Benchmarks of VM proving times made by Aligned¶

Table of Contents¶

  • About
  • GPU Benchmarks on RTX A6000
    • Fibonacci Benchmark
    • Keccak Benchmark
    • RSP Benchmark
  • CPU Benchmarks on AMD EPYC 8534P 64-Core Processor
    • Fibonacci Benchmark
    • Keccak Benchmark

About¶

These benchmarks aim to provide a neutral suite for measuring the efficiency of zkVMs.

The benchmarks test the following programs:

  • ✅ Fibonacci
  • ✅ Keccak
  • ️✅ RSP (Reth Succinct Processor)

The benchmarks are run on the following zkVMs:

  • SP1
  • Risc0
  • Pico

You can run this Python notebook yourself to create new plots or analyze the data.

The Githbut repository is available here.

Raw results can be found here.

The full Jupyter notebook can be found here.

GPU Benchmarks on RTX A6000¶

The data was collected on a server with the following specs:

  • GPU: NVIDIA RTX A6000 (48GB)
  • RAM: 44GB
  • OS: Ubuntu 22 LTS

Fibonacci Benchmark¶

In this section, we evaluate the performance of Fibonacci across various provers.

The data is sourced from bench_results/fibo_6_mar_25_A6000.csv.

For validation, we will display the data sorted by Prover and N.

Data sorted by Prover and N for validation:
              Prover         N     Time  Minutes
2              Risc0     10000     1.7s     0.03
5              Risc0    100000     5.6s     0.09
8              Risc0   1000000    48.9s     0.81
11             Risc0  10000000  7m45.5s     7.76
0           SP1-AVX2     10000    13.9s     0.23
3           SP1-AVX2    100000    13.3s     0.22
6           SP1-AVX2   1000000    22.8s     0.38
9           SP1-AVX2  10000000  1m46.8s     1.78
1   SP1-AVX2-Groth16     10000    53.1s     0.89
4   SP1-AVX2-Groth16    100000    53.7s     0.90
7   SP1-AVX2-Groth16   1000000   1m9.2s     1.15
10  SP1-AVX2-Groth16  10000000  2m27.1s     2.45

Plot: Log N vs Log Time¶

No description has been provided for this image

Time used at biggest sample taken¶

Prover N (Millions) Time
9 SP1-AVX2 10.00 1m46.8s
10 SP1-AVX2-Groth16 10.00 2m27.1s
11 Risc0 10.00 7m45.5s

Keccak Benchmark¶

In this section we evaluate the performance of Keccak across various provers.

The data is sourced from bench_results/keccak_6_mar_25_A6000.csv.

For validation, we will display the data sorted by Prover and N.

Data sorted by Prover and N for validation:
              Prover          N      Time  Minutes
2              Risc0        100      7.3s     0.12
5              Risc0       1000      7.5s     0.12
8              Risc0      10000     11.0s     0.18
11             Risc0     100000     43.5s     0.72
14             Risc0    1000000    6m8.0s     6.13
17             Risc0   10000000  64m31.4s    64.52
0           SP1-AVX2        100     21.5s     0.36
3           SP1-AVX2       1000     21.8s     0.36
6           SP1-AVX2      10000     22.0s     0.37
9           SP1-AVX2     100000     22.6s     0.38
12          SP1-AVX2    1000000     33.1s     0.55
15          SP1-AVX2   10000000   2m22.0s     2.37
18          SP1-AVX2  100000000  20m24.4s    20.41
1   SP1-AVX2-Groth16        100     57.2s     0.95
4   SP1-AVX2-Groth16       1000    1m2.6s     1.04
7   SP1-AVX2-Groth16      10000    1m1.6s     1.03
10  SP1-AVX2-Groth16     100000    1m2.6s     1.04
13  SP1-AVX2-Groth16    1000000   1m12.7s     1.21
16  SP1-AVX2-Groth16   10000000    3m2.9s     3.05
19  SP1-AVX2-Groth16  100000000  24m23.7s    24.39

Plot: Log N vs Log Time¶

No description has been provided for this image

KiB hased per second at max sample¶

Prover KiB/sec
0 SP1-AVX2 79.76
1 SP1-AVX2-Groth16 66.72
2 Risc0 2.52

Time used at 10M and 100M bytes¶

Prover Bytes (Millions) Time
15 SP1-AVX2 10.00 2m22.0s
16 SP1-AVX2-Groth16 10.00 3m2.9s
17 Risc0 10.00 64m31.4s
Prover Bytes (Millions) Time
18 SP1-AVX2 100.00 20m24.4s
19 SP1-AVX2-Groth16 100.00 24m23.7s

RSP Benchmark¶

In this section we evaluate the performance of RSP across various provers over blocks with different amounts of Gas.

Notice the cost is not purely lineal, since there can be blocks that require more keccaks and are more expensive to prove than others. For example, the block of 27M gas is cheaper to prove than the one of 18M Gas

The data is sourced from bench_results/rsp_12_mar_25_A6000.csv.

For validation, we will display the data sorted by Prover and N.

Data sorted by Prover and N for validation:
       Prover   N       Time  Minutes
1   RSP RISC0   1    8m31.8s     8.53
3   RSP RISC0   3    23m3.6s    23.06
5   RSP RISC0   9   57m39.1s    57.65
7   RSP RISC0  18  101m26.8s   101.45
9   RSP RISC0  27   48m37.8s    48.63
11  RSP RISC0  36  217m43.8s   217.73
0     RSP SP1   1      44.3s     0.74
2     RSP SP1   3    1m29.6s     1.49
4     RSP SP1   9    3m27.8s     3.46
6     RSP SP1  18    5m53.5s     5.89
8     RSP SP1  27    2m49.8s     2.83
10    RSP SP1  36   11m55.6s    11.93

Plot: Log N vs Log Time¶

No description has been provided for this image

Note: The 27Mgas block has fewer transactions than the other blocks, but one of them is a high cpu usage transaction.

Time used for a 36M gas block¶

Prover Megagas Time
10 RSP SP1 36.00 11m55.6s
11 RSP RISC0 36.00 217m43.8s

CPU Benchmarks on AMD EPYC 8534P 64-Core Processor¶

The data was collected on a server with the following specs:

  • CPU: AMD EPYC 8534P 64-Core Processor
  • RAM: 576GB
  • OS: Ubuntu 24 LTS

Fibonacci Benchmark¶

In this section we evaluate the performance of Fibonacci across various provers.

The data is sourced from bench_results/fibo_5_mar_25_epyc8534p_64c_576gb.csv.

For validation, we will display the data sorted by Prover and N.

Data sorted by Prover and N for validation:
                Prover        N     Time  Minutes
0          Pico-AVX512    10000    22.9s     0.38
4          Pico-AVX512   100000    35.5s     0.59
8          Pico-AVX512  1000000  1m46.0s     1.77
12         Pico-AVX512  4000000  5m18.6s     5.31
3                Risc0    10000    10.8s     0.18
7                Risc0   100000    50.9s     0.85
11               Risc0  1000000  7m53.1s     7.89
15               Risc0  4000000  31m2.0s    31.03
1           SP1-AVX512    10000    16.0s     0.27
5           SP1-AVX512   100000    34.0s     0.57
9           SP1-AVX512  1000000  1m42.8s     1.71
13          SP1-AVX512  4000000  5m21.2s     5.35
2   SP1-AVX512-Groth16    10000  1m18.5s     1.31
6   SP1-AVX512-Groth16   100000  1m36.4s     1.61
10  SP1-AVX512-Groth16  1000000  2m45.8s     2.76
14  SP1-AVX512-Groth16  4000000  6m29.4s     6.49

Plot: Log N vs Log Time¶

No description has been provided for this image

Time used at biggest sample taken¶

Prover N Time
12 Pico-AVX512 4.00 5m18.6s
13 SP1-AVX512 4.00 5m21.2s
14 SP1-AVX512-Groth16 4.00 6m29.4s
15 Risc0 4.00 31m2.0s

Keccak Benchmark¶

In this section we evaluate the performance of Keccak across various provers.

The data is sourced from bench_results/keccak_5_mar_25_epyc8534p_64c_576gb.csv.

For validation, we will display the data sorted by Prover and N.

Data sorted by Prover and N for validation:
                Prover         N      Time  Minutes
0          Pico-AVX512       100     28.8s     0.48
3          Pico-AVX512      1000     28.8s     0.48
6          Pico-AVX512     10000     29.8s     0.50
9          Pico-AVX512    100000     39.4s     0.66
12         Pico-AVX512   1000000   1m46.8s     1.78
15         Pico-AVX512  10000000  12m27.4s    12.46
18               Risc0       100     34.4s     0.57
19               Risc0      1000     37.1s     0.62
20               Risc0     10000    1m8.9s     1.15
21               Risc0    100000   6m19.7s     6.33
1           SP1-AVX512       100     53.5s     0.89
4           SP1-AVX512      1000     53.5s     0.89
7           SP1-AVX512     10000     53.3s     0.89
10          SP1-AVX512    100000    1m2.8s     1.05
13          SP1-AVX512   1000000   2m22.6s     2.38
16          SP1-AVX512  10000000  19m51.6s    19.86
2   SP1-AVX512-Groth16       100   1m56.0s     1.93
5   SP1-AVX512-Groth16      1000   1m56.0s     1.93
8   SP1-AVX512-Groth16     10000   1m56.0s     1.93
11  SP1-AVX512-Groth16    100000    2m5.9s     2.10
14  SP1-AVX512-Groth16   1000000   3m25.3s     3.42
17  SP1-AVX512-Groth16  10000000   21m3.5s    21.06

Plot: Log N vs Log Time¶

No description has been provided for this image

KiB hased per second at max sample¶

Prover KiB/sec
0 Pico-AVX512 13.07
1 SP1-AVX512 8.20
2 SP1-AVX512-Groth16 7.73
3 Risc0 0.26

Time used at 100K bytes and 10M bytes¶

Prover N Time
9 Pico-AVX512 0.10 39.4s
10 SP1-AVX512 0.10 1m2.8s
11 SP1-AVX512-Groth16 0.10 2m5.9s
21 Risc0 0.10 6m19.7s
Prover N Time
15 Pico-AVX512 10.00 12m27.4s
16 SP1-AVX512 10.00 19m51.6s
17 SP1-AVX512-Groth16 10.00 21m3.5s