ZKPU: A Novel NVMe-Based Accelerator for Scaling Zero-Knowledge Proofs in Blockchain Systems

Abstract

Zero-Knowledge Proofs (ZKPs) are becoming a foundational technology for scalable and privacy-preserving blockchain systems, especially through applications like zkRollups. However, the computational intensity of proof generation continues to limit real-world deployment. We present ZKPU, a hardware-software co-designed ZK accelerator that combines native NVMe integration—ensuring seamless compatibility across existing server and edge infrastructure—with a modular RISC-V System-on-Chip (SoC) architecture that opens the path to eliminating host–device communication bottlenecks. ZKPU is designed to flexibly support a wide range of ZK workloads; in this work, we demonstrate its capabilities by implementing and optimizing multi-scalar multiplication (MSM), a core bottleneck in many zk-SNARK systems. Built using the Chipyard framework and equipped with dedicated modular arithmetic units, ZKPU achieves significant performance and energy efficiency improvements over CPU, GPU, and FPGA baselines. Our results highlight ZKPU as a practical and forward-compatible foundation for scalable ZK acceleration in modern decentralized systems.