This page is collecting number of running SupraSeal C2 on a 32GiB sector, the runtime is in seconds. It depends on the hardware, hence there’s always a comparison between using CUDA and using SupraSeal. Most important part is the GPU that is used, hence it’s mentioned which one it is.
| User |
GPU |
Speedup |
CUDA (sec) |
SupraSeal (sec) |
Source |
| vmx |
Quadro RTX 6000 |
4.86 |
1104 |
227 |
‣ |
| vmx |
RTX 4090 |
3.05 |
397 |
130 |
|
| Reiers |
RTX 4090 |
3.67 |
525 |
143 |
‣ |
| Reiers |
TITAN RTX |
|
|
183 |
‣ |
| TippyFlits |
A4000 |
2.03 |
1091 |
538 |
‣ |
| Stuberman |
RTX 3090 |
|
|
127 |
‣ |
| zcfil |
RTX 3090 |
3.88 |
535 |
138 |
‣ |
Benchmarks for Non-interactive PoRep
10 PoReps (this is what interactive PoRep does) with CUDA:
- synthesis time: 82.938055025s
- prover time: 1020.361815339s
- total time including other overhead (from first synthesis to before next proof): 1105s = 18:25min
10 PoReps (this is what interactive PoRep does) with SupraSeal:
- synthesis time: 87.836821965s
- prover time: 149.925715356s
- total time including other overhead (from first synthesis to last proof finished): 245s = 4:05min
126 PoReps (this is what non-interactive PoRep does) with Supraseal:
- synthesis time:
87.836821965 + 84.757182069 + 84.737105604 + 84.750190676 + 84.705759966 + 84.63558107 + 84.759582445 + 84.838175901 + 84.571978607 +
84.803744137 + 84.589651577 + 84.500893984 + 81.972020394 = 1101.4586884 = 18:21min
- prover time: 149.925715356 + 94.218907837 + 121.449434091 + 118.061023323 + 120.269586734 +
121.15618083 + 124.251977965 + 94.169493368 + 92.452175002 +
93.070175029 + 84.141457732 + 84.439310483 + 55.074800451 = 1352.6802382 = 22:32min
- total time including other overhead (from first synthesis to last proof finished): 2546s = 42:26min
Conclusion:
- SupraSeal for interactive PoRep is 1105s/245s = 4.51x faster than with CUDA
- Non-interactive PoRep with
SupraSeal takes 2546s/245s = 10.39x more time than interactive PoRep
with SupraSeal (please note that the first prover time was expectionally slow, which makes the slowdown lower. When you take the average of all
proving times that proof 10 PoReps each (108.133786479s), then the
slowdown is 2546s/204s = 12.48x, which is similar to the one observed
with the RTX 4090 and also matches the theory of having about 12.6x the
proofs (hence work) to be done.