These experiments demonstrate the overhead incurred while running WASM or EVM interpreters in SP1 over native Rust. We evaluate programs that compute the n-th Fibonacci numbers increasingly for increasingly larger n, along with varying integer arithmetic in u32, u64, and u128, and compile them to .wat files and evm bytecode. As a baseline, we implement these programs in native rust. We run the evm bytecode through the revm interpreter, and we run the .wat through the wasmi interpreter.
We evaluate the number of sp1 cycles it takes to execute the Fibonacci programs for n= (1000, 100000, 1000000).
For the experiment you want to run, go to the [experiment]/program directory and run cargo prove build. Then, go to the [experiment]/script directory and run RUST_LOG=info cargo run --release. For example, if you run revm-32, you might get a result like this.
2025-03-25T17:22:09.018395Z INFO execute: clk = 0 pc = 0x21d39c
2025-03-25T17:22:09.018482Z INFO execute: ┌╴set up input
2025-03-25T17:22:09.019346Z INFO execute: └╴43,474 cycles
2025-03-25T17:22:09.019362Z INFO execute: ┌╴set up runtime
stderr: WARNING: Using insecure random number generator.
2025-03-25T17:22:09.019646Z INFO execute: └╴10,961 cycles
2025-03-25T17:22:09.019662Z INFO execute: ┌╴interpreter
2025-03-25T17:22:09.234599Z INFO execute: └╴9,018,117 cycles
2025-03-25T17:22:09.236082Z INFO execute: gas: 11389884
2025-03-25T17:22:09.236210Z INFO execute: close time.busy=220ms time.idle=2.33µs
Total cycles: 9080285
The setup costs are roughly fixed, and the "run interpreter" part is responsible for executing the actual instructions of the wasm/evm program.
| Experiment: u32 | Set up input | Set up runtime | Interpreter loop | Total cycles |
|---|---|---|---|---|
| Baseline | -- | -- | -- | 16,947 |
| Wasmi | 33,767 | 2,909 | 307,625 | 347,000 |
| Revm | 43,474 | 10,961 | 9,018,117 | 9,080,285 |
| Revm (optimized) | 32,531 | 10,961 | 2,398,175 | 2,449,400 |
| Experiment: u64 | Set up input | Set up runtime | Interpreter loop | Total cycles |
|---|---|---|---|---|
| Baseline | -- | -- | -- | 124,947 |
| Wasmi | 34,050 | 2,910 | 3,873,431 | 3,913,090 |
| Revm | 44,934 | 10,961 | 91,478,766 | 91,537,500 |
| Revm (optimized) | 32,091 | 10,961 | 23,885,232 | 23,940,971 |
| Experiment:u128 | Set up input | Set up runtime | Interpreter loop | Total cycles |
|---|---|---|---|---|
| Baseline | -- | -- | -- | 1,204,947 |
| Wasmi | 34,050 | 2,910 | 38,745,658 | 38,785,317 |
| Revm | 43,473 | 10,961 | 1,002,027,202 | 1,002,088,904 |
| Revm (optimized) | 33,817 | 10,961 | 237,555,380 | 237,603,026 |
fib-32-optimized.sol, fib-64-optimized.sol, and fib-128-optimized.sol contain the smart contracts that are compiled to evm bytecode for the Revm (optimized) benchmarks. Wasmi doesn't support 128-bit integer types, so we substitute with u64. The Solidity optimizations for the revm benchmarks were suggested in this post.
The moduli used are:
| Experiment | Modulus |
|---|---|
| u32 | 7919 |
| u64 | 65776547668456965 |
| u128 | 170141183460469231731687303715884105727 |
The wasmi version used is 0.42.1 and the SP1 version is 4.1.0.