Authors: Xinwei Qiang, Liukun Yu, Xiyu Wang, Zhengyi Li, Shixuan Sun, Jingwen Leng, Chen Chen, Jiaping Gui, and Zhenzhe Zheng (Shanghai Jiao Tong University); Jin Dong (Beijing Academy of Blockchain and Edge Computing); and Minyi Guo (Shanghai Jiao Tong University)
Abstract: Zero-knowledge proof (ZKP) protocols enable privacy-preserving verification but face critical performance bottlenecks at the proof generation stage, particularly when handling large-scale computations that exceed GPU memory capacity. We present FluxZK, a framework that addresses scalability and efficiency challenges through ZKP-specific GPU-based optimizations. Our framework introduces FluxMSM, an efficient multi-scalar multiplication (MSM) implementation that leverages fine-grained parallelism and overlapped data transfer with computation. Additionally, we develop FluxNTT, a high-performance number theoretic transform (NTT) solution supporting both out-of-core and on-GPU computation paradigms. Extensive evaluations on data center GPUs demonstrate that FluxZK scales well to large input sizes of up to 2³⁰ elements, delivering an estimated performance improvement of up to 2.16x over existing solutions in real-world ZKP applications.
Our code is open-sourced at https://github.com/SJTU-Liquid/FluxZK.
Presentation: file