Need for zkSpeed: Accelerating HyperPlonk for Zero-Knowledge Proofs

Alhad Daftardar, Jianqiao Mo, Joey Ah-kiow, Benedikt Bünz, Ramesh Karri, Siddharth Garg, Brandon Reagen

Abstract Zero-Knowledge Proofs (ZKPs) are essential for privacy-preserving computations in blockchain and secure voting, yet they are hindered by significant prover overhead. HyperPlonk, a state-of-the-art ZKP protocol, offers universal setup and small proof sizes but introduces intense computational and memory bandwidth challenges. This paper presents zkSpeed, the first hardware accelerator specifically designed for HyperPlonk. By utilizing specialized kernels for SumCheck and Multi-Scalar Multiplication (MSM), zkSpeed achieves an 801× geometric mean speedup over high-end CPUs, making complex ZKPs practical for real-world deployment.

Introduction

Zero-Knowledge Proofs (ZKPs) allow a prover to demonstrate the correctness of a computation to a verifier without revealing sensitive underlying data. Despite their potential for decentralized finance and private identity, ZKPs face a massive "prover bottleneck." Generating a proof for complex circuits can take minutes or hours on standard hardware, limiting their scalability.

While earlier hardware accelerators targeted protocols like Groth16, these often require a "trusted setup" or produce proofs too large for efficient blockchain verification. HyperPlonk has emerged as a "Goldilocks" solution—providing a universal setup and compact proofs. However, HyperPlonk replaces the traditional Number Theoretic Transform (NTT) with the SumCheck protocol. While asymptotically more efficient, SumCheck imposes severe hardware requirements, including massive bitwidths (255-381 bits) and extreme memory bandwidth demands.

The zkSpeed Architecture

To address these challenges, we introduce zkSpeed, a modular ASIC architecture featuring eight unique kernel accelerators. The design focuses on optimizing the most computationally expensive operations in the HyperPlonk workflow.

Overview of the zkSpeed architecture and the four-step HyperPlonk protocol workflow.

The architecture is divided into specialized units:

  1. Unified SumCheck Unit: Rather than implementing separate hardware for every SumCheck variation, zkSpeed utilizes a unified Processing Element (PE). Through High-Level Synthesis (HLS), resources are shared across different protocol stages, reducing silicon area by nearly 50%.
  2. Multifunction Tree Unit (MTU): Polynomial evaluations often lead to "memory explosions" in standard hardware due to Breadth-First Search (BFS) traversal. The MTU implements a hybrid DFS/BFS strategy to maximize on-chip data reuse.
  3. FracMLE and Modular Inversion: zkSpeed accelerates modular inversion—a notoriously slow operation—using Montgomery batching and a constant-time hardware inverter to enable parallel processing.
  4. Memory Compression: Recognizing that approximately 90% of witness data in HyperPlonk is sparse (containing mostly 0s and 1s), zkSpeed employs an on-chip compression scheme that reduces storage requirements by 10–11×.

Technical Innovation: Accelerating SumCheck

The core of HyperPlonk's efficiency lies in the SumCheck protocol, which operates on Multilinear Extensions (MLEs). zkSpeed optimizes the computation of these polynomials by streamlining the evaluation and combination steps.

Detailed breakdown of the SumCheck computation process, illustrating the transition from table-based MLEs to polynomial evaluations.

By shifting the bottleneck from pure computation (NTT) to memory bandwidth (SumCheck), zkSpeed leverages High Bandwidth Memory (HBM3) to provide up to 2 TB/s of throughput. This ensures that the high-performance SumCheck units are not starved for data, a common failure point in standard DDR5-based designs.

Evaluation and Results

We evaluated zkSpeed using a combination of mock circuit workloads and real-world benchmarks, including Zcash transactions, private auctions, and blockchain rollups.

  • Performance: zkSpeed achieves a geometric mean speedup of 801× over a 32-core AMD EPYC CPU baseline.
  • Throughput: A single 366mm² chip can generate a proof for a 1-million-gate circuit in 11 to 86 milliseconds, depending on the specific workload.
  • Comparison: Compared to previous accelerators like NoCap, zkSpeed offers a three-order-of-magnitude reduction in proof size for a similar silicon area, making it significantly more viable for decentralized networks like Ethereum.
  • Power Efficiency: The design operates at approximately 170W, maintaining a power density comparable to modern server-grade CPUs.

Conclusion

zkSpeed demonstrates that "algorithm-aware" hardware design is essential for the next generation of ZKP protocols. By specifically targeting the memory and computational patterns of HyperPlonk, zkSpeed removes the prover bottleneck that has long hindered the adoption of privacy-preserving technologies. Future work will focus on extending support for higher-degree gates (Jellyfish) and further optimizing the memory wall for non-HBM systems.

Summary updated

BibTeX
@inproceedings{daftardar2025need,
  author = {Alhad Daftardar and Jianqiao Mo and Joey Ah-kiow and Benedikt Bünz and Ramesh Karri and Siddharth Garg and Brandon Reagen},
  title = {{Need for zkSpeed: Accelerating HyperPlonk for Zero-Knowledge Proofs}},
  booktitle = {ISCA},
  year = {2025},
  doi = {10.1145/3695053.3731021},
}
Figure