On the Performance of Cloud-based ARM SVE for Zero-Knowledge Proving Systems
This work addresses the significant performance gap of ARM-based cloud servers versus x86-64 under zero-knowledge proof (ZKP) workloads—particularly in critical paths such as Merkle tree construction. We systematically identify the root cause: the 128-bit vector width and lower clock frequencies inherent to current ARM implementations. To overcome this, we propose and empirically validate a 512-bit scalable vector extension (SVE/SVE2)-based acceleration framework, integrating optimized Poseidon hashing, Goldilocks field arithmetic, and parallelized Merkle tree construction. On AWS Graviton4 and Axion instances, our approach achieves a 1.4–1.6× speedup over AVX-512-optimized x86-64 baselines. Theoretical analysis confirms that 512-bit SVE can surpass x86-64 in ZKP throughput while sustaining >10% cost advantage. This work establishes the first systematic performance modeling and vectorization optimization paradigm for ARM architectures in production-scale ZKP infrastructure.