Bladebit Cuda - Low Proofs Ratio [Bug]

[FFurno]

What happened?

I have created a few uncompressed plots (-z 0) using bladebit cuda (various alpha versions, latest 4.4). The plots look okay and are recognized when farming, but if I run chia plots check I get very low ratios despite high -n (<0.4). I have win 11, rtx 3080 10gb, a threadripper 1900x, x399 mobo with 256gb ram ddr4. I used chia blockchain 1.8 to check the plots.

Version

1.8

What platform are you using?

Windows

What ui mode are you using?

GUI

Relevant log output

No response

[wjblanke]

you could try 2.0.0-b5 and see if uncompressed cuda plots result in better proof ratios. for further assistance reach out to discord

[Ron-ski]

I've also noticed this, I sometimes get invalid partials, and when checking the logs I find the following, which includes a plot ID.

ERROR Invalid proof of space

When I checked the plots, I get a lot of the following warnings.

WARNING Quality doesn't match with proof. This can occasionally happen with a compressed plot.

And the ratios are very low.

073a9fba4d7d71924ee031dcf46c0af6e8253e1027bc0fd550bff56a1a5bb14a = Proofs 9 / 60, 0.15	
7679b2f0838ecdd830e4e7f072dcb05359acd419bf643530153d93ef29c47f9b = Proofs 5 / 60, 0.0833
a0e27263db5c6c0f45c4b69bffe2249df3d34c8379291c69b2cf742d2b913cb8 = Proofs 14 / 60, 0.2333
b7abc6feae8d16022f5ab5c0acc0ab7cc20e7c4229a93ade919dee768f5a658b = Proofs 16 / 60, 0.2667 
2095ab1561a17d60d4471652ed17703dee05a8897a6715ef7463305a65e4434a = Proofs 22 / 60, 0.3667
25a1fafd5f44f8d27473e8a1b33adfafc7b58868ab7acf5a30d840b47ffd0746 = Proofs 14 / 60, 0.2333	
40d8b4630405f737ce0374a37b17f1e8bb81b32b0e7cb2774c437825446cba9d = Proofs 13 / 60, 0.2167		
09484bfe789caea3ad37a1d826793b98fd0bff675e4b4869fbd48bea90a831fe = Proofs 17 / 60, 0.2833
6baf8079269b71d38be544181f25988f5f9c39ca4af00537a182104a8602cec7 = Proofs 24 / 60, 0.4	

None of the above plots were deemed invalid, but I've now replaced them, and will see if I get any more errors/invalid partials.

These were plotted on Windows 11, using a 3080 and Bladebit Alpha 4.3
Farmed on Windows 10 on the GUI

[mjoach]

I've been trying to find out why the pool underestimates the size of my C7 farm by about 10-15%. I identified hundreds of low proof-ratio plots and replotted them, it helped, but the size is still underestimated. So I did a histogram from 30 challenges test on all plots (slow!). It is not centered on 30, so I think it suggests that it is not just a few bad plots, but rather, I was generating plots with various percentages of invalid proofs and right now I don't see an easy way to identify them. Plotted with Alpha 4.3/4.4/4.5 on Ubuntu 22.04.

[github-actions]

This issue has not been updated in 14 days and is now flagged as stale. If this issue is still affecting you and in need of further review, please comment on it with an update to keep it from auto closing in 7 days.

[jmhands]

we do we want to follow up on this, thanks again for getting the data. From what I can gather with a large sample size (thousands) you really shouldn't be getting under say, 97-98%. But we have no explanation of how this could happen and plot still be not corrupt. We will take a look

[mjoach]

Thanks, good to hear. I have seen at least 2 other users report an average of 28/30 for their farms on the Discord channel. So either a plots check specific bug or something potentially more serious, plotting related. Myself, I have never found out why my C7 farm's size is underestimated by the pool, though this could still be related to using GTX 1070 for farming and twice-daily harvester crashes/GRResult errors. Plotting and the above plots check were however all done on 3090.

[FFurno]

Tried to plot uncompressed with newly released bladebit 3.0. Same problem as before. Plots are recognized as valid but only have 30-40% valid proofs in Win 11.

[jmhands]

I reproduced this internally, but it has a very low occurrence rate. Will be following up on this soon
Chia-Network/bladebit#393

[llowrey]

I opened a ticked the other day about non-reproducible plots: Chia-Network/bladebit#390

It seems that cudaplot does not produce the same plot data every time when the plot id (seed) is the same (by specifying the plot id using the -i parameter when plotting). For diskplot the plots are always the same (except for a small bit in the header). Since the point of the -i parameter is to reproduce plots, they must be fully deterministic given the plot id. The fact that cudaplot does not produce the same contents for a given plot id is concerning.

[FFurno]

I have also tried to produce compressed -z 5 plots using bladebit cuda 3.1 beta on the same Threadripper platform described above. Proofs are close to zero, plots are detected as invalid.

[emlowe]

I believe the original report was fixed by Chia-Network/bladebit#410

closing