GPTX: Comparing multi-processing features of Rust and CPP using GPT-2
[ View on Github ]
In this project, we benchmark the multiprocessing features of Rust and CPP and report the metrics as we increase threads. We use Rayon+Rust and CPP+OpenMP for these experiments.
Machine Details
We benchmark on the NYU HPC crunchy-5 machine with the configuration of 64 cores, 256GB memory.
Observations
- Rust achieves the best peak tokens-per-second highlighting the power of Rust optimizations along with Rayon’s multi-processing.
- CPP scales more gracefully, and improves speeds even as threads/cores > 1.
Results