Experiment Description
These benchmarks compare two different mechanisms for verifying that a set of random scalar values lies within a certain range:
- SignExpr through inequality (treating scalars as full width, 2^248)
- Range check construction (on both u16 and u8)
Both experiments were run using the dynamic Dory proof system with a nu value of 16 on an NVIDIA A100 multi-GPU setup. The table below summarizes the timings for sequential and fully-random inputs, as well as varying input sizes.
Single-Value Results (size = 1_000_000)
| Experiment |
Time |
| trace fully random SignExpr through inequality |
52.3s |
| trace fully random rangecheck on u16s |
8.97s |
| trace sequential 0 to 1 million SignExpr through inequality |
7.35s |
| trace sequential 0 to 1 million rangecheck on u16s |
8.7s |
Fully Random Inputs
SignExpr (2^248) Through Inequality
| Input Size |
Time |
| 1000 |
532.22ms |
| 1,000,000 |
52.3s |
| 10,000,000 |
3m 40s |
Rangecheck on u16s
| Input Size |
Time |
| 1000 |
2.78s |
| 1,000,000 |
8.97s |
| 10,000,000 |
36.4s |
Rangecheck on u8s
| Input Size |
Time |
| 1000 |
352.8ms |
| 1,000,000 |
14.3s |
| 10,000,000 |
1m |
