Benchmark: Proving with Risc Zero

Determination of the expected proving speed (in TPS) for a given machine and configuration using the Risc Zero zkVM. This benchmark produces results with a tradeoff between proof latency and max throughput.

Setup

Before running the benchmark, make sure your machine is set up and ready to go. Refer to the Setup Guide for how to set everything up with dependencies.

Running

The benchmark currently supports both CPU and GPU proving with Risc Zero.

For CPU proving, run the benchmark with the following command:

cd benchmarks/zk-proving/risczero
RUSTFLAGS="-Ctarget-cpu=native" cargo run --release

For GPU proving, run the benchmark with the following command:

cd benchmarks/zk-proving/risczero
RUSTFLAGS="-Ctarget-cpu=native" cargo run --release -F cuda

Note: GPU proving takes a VERY long time to compile due to having to build the GPU kernels

Options

The following options are available:

Options:
      --num-accounts <NUM_ACCOUNTS>
          The number of accounts to fill the state trie (default: 1_000_000) [default: 1000000]
      --max-proof-duration <MAX_PROOF_DURATION>
          The maximum amount of proof generation time to target (default: 600 seconds) [default: 600]
  -h, --help
          Print help
  -V, --version
          Print version

Output

The benchmark outputs two tables. One lists the raw proving time and the number of transactions/events that were proven. The other lists estimated TPS values for a range of proof latencies.

Example:

Proof generation benchmark results:
Events  Time (s)        TPS
1       3.84    0.26
2       3.92    0.51
...
10000   566.03  17.67
20000   1021.02 19.59

Estimated TPS at various latencies:
Latency (s)     TPS
5       3.07
10      7.94
...
120     14.50
300     16.10

Benchmark Results

The following results were obtained on the various reference machines.

AMD Ryzen 9 9950X + RTX 5090

GPU proving

Proof generation benchmark results:
Events  Time (s)        TPS
1       3.84    0.26
2       3.92    0.51
5       4.47    1.12
10      4.23    2.36
20      5.66    3.53
50      7.96    6.28
100     11.43   8.75
200     19.53   10.24
500     39.54   12.65
1000    74.62   13.40
2000    135.90  14.72
5000    309.93  16.13
10000   566.03  17.67
20000   1021.02 19.59

Estimated TPS at various latencies:
Latency (s)     TPS
5       3.07
10      7.94
30      11.90
60      13.19
120     14.50
300     16.10

AWS g6.xlarge

GPU proving

Proof generation benchmark results:
Events  Time (s)        TPS
1       6.23    0.16
2       6.26    0.32
5       6.79    0.74
10      7.78    1.29
20      9.23    2.17
50      13.66   3.66
100     19.64   5.09
200     32.81   6.10
500     69.10   7.24
1000    126.09  7.93
2000    234.46  8.53
5000    520.38  9.61
10000   942.56  10.61

Estimated TPS at various latencies:
Latency (s)     TPS
5       --
10      2.52
30      5.96
60      7.08
120     7.89
300     8.96

AWS c6i.8xlarge

CPU proving

Proof generation benchmark results:
Events  Time (s)        TPS
1       66.61   0.02
2       66.52   0.03
5       84.84   0.06
10      125.10  0.08
20      152.68  0.13
50      297.24  0.17
100     472.23  0.21
200     848.63  0.24

Estimated TPS at various latencies:
Latency (s)     TPS
5       --
10      --
30      --
60      --
120     0.08
300     0.17