Benchmarks code mirrors the reference implementation by Microsoft Research – https://github.com/microsoft/Nova/tree/main/benches.
Benches comparison-results:
| StepCircuit size | Microsoft-Prove | Nexus-Prove | Microsoft-Verify | Nexus-Verify |
|---|---|---|---|---|
| 0 | 44.201 ms | 221.17 ms | 36.907 ms | 108.81 ms |
| 6399 | 53.304 ms | 229.65 ms | 46.401 ms | 116.76 ms |
| 22783 | 74.887 ms | 244.27 ms | 69.095 ms | 143.06 ms |
| 55551 | 120.05 ms | 277.11 ms | 116.09 ms | 192.72 ms |
| 121087 | 196.03 ms | 346.58 ms | 186.52 ms | 282.84 ms |
| 252159 | 335.71 ms | 498.95 ms | 319.76 ms | 476.57 ms |
| 514303 | 630.94 ms | 776.78 ms | 596.13 ms | 824.23 ms |
| 1038591 | 1.1913 s | 1.3661 s | 1.1284 s | 1.5852 s |
Both were executed on the same hardware, commits used:
- Nexus – https://github.com/nexus-xyz/nexus-zkvm/tree/6045c596f2e136aca58e248c993d85a370f983f9
- Microsoft – https://github.com/microsoft/Nova/tree/4f8f3e782b172e98d6d741b29e5bc671ab5b93a6
Results may vary based on your hardware, but the performance ratio should look similar. Microsoft implementation has significantly lower recursion overhead, and as the step circuit size grows, the gap becomes smaller from ~3-4x to 20-30% slowdown.
Hardware used is Ryzen 5900x with 64gb 3600MHz RAM, GPU acceleration disabled.
From the nova-benches directory
cargo bench --bench=recursive-snarkFlamegraphs are generated with
cargo bench --bench=recursive-snark -- --profile-time=10
# saves to nova-benches/target/criterion/RecursiveSNARK-StepCircuitSize-*/[Prove/Verify]/profile/flamegraph.svg