Hi If you were starting to build a plotting machine from scratch, what would be the best cost-benefit ratio? From me lurking around it seems
5900/5950X
32GB - 64GB Ram (not sure about speed and timings)
2 - 4 SSD 2TB with asus raid card
Cheapest GPU possible
Some PC case with decent cooling.
What I am not sure about is a lot of people going for consumer-grade SSD and honestly, they are much easier to find at some places around the world but sustainable write seems to be the main bottleneck here
Yeah obviously ideally best is what you already have
I’ve seen people also going for multiple smaller machines. For example I am skeptical that going above something like 5950x is worth it. Maybe there is some good thread ripper build…
Buying used decommissioned servers is probably your best bet on plot value per dollar. Lots of cores in those things, and the dropoff on Intel perf isn’t too bad as you go back in generations, per my topic.
The build that you can manage and troubleshoot with the most efficiency is the ideal plotting build, in my opinion.
There is at least one whale out there who had zero IT experience so he literally bought tons of simple builds and just scaled up that way rather than managing servers, and there are other farmers with more technical expertise who use more advanced setups.
There are a couple of things to consider. First of all, the Asus cards are sold out, and secondly they are not that great: you need to make sure that your motherboard has 4x4 bifurcation support (and enabling it normally is a pain in the butt). Here’s what you want:
a big chassis that you can stuff with drives and nvme cards or a motherboard with a bunch of nvme slots.
a lot of cheap ram: 64 or 128 gigs if it’s super cheap
as fast a processor as you can afford (or two or four) with a lot of threads – something like 64.
good driver support
Hunt around on Craigslist – I found some really nice deals on older dual processor corporate workstations like DELL Precision T7910 – that one is my favorite so far, but there are plenty of other older high end workstations from HP and DELL that can be had for a song (with the caveat that some people want too much for them). I normally spend about 1200 for the server and 750-1400 for nvmes + cards. These machines are theoretically capable of doing 50-100 plots per day, but of course in reality I’m only getting 20-40 per day as I’m still dialing things in.
I am not sure as fast a process as you can afford is way to go. From my rough estimation it seemed that at some point its actually better to buy a second plotter because you don’t squeeze as much performance.
Unfortunately craiglist doesn’t help me as its not used in my area. Thats why I am curious about build from scratch all else being equal
I reckon a Mac Mini M1 with an external NVME SSD is pretty good value for this. Fast, small, simple, pretty good IO options given its size, and they have good resale value if you ever want to sell it.
Also easy to scale if you want to increase throughput.
I only just got the Minis set up and still figuring out optimal config. But I think 1.2 - 1.5 TB 2 day. Will report back when I see the run rate in the morning.
Not a fan of statements that have no data to back them up! I can unequivocally say this statement is false. I’m personally running one the fastest plotters on the community leader board, and every system I’ve built uses the Asus hyper m.2. I’ve also now built three additional systems based on my original design for clients. I don’t deny there are MANY successful approaches to this endeavor, but why the random shade??
now there’s a true statement backed by facts , but like the ryzen 5900x they can still be found on the second-hand market and even with the markup still make a great value/per plot
The biggest bottleneck is the growth of the network and the time that it takes to build the machine – we probably have only a few more days of reasonably easy plotting. I’d say at this point cost/benefit calculation is not as important as getting the machine up and running quickly, and used corporate machines are the best bang for the buck. That being said, I realized that I’m not answering the question that you’ve asked. I’ve built 10 machines for plotting, and 2 of them I scratch built.
first system
used Gigabyte AORUS X570 Elite
used 1X AMD Ryzen 9 3900X 12-Core Processor @ 3.79 GHz
64 Gigs of ram
2X Samsung 970 Evo 1TB
1X .5TB Hynix NVME
ASUS Hyper M.2 X16 PCIe 4.0 X4 Expansion Card
4X Sabrent Rocket NVME 2TB
Running on Windows 10, typical run: 15 plots 40 min 3000 4 threads, when run in strings plots take about 8 hours to complete
System cost ~$1,300, NVME cost – $1461.92
Second system –
MSI B550 MEG Unify AMD AM4 ATX motherboard
1X AMD Ryzen 7 3800X 8-Core Processor @ 3.90 GHz
ASUS ASUS Hyper M.2 X16 PCIe 4.0 X4 Expansion Card
4X Hynix 1TB
MSI B550 MEG Unify AMD AM4 ATX motherboard
Typical run – 12 plots at a time, 40 min offset, 4000 ram, 3 threads, 7.5 hours plot time
System cost - ~$1100 NVME cost – $536
I had spent waaaay to much time trying to get these machines to POST, assembling them, figuring out BIOS settings, trying to get all the nvme cards to show up, etc. The Ryzen 9 machine can definitely do more, but keep in mind that an over-ambitious gets you nothing and a conservative one puts plots onto harvesters. Things that contribute to the better performance of the Ryzen 7 machine is the fact that it has 3 on-board nvme slots in a more expensive motherboard, and those Hynix nvmes are small, but very very good. I’d say that as bang for the buck goes, these are about equivalent, but you got to figure that build and tuning time is not free, so probably go for a bigger / faster machine, but don’t get too hung up on specifics. Also, go for 64 gigs of ram – the second machine is ram-bound.
“Not great” is an expression of opinion, and as such is 100% backed up by a singular fact that I hold this opinion. Here’s why I hold this opinion: as a card without an on-board raid controller it only works correctly if you have bifurcation support enabled or if your motherboard has it enabled by default. It simply does not work with many motherboards. Figuring out what’s happening when only one or sometimes two nvmes show up, but others do not is pretty frustrating. You are also need a 16x slot – so in most cases you can only have one of these per machine. The silicon strips this card uses are pretty annoying – replacing nvmes is a pain in the ass. This is all very good if the card is $60, but if you need to spend $150 and wait a week to get one, you are better off picking up a HighPoint SSD7204 – that card works right out of the box, has hardware raid and can go into either x8 or x16 slot. I guess these will sell out soon too.
I’m doing 10tb/day in total on 2x Mac Mini M1s, an i9 PC, and an R7 PC. I’ve got scripts moving plots to the farmer all the time so it’s not easy to figure out how many plots each Mini is completing but I’ll see if I can work it out today.
I’m also still fine tuning this set up - I think I can get to 14tb/day with 2tb/day each from the Minis.
There’s another way to look at “best bang for the buck”
and that is to ask yourself, can your bankaccount and house/garage space keep up with what you are plotting.
The cost of plotting is almost insignificant compared to the cost of storing the plots.
6-20 cent to create a plot and 2-3 dollar or so to store them.
So I think a better question is: how fast do you want to create plots? or how many harddrives do you want to buy per month?
Specific toy your system suggestion
5900x has 12 cores, that means you could do 12 plots in parallel, needing:
3+TB ssd, 48-64 GB RAM
5950x has sixteen cores, so can do 16 plots parallel:
4+TB ssd, 64 GB RAM
I don’t know about about Asus raid cards, but there are many options for ssd’s
Very expensive enterprise ones that are fast and will last long time. Better price/plot, but only if you intend to spend around 200.000 dollars on hardrives in the next two years or sell plots to others.
Lower quality enterprise ssd’s, still good durability and price
High end comsumer ssd, higher price/plot but less initial investment. (what I am using)
I’m starting to ramble sorry
My point is, you want to ask yourself how fast and how long you want to plot, considering the price of hhd you need to store your plots.
Best bang for the buck is generally not the newest top of the line hardware.
You need to balance your cores/ram/ssd so it al matches the amount of plots. If the system is balanced you can build a good value system for 6,8,12,16 cores depending on your budget
This is truth. The real issue is long term storage. That’s why I’m so happy we racked a 45-drive Supermicro JBOD and I have two of them built up, only one up and racked at the moment at a cost of $250/month … I can just drive over to the datacenter (10 mins) and slot the drives in as they full up, my ideal is to have 45 18tb drives, or 810tb. I won’t quite make that since I’m getting a mixture of drives.
The electricity alone to run that is $40/month btw, and that’s with a power efficient 30w “brains” server, kaby lake Xeon.
Don’t listen guys above. You shouldn’t buy AMD CPU. Most of all AMD doesn’t have integrated graphic chip. And right now you can’t buy good videocard (because of ethereum miners they costs x3). So you’ll need to waste 100-200$ for useless gtx710 or something like this. Better buy Intel, because most of all of them have good enough integrated videochip.
, but like the ryzen 5900x they can still be found on the second-hand market and even with the markup still make a great value/per plot
This is truth. The real issue is long term storage. That’s why I’m so happy we racked a 45-drive Supermicro JBOD and I have two of them built up, only one up and racked at the moment at a cost of $250/month … I can just drive over to the datacenter (10 mins) and slot the drives in as they full up, my ideal is to have 45 18tb drives, or 810tb. I won’t quite make that since I’m getting a mixture of drives.