Performance Benchmarks

Pregenerated benchmarks

We have a performance datasheet with runtime and memory performance information. This information is organized by both program size and hardware target, and includes both CPU and GPU targets. The datasheet also includes the RISC Zero zkVM version used and the size of several example programs to give context for the cycle count size data.

Generating your own benchmarks

You can generate benchmarks yourself to get performance data for your exact system, or if you want benchmarks for a different RISC Zero zkVM version than used in the datasheet.

To generate benchmarks, first install Rust and the RISC Zero toolchain and clone the risc0 repository if you haven't already done so. Navigate to the risc0 source code directory and, if necessary, check out the branch or tag you wish to generate benchmarks for. Then run

cargo run --release --example loop

This will produce the benchmark data shown in the datasheet for your system (using the CPU) on the checked out version of the RISC Zero zkVM.

If you want to benchmark a GPU, you will need to build with the cuda or metal feature enabled (whichever is appropriate for your hardware). To do this, use the following commands:

• Metal

cargo run --release -F metal --example loop

• CUDA

cargo run --release -F cuda --example loop

We also have a Fibonacci computation benchmark, which you can run with

cargo bench --bench fib

This will compute the 100th, 1000th, and 10000th Fibonacci numbers modulo 2^64 (ten times for each). It will report both time and throughput (how many numbers were added per second) with separate statisics for execution and proving. As with the loop benchmark, the Fibonacci benchmark will use the CPU by default, and you can benchmark a CUDA or Metal GPU by setting the appropriate feature flag.