45.1	54.0	10.2
Phind-V2	34B	71.9	65.6	32.1
Gemini-Pro	—	67.7	59.3	25.1
GPT-3.5	—	61.5	76.0	39.6
GPT-4	—	84.1	76.1	37.8

Last updated March 5, 2024

If you would like a model added you can reach out to dnicho@umd.edu.

Citing ParEval

@misc{nichols2024large,
      title={Can Large Language Models Write Parallel Code?}, 
      author={Daniel Nichols and Joshua H. Davis and Zhaojun Xie and 
              Arjun Rajaram and Abhinav Bhatele},
      year={2024},
      eprint={2401.12554},
      archivePrefix={arXiv},
      primaryClass={cs.DC}
}