Year: 2026
Type: article
Abstract: Zero-knowledge proofs (ZKPs) enable verifiable computation without exposing private data and are increasingly deployed in blockchains and privacy-preserving machine learning. A major computational bottleneck in modern ZKP systems arises from algebraic hash functions and Merkle tree operations, whose heterogeneous round structures and irregular data dependencies expose limited parallelism and conflict-prone access patterns that are poorly aligned with the execution model of modern GPUs. Existing GPU frameworks accelerate specific hashes or static Merkle constructions, yet lack a unified execution model for structurally diverse hash designs and irregular dynamic Merkle workloads. We present MARS , a GPU acceleration framework for hash-intensive ZKP workloads. MARS introduces a shared operator-level abstraction for algebraic hashes, expressing diverse round structures as reusable algebraic operator skeletons and materializing static execution plans through compile-time specialization. It further proposes an index-aligned warp-cooperative execution model that preserves uniform control flow under asymmetric round designs and sustains high SIMD utilization across heterogeneous hash structures. It also builds a conflict-aware Merkle processing pipeline that decouples path extraction from depth-wise hash updating to handle irregular dynamic updates efficiently. Evaluation shows that MARS improves ZK-friendly hash throughput by up to 1.71 × and Merkle tree update performance by 6.99 × on average compared to state-of-the-art GPU implementations. It also offers a 1.36 × speedup over a multi-core CPU baseline for a production zkEVM prover.
DOI: 10.1145/3864374
Language: English
Subfield: Hardware and Architecture
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
Award: G618317349
OpenAlex ID: W7236218072
