Paper | Benchmarks

Stwo

⚠️ Work in Progress

Stwo is currently under active development and is not recommended for production use.

Table of Contents

• About
• Key Features
• Getting Started
  • Prerequisites
  • Building
  • Running Examples
• Project Structure
• Real-World Applications
• Benchmarks
• License

🌟 About

Stwo is a next-generation implementation of a Circle STARK (CSTARK) prover and verifier framework, written in Rust 🦀. While it includes example implementations for demonstration purposes, the primary purpose of this repository is to serve as a foundational framework for building more complex, production-grade provers.

The framework provides a high-performance, flexible foundation for generating and verifying STARK proofs using the latest cryptographic research and innovations in STARK proof systems. It is designed to be extended and customized for specific use cases beyond the simple examples included in this repository.

🚀 Key Features

• Circle STARKs: Implementation based on the latest cryptographic research and innovations in STARK proof systems
• High Performance: Designed for extreme speed and efficiency with optimized SIMD operations

🛠️ Getting Started

Prerequisites

• Rust Toolchain: This project requires Rust nightly (see rust-toolchain.toml for the specific version)

Building

Clone the repository and build the project:

git clone https://github.com/starkware-libs/stwo.git
cd stwo
cargo build --release

Running Examples

The project includes example implementations in the crates/examples directory:

# Run examples
cargo run --release --example <example_name>

📁 Project Structure

This is a Rust workspace containing multiple crates:

• crates/stwo - Core library implementing the Circle STARK prover and verifier
• crates/constraint-framework - Framework for expressing and evaluating constraints
• crates/air-utils - Utilities for working with Arithmetic Intermediate Representation (AIR)
• crates/air-utils-derive - Derive macros for AIR utilities
• crates/examples - Example implementations demonstrating basic usage
• crates/std-shims - Standard library shims for no_std compatibility

🌍 Real-World Applications

While the examples in this repository demonstrate basic usage patterns, Stwo is designed as a framework for building sophisticated provers that handle complex real-world scenarios. A prominent example of such an application is:

stwo-cairo

stwo-cairo is a production-oriented prover built on top of the Stwo framework that enables proving Cairo program executions.

📊 Benchmarks

Quick Benchmark

Run a single-threaded Poseidon2 hash proof benchmark:

./poseidon_benchmark.sh

Comprehensive Benchmarks

Run all benchmarks using Cargo:

cargo bench

Note: to keep benchmark output clean from optional println! noise, set STWO_QUIET=1 when running benchmarks, for example: STWO_QUIET=1 cargo bench ....

Benchmark Reports

Visual representation of benchmark results can be found at: https://starkware-libs.github.io/stwo/dev/bench/index.html

GPU Performance

Reference Machine

• 1 * NVIDIA GeForce RTX 4090
• CPU: AMD EPYC 9224 with 16 cores
• Memory: 94GB
• CUDA Toolkit: 13.0.1_580.82.07

Wide-Fibonacci Test

• Blake2s commit channel
  • Tests (Prove)
    • SIMD: RUSTFLAGS="-C target-cpu=native -C opt-level=3" MIN_LOG=16 MAX_LOG=24 RUST_LOG=info RAYON_NUM_THREADS=16 cargo test test_wide_fib_prove_with_blake_simd --release --features parallel -- --nocapture
    • GPU: MIN_LOG=16 MAX_LOG=23 RAYON_NUM_THREADS=16 cargo test --release test_wide_fib_prove_with_blake_cuda --features parallel -- --nocapture
  • Benchmarks (Throughput)
    • SIMD: STWO_QUIET=1 LOG_N_INSTANCES=23 RUSTFLAGS="-C target-cpu=native -C opt-level=3" RAYON_NUM_THREADS=16 cargo bench --bench wide_fibonacci_blake2s --features parallel -- --nocapture
    • GPU: STWO_QUIET=1 LOG_N_INSTANCES=23 RAYON_NUM_THREADS=16 cargo bench --bench wide_fibonacci_cuda_blake2s --features parallel -- --nocapture

Log(Size)	Prove SIMD ms	Prove GPU ms	Speedup	Thr SIMD (Kelem/s)	Thr GPU (Kelem/s)	Thr Speedup
16	199	222	0.90x	429	2131	4.96x
17	267	34	7.85x	478	3844	8.05x
18	450	47	9.57x	671	6672	9.94x
19	757	57	13.28x	909	10123	11.14x
20	1390	87	15.98x	973	12638	12.99x
21	2670	139	19.21x	1127	15007	13.32x
22	5166	254	20.34x	870	16435	18.89x
23	11014	488	22.57x	783	17898	22.86x

• Poseidon commit channel
  • Tests (Prove)
    • SIMD: RUSTFLAGS="-C target-cpu=native -C opt-level=3" MIN_LOG=16 MAX_LOG=24 RUST_LOG=info RAYON_NUM_THREADS=16 cargo test test_wide_fib_prove_with_poseidon_simd --release --features parallel -- --nocapture
    • GPU: MIN_LOG=16 MAX_LOG=23 RAYON_NUM_THREADS=16 cargo test --release test_wide_fib_prove_with_poseidon_cuda --features parallel -- --nocapture
  • Benchmarks (Throughput)
    • SIMD: STWO_QUIET=1 LOG_N_INSTANCES=23 RUSTFLAGS="-C target-cpu=native -C opt-level=3" RAYON_NUM_THREADS=16 cargo bench --bench wide_fibonacci_poseidon --features parallel -- --nocapture
    • GPU: STWO_QUIET=1 LOG_N_INSTANCES=23 RAYON_NUM_THREADS=16 cargo bench --bench wide_fibonacci_cuda_poseidon --features parallel -- --nocapture

Log(Size)	Prove SIMD ms	Prove GPU ms	Speedup	Thr SIMD (Kelem/s)	Thr GPU (Kelem/s)	Thr Speedup
16	2566	278	9.23x	34	798	23.73x
17	4919	105	46.85x	34	1301	37.75x
18	9745	133	73.27x	35	1939	54.65x
19	19146	196	97.68x	36	2642	74.27x
20	38325	321	119.39x	36	3282	91.78x
21	76322	558	136.78x	35	3695	104.18x
22	152660	1037	147.21x	36	3975	111.35x
23	305667	2010	152.07x	36	4124	116.10x

Poseidon Test

• Blake2s commit channel
  • Tests (Prove)
    • SIMD: RUSTFLAGS="-C target-cpu=native -C opt-level=3" MIN_LOG=16 MAX_LOG=23 RUST_LOG=info RAYON_NUM_THREADS=16 cargo test test_simd_poseidon_prove --release --features parallel -- --nocapture
    • GPU: MIN_LOG=16 MAX_LOG=22 RAYON_NUM_THREADS=16 cargo test --release test_poseidon_prove_with_blake_cuda --features parallel -- --nocapture
  • Benchmarks (Throughput)
    • SIMD: STWO_QUIET=1 LOG_N_INSTANCES=22 RUSTFLAGS="-C target-cpu=native -C opt-level=3" RAYON_NUM_THREADS=16 cargo bench --bench poseidon_blake2s --features parallel -- --nocapture
    • GPU: STWO_QUIET=1 LOG_N_INSTANCES=22 RAYON_NUM_THREADS=16 cargo bench --bench poseidon_cuda_blake2s --features parallel -- --nocapture

Log(Size)	Prove SIMD ms	Prove GPU ms	Speedup	Thr SIMD (Kelem/s)	Thr GPU (Kelem/s)	Thr Speedup
16	3545	198	17.90x	207	1280	6.19x
17	6653	221	30.10x	250	2179	8.73x
18	13355	231	57.81x	288	3497	12.12x
19	26097	272	95.94x	363	4962	13.66x
20	51518	312	165.12x	457	6312	13.82x
21	103045	403	255.69x	331	7186	21.71x
22	206096	579	355.95x	309	7763	25.12x
23	422364	N/A	N/A	N/A	N/A	N/A

• Poseidon commit channel
  • Tests (Prove)
    • SIMD: RUSTFLAGS="-C target-cpu=native -C opt-level=3" MIN_LOG=16 MAX_LOG=23 RUST_LOG=info RAYON_NUM_THREADS=16 cargo test test_simd_poseidon_prove_with_poseidon --release --features parallel -- --nocapture
    • GPU: MIN_LOG=16 MAX_LOG=22 RAYON_NUM_THREADS=16 cargo test --release test_poseidon_prove_with_poseidon_cuda --features parallel -- --nocapture
  • Benchmarks (Throughput)
    • SIMD: STWO_QUIET=1 LOG_N_INSTANCES=22 RUSTFLAGS="-C target-cpu=native -C opt-level=3" RAYON_NUM_THREADS=16 cargo bench --bench poseidon_poseidon --features parallel -- --nocapture
    • GPU: STWO_QUIET=1 LOG_N_INSTANCES=22 RAYON_NUM_THREADS=16 cargo bench --bench poseidon_cuda_poseidon --features parallel -- --nocapture

Log(Size)	Prove SIMD ms	Prove GPU ms	Speedup	Thr SIMD (Kelem/s)	Thr GPU (Kelem/s)	Thr Speedup
16	3341	245	13.64x	17	542	31.27x
17	6613	290	22.80x	18	845	47.19x
18	13384	348	38.46x	18	1208	66.46x
19	26085	436	59.83x	18	1543	84.68x
20	51397	598	85.95x	18	1803	98.11x
21	103973	964	107.86x	18	1957	107.54x
22	207684	1623	127.96x	18	2060	112.94x
23	411598	N/A	N/A	N/A	N/A	N/A

Notes On Recent Changes

• Added Poseidon commit channel variants for both Wide-Fibonacci and Poseidon examples (SIMD and CUDA), and corresponding benches/tests.
• Renamed previous Blake2s-channel benches to *_blake2s to distinguish from *_poseidon benches.
• Tests now reuse shared example prove functions across SIMD/CUDA and Blake2s/Poseidon channels, reducing duplication.

🥳 Acknowledgements

• stwo-gpu : The m31 field arithmetic and extended field operations, FRI operations and quotient accumulator are inspired by stwo-gpu.
• era-bellman-cuda : Low-level field arithmetic code was referenced when implementing Poseidon252 hash computations.

📜 License

This project is licensed under the Apache License 2.0.

See the LICENSE file for more information.