So, I started with a cheap SFF Lenovo M725s. It takes 90 minutes per plot. My desktop is a XPS8950, but I really don’t want to use it for plotting or farming. I was thinking of using the 725 to farm and buy something else used to GPU plot. I’m looking at a Xeon workstation like the T5810, with 1 Xeon and 256gb of RAM. But I also have seen 1U servers like the Hyve Zeus with 2 Xeon cheaper. I don’t think 2 Xeon help, but they are CHEAP.
My question is, obviously a good GPU to plot with wont’ fit in a 1U server case. BUT, could I use GPU risers (I already have) and hook up a GPU that way. Would it be a bottleneck or would that work?
I don’t need sub-10 minute plots. I just want C8 plots to get extra plots per HD. If the plot times are sub 30 minutes that’s cool. I’ve played and played with the 725s to get the plots from 160 minutes to 90 minutes, so 30 would be awesome.
Thanks!
1 Like
A GPU riser will work for this. This should not cause a bottleneck in anyway - as the riser is plugged directly into a USB port.
I actually did a comparison on Evergreen miners VS Hard Drive mining a few days ago.
1 Like
Post this question here as well,
Good chia related forum.
I have a Dell T5810 with 256 GB Ram and a Tesla P4. It plots k32 plots in 8 mins all day long.
Tesla P4’s go for cheap ($100). So whatever work station you buy, a Tesla P4 will give you sub 10 min k32 plots when using 256 GB RAM.
Do note that the Tesla P4 does not have any video out plug ins. So if you are not using a headless OS, you’ll have to find another way to get your monitor hooked up. And older GPU’s that use Nvidea’s legacy drivers won’t work alongside the Tesla P4 as it works with Nvidea’s newer drivers… stuff you don’t think about until you find out it won’t work…
Again, do a bit more research and head over to https://chiaforum.com/ to get a lot of good advice!
Definitely much cheaper and also gives you the ability to take full advantage of compressed plots when going the non-evergreen route.
2 Likes
I got the T5810 and it came and a Tesla K40. I need some kind of fan on it and waiting for a 825W PSU to come. I got the T5810 for $75 so I can’t complain it only came with a 425W PSU. I had an old GTX970 laying around in the meantime, and damn, it does 10-11 minute plots. It will randomly crap out after 30-50 C8 plots but I am amazed!
2 Likes
It’s pretty sweet isn’t it? You won’t look back!
Here are some tips for the crashing, MadMax’s discord also has a lot of tips for preventing the crash,
Gigahorse Cuda Plotter works best on kernel 5.15 currently… 5.19 manages the memory differently and for the worse.
If you do have 5.19, you’ll need to play with -M setting. Start with M setting half of your total system RAM. If you still have issues, lower it to a quarter of your System RAM.
Also, another option that you may also have to do is remove/disable OOMD as well to prevent it from crashing.
If using 5.15, it may just work without doing anything with M settings. Although M setting still may be needed with 5.15, the setting can be set higher with 5.15 vs 5.19 and hence an improved plotting speed.
I’m getting roughly 1.5 minutes faster with 5.15 vs 5.19 plotting K33 plots using half RAM mode (256GB). As well as more stable plotting.
5.19 does work, it’s just slightly slower and takes a bit more tweaking currently…. Either lower M setting which slows plotting, or disabling/removing OOMD.
1 Like
I tried Ubuntu, but seeing as I had half C8 plotted drives formatted NTFS, I was getting HORRIBLE transfers to the HD in Ubuntu…I was told ext4 would be much faster, but I’d have to start over with 13 8-12tb harddrives. I’ve got the plotting sorted in Windows on the T5810. I did spring for a 3060, but now plot sink feeds plots to the HD about as fast as the 3060 plots them. About 6-7 each. I could plot faster, but it outruns plotsink and bogs down.
I have the farmer running on the 725s in WSL. It’s cranky starting up reading a bunch of HD at once, but I eventually will start farming on the 5810 as well or maybe solely.
1 Like