cuda-sumcheck/README.md

Overview

This project aims to implement Sumcheck protocol to run on GPU via CUDA, following the algorithms described in the paper. Currently algorithm 1 is implemented and tested with bn254 scalar field. Future plan is to implement algorithm 3 and support small prime fields such as Goldilocks.

Only kernels are written in C++ and uses Rust to launch kernels and handle device datas by using bindgen and cudarc.

Environment

The implementation has been tested on the following spec.

OS

Ubuntu 22.04.4 LTS

GPU

1 NVIDIA A10G Tensor GPU, 24GB Ram

CUDA version

• nvcc version: 12.1
• C++ compiler version: clang-16
• C++ language standard: C++20

Architecture