Benchmarks

Looking for the indicator library's numbers?

This page is about Wickra ZK. Wickra's own indicator benchmarks — the comparison against TA-Lib, talipp, pandas-ta and the other Rust TA crates — live at wickra.org.

Measured by the wickra-zk-bench Criterion suite (crates/wickra-zk-bench/benches/prove.rs) on the momentum golden case -- a fast/slow SMA cross over sym-01, 48 OHLCV bars -- on a GitHub-hosted ubuntu-latest runner (4 vCPU, CPU prover, no GPU). Medians from the nightly bench.yml run of 2026-09-15, run 35007405126; the whole job took 3 h 39 min, build included.

Measurement	What it times	Samples	Median
exec/momentum	Dev-mode proving: the guest runs the backtest inside the zkVM and produces the journal, no real proof (RISC0_DEV_MODE=1).	100	237.4 ms
prove/momentum	Production proving: a real RISC Zero receipt for the same case, in-process on the CPU.	10	1056.6 s (17 min 37 s)
verify/momentum	Checking that receipt against the pinned guest id (WICKRA_ZK_GUEST_ID).	100	19.09 ms

Proving dominates by four orders of magnitude over execution and five over verification, which is the shape a zkVM proof is supposed to have: the backtest itself costs a quarter of a second inside the guest, producing the proof of it costs minutes of STARK arithmetic, and anyone can check the result in twenty milliseconds without the data or the strategy.

Reading the numbers

• Execution is what a caller pays on every prove call before proving starts, and what the dev-mode tests pay in full. It scales with the number of bars and the indicators the strategy names.
• Proving scales with the guest's cycle count in segments; a longer history or a heavier strategy adds segments and proving time roughly linearly. The CPU prover on a hosted runner is the floor, not the ceiling: a machine with more cores, or the CUDA / Metal backends risc0 ships, proves the same case several times faster. The receipt is the same bytes either way.
• Verification does not depend on the history length or the strategy; it is the cost of checking the receipt's STARK, and it is what a counterparty runs.

Reproducing

# execution only, seconds:
cargo bench -p wickra-zk-bench
# execution, proving and verification, hours (one real proof per sample):
WICKRA_ZK_BENCH_PROD=1 cargo bench -p wickra-zk-bench

The prover needs clang (CC=clang CXX=clang++; the C++ kernels take minutes with g++ and seconds with clang) and no Windows host. bench.yml runs the production suite nightly under a six-hour budget: Criterion's minimum of ten samples at seventeen minutes a proof, plus one warm-up proof and the receipt the verify bench checks, is close to three and a half hours.

The numbers above are the ones in the repository's BENCHMARKS.md, measured with the commands it names.