Hi everyone,

I'm currently applying a new version of the loneseed application (1.07) with the following changes/improvements:

1. Smaller batch sizes when we detect a GPU on the Maxwell architecture
(Resolves the 500 or so instances where older 900 and 700 series GPUs cannot complete the work since the kernel runs for too long and gets killed due to timeout)
2. More robust error handling for cuda functions prior to work beginning
(Should help us diagnose the remaining segmentation faults we're getting from some users)

Please let us know if you encounter any issues using this new version.


Thank you!

Hi everyone,

I'm going to push a 1.06 version which reverts a change we made to the maxrregcount argument Nvidia's compiler uses.
It sped up certain cards (Titan X Maxwell, RTX 3090, V100, GTX 1060 tested) but RTX 4090, RTX 2080Ti, and seemingly GTX 750Ti have been negatively impacted.

For the error on GTX 750Ti, it's just a timeout error as there's a limit to how long you can run a single kernel wave. Since we changed the max register count, it's possible this impacted the weaker GTX 750Ti enough to fail this timeout.

Hopefully this addresses both issues.

Thanks!

Hello,

We've pushed another update (1.05) that may address the speed change.

Please let me know if this helps.

Thanks!

It is not about checkpointing. When a task is paused, I so far assumed it is just halted and kept in memory. The experience I described above, however, suggests that Minecraft discards the entire data after a certain period of time? Which I found baffling. ;-)

Michael.

Please remove the useless once a minute scheduler call from your settings.

If there is a real need for the client to call home make the interval at most once per hour.

Removed, this was dangling from testing, apologies