Well 16GB mode on linux might be enough for me to finally pull a refurb at the shop and put linux on it to plot. I know plotting costs money as well but I really don’t care about how long plotting takes as long as it is under 1 hour per plot. I think that is fair.
So 16GB and a Nvidia card + Linux is what I really need then from the sounds of it. Or just be lazy and buy the cheapest kit of 128GB DDR4 ram I can get just for plotting. Ram is like the only thing in that system that is not a total nightmare to swap out.
yeah I probably wouldn’t be mining anything at 0.30 per kWh tbh. That sounds awful. Luckily for me I am in america where we just light the world on fire for cheap power, which is also unlucky for the rest of the world lol. I made a solar power plant for my boat and it works awesome, I am also going to get some solar on my roof as well, the only thing really stopping me is the city planners not understanding anything about solar but I would hope that is rapidly changing as I am finally starting to see people pop solar on their roofs in my area (took long enough).
With you’re small farm size I don’t think its worth spending money buying a 128GB kit, unless you need it for something else.
Just install Linux Mint on a spare SSD and boot off that when you need to plot. Linux is way faster at plotting than Windows, so probably not much difference between 16GB Linux and 128GB Windows. Get a cheap enterprise SSD for the temp drive, it won’t be as fast as an NVME, but again it’s not worth spending a lot on.
My gas and electric bill for 2023 is negative due to the solar, and various other things I take advantage of, but the solar and batteries was a huge investment.
I just fired up a ‘version 1’ NOSSD instance the other day, C4 with the same thought in mind (not interested in running a GPU to farm 25 TB). It plots about 2TB a day just banging away on hard drives, no GPU or SSD needed .
The command string I use in the .bat file (I limit ram to 9GB so I can use my machine while plotting):
client.exe -a xch1zgdawm8c9c3vulr4mzr6nz6sy9z3gwmygudg7haur8eghjhzpztqqlq3cr -d H:\ -d G:\ -d D:\ -c 4 -m 9GB -w GeneIsAdork -u 0
I would still use m.2 NVME’s to plot. I plotted 4TB since I posted this 21 hours ago using an really old plot manager on win 11.
Using no SSD is only temping because they let you use compression without insane requirements but then I would also need to replot all of my drives because I don’t want to be on 2 different pools and not allowing any of your code to be open source for auditing doesn’t make me want to run that program on my office network.
CUDA
An NVIDIA GPU is required for this mode. This mode is exposed via the cudaplot command in a separate executable "bladebit_cuda". This mode has mainly been tested on consumer cards from the 10xx series and up.
Mode OS DRAM VRAM CUDA capability
In-RAM Linux, Windows 256G 8G 5.2 and up
Disk-hybrid 128G Linux, Windows 128G 8G 5.2 and up
Disk-hybrid 16G (WIP) Linux 16G 8G 5.2 and up
**NOTE: 16G mode currently a work in progress and at this stage it only works in Linux and direct I/O is unavailable in this mode.**
So looking at the github it does say 16GB mode is a work in progress, so sitting around and waiting for development might be another option as well. But also makes me wonder what direct I/O is unavailable in this mode means as well? Any idea?
I think you’ll have a long wait, CNI are still working on stuff, but other things have much higher priorities than plotting. Make sure you run plot checks on Bladebit plots, some can be bad.
I’m sure 16GB mode does work on Linux, but I’ve no idea what direct I/O is. I’m a NOOB when it comes to Linux, only installed and and set it up because plotting is so much faster than Windows, and my workstation has 512GB of ram, so with two GPU’s I can run two plotters at once. My 3080 plotting time halved, to just over two minutes, and my P4 Tesla went from 10 minutes to 6 or 7 minutes. With GH C19 plots I was creating about 900 a day, or in the region of 56TiB.
Jesus lol. Man it would be stupid nice to plot a 8TB drive in under a day. I’ve also debated getting a super cheap server just to plot with since the ram is so much cheaper and I could fire that up and let it plot without affecting anything else I am doing that day on my systems. I also am not growing my farm nearly enough to really justify anything besides a few drives here and there for fun.
Might still try a second refurb box dedicated to it where I just throw it up, connect the drives and forget it exists until it is done on linux with a cheap 8GB nvidia card. Grabbing a tesla P40 sounds like a decent idea as well as those are pretty dam cheap.
When you use a standard file handling, the OS level cache is in place. For most tasks, that is more than good enough. However, in case of a plotter it may be beneficial to do your own RAM caching, as the program better understands what needs to touch the SSD (briefly) and what may be needed for longer or used more often.
Most likely, for 16 GB there is just not enough RAM for BB to do much if any caching (whether by OS or by the plotter).
On the Linux side, OS caching is really well optimized, so usually trying to improve on that is a challenge. Therefore, it would be nice to see a comparison of BB runs with and without direct IO as benefits of using it may be overstated. (If the program knows that something should not touch the SSD, it should keep that in RAM, rather than writing it to a file.)
On Win side, OS caching may not be as good as on Linux, so potentially benefits of direct IO could be greater. Although, PrimoCache has an option to defer writes almost completely. This may give about the same benefits as using a direct IO by a given program. (I tried to use it with MM 1.0, and it worked really great, but ended up buying more RAM.)
My system was relatively slow, fastest plot time I think was 49 seconds.
There are some great cheap systems out there, pretty cheap now, unfortunately when I bought mine prices had gone up, so it all added up rather expensive.
Ideally you want a 20 series GPU or better, Bladebit plotting and farming has issues with 10 series and cards like the P4 Tesla, although I didn’t experience it when I originally plotted BB C7 plots. Search Chia GRR error.
Let me add my 2 mojos on this from a very small ( <Pb ) farmer perspective who has been farming since just before mainnet.
I’ve been on OG then NFT and now almost done converting to compressed plots to take advantage of “increase” in rewards which sounds good but really does not exist and before you take out the pitch fork hold of and reading
When the compressed plots came out and especially once GH went mainnet I did notice that my farmer was starting to make less and less rewards overtime as the ETA to block kept on rising so it was only natural to convert to compressed plots just to keep the rewards close to previous C0 levels. Note that I closely monitor my farm and all harvesters with custom solution I developed to keep track of common issues and to ensure my farm is at peek performance 24/7.
When I decided to switch to compressed plots I was watching what develops for couple of months before jumping in as I did not like the fact that GH uses proprietary farmer and there is only 1 dev supporting it. The reason I do not like that is because I decided to stay with XCH for a long time and this is definitely a big risk, same goes with NoSSD but in their case NoSSD solution look too much like HPool to me which for me is a negative.
The way the fee is applied with GH is also a big negative for a small farmer. GH takes your .25 XCH on every 1 of 4 blocks farmed (when GPU farming) or 1 of 8 (when CPU farming). The randomization in this fee eventually works out for a large farm over time but keep in mind that small farmer is looking at only getting single or low double digits blocks over it’s life time and randomization may not be favorable to you depending on your farm’s luck and whether it is running properly 24/7 which in case of very small farms may be an issue. This adds yet another layer to uncertainty to farming which I did not want to get in to.
This left me with only one choice which is BB and fortunately for me the 128GB RAM edition was finished right around the time I decided to convert to compressed plots. The reason 128GB version is important for me is because that is typically the maximum RAM you can get in a common budget end user PC these days without getting a specialized workstation or server and because I already had this hardware from NFT re-plot.
With BB I’m reploting to a combination of C3 & C4 so that my CPU harvesters hit <3sec max response time with current 256 filter. Once Chia increases that to 512 (June’24) I expect this to double so I hope to stay at 5-6sec max then. Number of GPU farmers will likely be forced to either 2x their farming GPU power or downgrade their compression level in June once 512 filter hits and if the price of XCH does not change it may be hard to justify GPU farming with increased power usage at that time so I will be interested to see what happens with the compressed plots and GPU farming after June. All you GPU farmers start cranking these numbers if you haven’t already
@Toaster_Potato if you decided to go with BB I can help you get that setup and test 16GB RAM mode on my setup which I currently do not use since I have 128GB of RAM. In my case with 5950X CPU, 128GB@3600 RAM, 2TB temp NVME Gen 4, and GTX1080 I’m creating plots in ~5-6min + transfer time to final destination. Note that my final destination is a high speed NAS over 10Gb NIC and I’ve noticed that plot generation in 128GB mode does start to write to the destination before plot is finished so the 5-6min may be different on your setup depending on your destination. This is on Linux desktop (Debian 12) but all done from terminal (aka CMD for Win folk) so if you are willing to learn I can guide you through it. I completely gave up on Windows over a year ago when I was reploting to NFT and noticed 20%-30% better performance on Linux with the same hardware. Since then I used various Linux distros as my main desktop and I was not new to Linux before then as I frequently used it for other server & SBC uses but I never considered using it for my primary desktop as I wanted to game on it as well. With major improvements that happened over the recent years that is no longer a problem as I’m happily gaming on the same setup with my primary AMD GPU even when I’m plotting on secondary GTX (depending on how much RAM the game needs).
If you got to the end of this book I hope this info helps you
Yeah I would definitely be interested in learning. The biggest issue is using linux on the rig with the 3080, it runs my business 24-7 so even dual booting it to linux isn’t that appealing. I do have a ton of other desktops around me though, just not qualifying video cards.
The home system with the 7900x and 7900xtx I would most like to use the linux to plot but then the 7900xtx can’t be used and I can’t pull the 3080 out either.
Most likely if I really want to keep building a farm I will probably have to upgrade my work rig to win 11 and get 128GB of ram for it from the sounds of the options here because it is also DDR4 and has the rest of the requirements. I also transfer to the drives using SATA3 ports.
But then also I have been reading into the halving coming up and have been thinking about is it really worth putting money into it when soon I will make even less.
I’ve used linux in years way back, I am very familiar with command prompts and scripts. But what almost sounds the most appealing is just figuring out how to get a system running linux around me that meets the requirements that I can pull out to plot when needed and throw back into a closet when done.
I do really appreciate your detailed input and how you wrote the fees about GH is exactly how I was trying to say it. It’s just not appealing besides the ease of use and it’s RNG which could be really fkkin bad feeling if it triggers twice in a row on me.
I need to sleep on the options and think about if I really even care enough to waste your time helping me get setup with a linux box.
I have tons of PCI4 NVME drives and get like 30 min plots currently (was 20 min last time I used this, idk what changed) but really just using c3-4 just sounds like the much better way to be plotting new drives now especially with everyone doing it.
Maybe getting a 1080 or something really cheap to plot with and pulling the XTX to plot might be the best way? Idk I have alot of parts but not a single ready to go system meeting the requirements currently is the biggest hurdle.
With GPU plotting how important is the CPU now? Like can a potato CPU plot great with a 1070?
Direct io is “starting copying the plot file to the disk (if the disk has the ability) while plotting. When the plot is completed, the copy process goes on and it takes about 3-4 minutes more. If you copy the file after completion, it takes 20-30 minutes for one file copy! Most farmers copy the new plot to an SSD then move the file to the disk. I have tested this on exos 16 TB disk. By direct io, total copy time is nearly half of copying after completion.
It is not as important now as GPU does most work. On my setup I see it using 2-3 cores. Note that Nvidia 10-series or newer with at least 8GB of GDDR is a must for BB. I’ll do some tests to compare 16GB vs 128GB RAM plotting on my setup to give you an idea on what difference to expect.
Bad news on 16GB mode so far as it fails for me same way as BB issue #445. This is with the current BB v3.1.0-dev compiled on the same machine using Nvidia driver 545.23.08 & CUDA 12.3
I’ve tried C4, C3, and C1 with 16GB mode and all failed the same way. 128GB mode works fine though.
Any idea what the cheapest video card that will work for it is?
That depressing about the 16GB mode. I also just got another bench system at the shop but it is again DDR5 R7 7000 series. That would be the best system to use as a dual boot to linux when not in service.
128GB of DDR5 is pretty pricey just for a chia plotter.
Honestly the best move might just be to wait it out. When the halfling hits, DDR5 could fall or grow in density to the point grabbing 128GB isn’t as bad. A big point of DDR5 is higher density afterall.
.
