HackMDPonos Technology’s HW-Accelerated Plonky2 GoldiBear: Unleashing Plonky2 Over BabyBear Throughput
With the emergence of recursive proving and the convergence of finite fields to smaller sizes, Zero-Knowledge (ZK) proving is making significant strides in reducing latency, increasing throughput, and cutting costs.
But why we would settle for less?
By integrating highly efficient hardware acceleration, we can push these improvements even further. Ponos Technology once again showcases how our cutting-edge FPGA designs can take ZKP performance to unprecedented heights!
Telos R&D team has successfully adapted Polygon Plonky2 to the BabyBear field for the recursive circuit, unlocking faster and more efficient zero-knowledge proofs (ZKPs) aggregation. Building on that foundation, Ponos Technology has introduced HW-accelerated Plonky2 GoldiBear, a groundbreaking solution that merges hardware acceleration with optimized software to deliver unprecedented performance in NTT and Poseidon2-based Merkle tree computations.[1]
A Two-Pronged Innovation
-
HW-Accelerated Plonky2 GoldiBear for the AMD Alveo U55C
Ponos has custom-built a pipeline on the AMD Alveo U55C FPGA, targeting BabyBear field operations for high-throughput NTT/INTT and parallelized Poseidon2 hashing. These specialized circuits operate in tandem with on-chip buffering to minimize latency and cut down polynomial processing times to a fraction of the CPU baseline. -
Zero-Overhead Software Memory Management
By reshaping the Plonky2 data-transfer routines, Ponos ensures polynomial coefficients reach the FPGA and inverse NTT results as well as Poseidon2 hashes reach the host with virtually no overhead. This fluid host-device data flow underpins the remarkable gains in efficiency.
Benchmark Highlights
Measuring polynomial-batch-intt-ext-ntt-merkelize at batch sizes of 16, 40, and 334 polynomials illustrates the dominance of HW-accelerated Plonky2 GoldiBear:
| Batch Size | Software (ms) | FPGA (ms) | Speedup |
|---|---|---|---|
| 16 | ~12.9 | ~0.32 | ~40X |
| 40 | ~25.6 | ~0.90 | ~28X |
| 334 | ~190 | ~12.2 | ~15–16X |
Across all sizes tested, HW-accelerated Plonky2 GoldiBear surpasses CPU performance by 15X to 40X, underscoring its effectiveness and scalability.
Unparalleled FPGA-Driven ZKP Acceleration
By tightly integrating Telos’ Plonky2 (adapted to the BabyBear field) with Ponos Technology’s tailored FPGA acceleration, HW-accelerated Plonky2 GoldiBear positions itself as a powerful solution in the ZK proof aggregation landscape. Memory and pipeline optimizations ensure minimal transfer overhead, transforming once time-consuming polynomial and Merkle operations into swift computations.
Conclusion
The HW-accelerated Plonky2 GoldiBear from Ponos Technology isn’t just a high-speed outlier—it’s yet another hugely successful HW-accelerated ZK prover backend, standing alongside the company’s unmatched price-performance Polygon Hermez zkEVM FPGA accelerator and its Aleo MSM FPGA accelerator (featuring significant software prover backend parallelization).
