I have got hold of some old server hardware recently and have been experimenting on servers with multiple CPUs and multiple cores and here is a summary of my findings.
I have been using cheap hardware so no SSDs, but lots of memory (as old server memory is cheap, and most old servers have lots of slots)
Linux - Not Windows - lots of people say that Linux is more efficient at plotting than Windows, on desktop hardware this might be 5-10%, but on server hardware it will come down to how your OS handles the NUMA architecture. I found on Windows that MadMax would only use half of the cores, but all of the cores in Linux (Ubuntu 21.04). I am sure you can overcome this with tweaking in Windows, or running plots in parallel, but it is so much simpler in Linux to access all of the hardware (I am a windows person, but getting Linux up and running with MadMax was really simple). With the older HP server Ubuntu 21.04 does not install well in desktop mode - so you either need to do the trick described by Fatviking above, or just install 21.04 server and use command line only - bit of a faff, but doable with Google support!
No SSD, so be creative with your server disks - SAS disks are cheap, so buy as many as you can and make a RAID0 array. I am finding that 4 x 15k SAS drives is a good middle ground of cost and performance. I use this for my 1st TMP drive.
No SSD, so RAM Drive is important for TMP2 - so use 110GB RAM drive, this only leaves 18GB of memory if you have 128GB - I have noticed this is quite tight, especially with lots of cores running, so if possible load up with even more server memory - 144GB is a sweet spot for non-parallel plotting.
USB3 or Network for destination drive. I have some really old servers with only USB2, this is a bottle neck, so on these servers copy the file to your farmer using network - 1GBps if possible. Obviously you can leave the plot on the SAS RAID for maximum plotting speed, but they will soon fill up.
So my approach is: Server → USB 3 for plotting (off network in the garage!). When disk is full, move disk to low powered laptop for Farming. You don’t want to farm on a server as heat, power and noise will be a problem, unless your are into industrial level plotting. I am finding that moving from SAS to USB using the destination setting in MadMax adds about 10 mins per plot (after the first plot) as it is copying the files off the disk while creating the next plot.
I am only up to 30TB of plots, but with a server doing the plotting, this can grow really fast. Prior to the server hardware I was doing 12Hr Plots on my farming laptop, 2.5Hr Plots on my desktop PC Ryzen 2400G (but only when not using the PC). But with servers I can do 30 min plots on my best server or 60 min plots on a really cheapo G7, so filling up a 5-6TB drive is 1 or 2 days work.
Some examples are:
HP DL360 G9 - 2 x E5-2650v3 - 256GB = 30 Mins Per Plot (2 x 10 Cores with Hyperthreading)
HP DL360 G9 - 2 x E5-2620v3 - 192GB = 40 Mins Per Plot (2 x 8 Cores with Hyperthreading)
HP DL360 G8 - 1 x E5-2620v2 - 96GB = 90 Min Plot (single CPU and no RAM Drive)
HP DL360 G7 - 2 x E5645 - 144GB = 60 Mins Per Plot (2 x 6 Cores with Hyperthreading)
All of the above using SAS Raid 0 for the TMP drive and RAMDisk - where enough memory is available
I figure you can get some half decent G7 dual CPU kit for < £200/300 in the UK and that will churn out plots in 60 mins - so run that for a few days to get your farm going - run it for a week or two and get a decent size home farm. Then move your disks to an old laptop or Pi and do your framing - when ever you get a new large disk - power up the server and fill the disk up and move it again. Obviously if you can get hold of G9 kit it is easier to install Ubuntu and much faster to plot, but much more expensive to buy…