ezkl
ezkl: Zero-Knowledge Proofs for Deep Learning (open-source toolkit)
Reported benchmarks
| model | params | proving time s | proof size mb | verification time s | peak ram gb | proving key gb | verification key mb |
|---|---|---|---|---|---|---|---|
| MNIST MLP | 2M | 1310 | 0.127 | 5.4 | 21.15 | 8.3 | 4.1 |
| LeNet5 | — | 535 | 0.127 | 2 | 5.45 | 2.1 | 1.1 |
| VGG11-mini | — | 2650 | 0.478 | 9.7 | 41.85 | 17 | 8.1 |
| ResNet18 | — | 6840 | 0.18 | 25.3 | 103.5 | 34 | 17 |
| MobileNetV2 | — | 14320 | 0.175 | 40.6 | 203.95 | 76 | 37 |
proving_time_note: Four hours to prove one MobileNetV2 forward pass, in 204 GB of RAM, for a proof that takes 40 seconds to verify. Bionetta proves the same network in 24.5 seconds on the same machine and 23 seconds on an iPhone. That is the widest single gap in this corpus, and it is not a protocol result -- most of it is that Bionetta's model is public, so its weights are circuit constants and its linear layers cost nothing.
Notes
Not a paper, but the de facto baseline that Jolt Atlas, NANOZK, Bionetta and others benchmark against. On nanoGPT (per Jolt Atlas): 237s proving, >400s key generation, 0.34s verify. WHAT THE BIONETTA NUMBERS DO TO THE STANDARD STORY ABOUT HALO2. The received account -- ours included, on the proof-systems page -- is that PLONKish is the slow-prover/fast-verifier option, and that you tolerate its prover because kilobyte proofs and millisecond verification are what make on-chain settlement possible at all. As configured by ezkl and measured by Bionetta, that second half is simply not true: the proofs run to hundreds of kilobytes, the verification key reaches tens of megabytes, and VERIFICATION TAKES SECONDS TO TENS OF SECONDS. None of that goes on a chain. Halo2 can of course be configured for small proofs; the point is that the toolchain people actually deploy is not, and nobody had published the numbers. Groth16 (Bionetta, ZEN, vCNN) is the line that actually delivers the constant-size proof and the cheap verifier -- at the price of a per-circuit trusted setup.
Our reading
Citation neighbourhood
Builds on
- none recorded
Cited by, in this corpus
Edges are a proxy: paper A's text mentions B anywhere (body or bibliography). See the full graph.
Discussed in
Other recorded fields
numbers_note: EVERY NUMBER BELOW WAS MEASURED BY BIONETTA (arXiv:2510.06784v2, Table 4), not by ezkl. Bionetta picked the models, picked the hardware, and is the system being flattered by the comparison. They are recorded here because they are the ONLY primary-source ezkl benchmarks in this corpus -- ezkl has no paper -- and because a baseline that everyone cites and nobody measures is exactly how a field ends up with a folk ranking. Cite them as "Bionetta measured ezkl at ...", never as "ezkl reports ...".