Omni: A Scalable Multi-FPGA Number-Theoretic Transform Accelerator with Streaming Global Transpose for Large-Scale Zero-Knowledge Proofs
Student Contest:
No
Affiliation Type:
Academia
Keywords:
FPGA Accelerator, Number Theoretic Transform, Zero Knowledge Proof, Multi FPGA System
Abstract:
Omni is a scalable multi-FPGA number-theoretic transform (NTT) accelerator for large-scale zero-knowledge proof generation. As polynomial domains approach 2^30, single-FPGA designs are constrained by off-chip memory capacity and host-device transfers, motivating intra-NTT parallelism across nodes. Omni turns the unavoidable global transpose into a streaming execution path on eight Alveo U280 FPGAs. Its distributed flow overlaps inter-node ownership redistribution and intra-node layout conversion with local computation. One-to-One Permutation Synchronization (OPS) maintains contention-free one-to-one transfers to improve aggregate bandwidth, while Communication-Aware Streaming Layout (CASL) embeds layout permutation into normal HBM accesses. The 256-bit, 180-MHz prototype completes a 2^30-point NTT in 1.469 s, achieves 3.96× strong scaling from 2 to 8 nodes, and reaches 32.89 GOP/s, outperforming reported FPGA NTT accelerators by 4.8–11.7×.
