2026-05-09 –, ETHRes
Zero-knowledge proofs are now operational infrastructure for rollups, private payments, verifiable execution, and zkVM-based applications, yet published benchmark results often compare different workloads, different proof modes, and different hardware paths. This paper presents a unified host--agent framework for benchmarking zero-knowledge proving systems under a common execution and reporting methodology. Each prover is isolated in a containerized agent, selected through a scheme-adapter layer, and required to report canonical JSON artifacts linked to a run manifest and to workload-specific semantic contracts. We use the framework to compare representative circuit-based, transparent, GPU-accelerated, folding, and zkVM systems: Groth16, ICICLE, Halo2, Plonky3, Winterfell, Nova, RISC Zero, RISC Zero CUDA/Groth16 variants, SP1, and SP1 without AVX. The benchmark suite separates six primary workloads with comparable semantics from secondary hash workloads that expose specialization and acceleration effects. Across the primary workloads, no prover family dominates all deployment criteria. At 8 CPU cores, Plonky3 and Winterfell achieve the lowest median proving latency and memory footprint; Groth16 and ICICLE produce sub-kilobyte proofs and the lowest median verification latency; and zkVM systems trade larger proofs and higher memory use for programmability. GPU acceleration is most beneficial once fixed initialization and setup costs are amortized, while CPU-oriented provers show strong scaling from 1 to 8 cores. The results support a deployment-oriented view of ZKP selection: benchmarks should disclose security profiles, distinguish comparable from specialization-driven workloads, and report resource limits as primary outputs rather than treating proof generation time alone as the ranking criterion.
