I’m gonna do it! I need to build a new WS, but its been awhile (a very long while) since my last build and I’m not up to speed on the latest hardware. So, this video seems like a good place to start:
A Workstation for what? Programming? Audio? Blender? Digital art (of which, Photoshop is still the unrivaled king)? Video editing?
Linux is great at many tasks and can do more and more tasks without hiccup but it is best at Programming and Audio stuff. 3D modelling, it is getting there with Blender although still not quite where it needs to be. Drawing and video editing, alternatives exist but you need to cut the cord on Adobe first for a sub-par experience. There are still no real competitors to the Adobe pipeline and until there is they will rule the roost as sure as Microsoft rule the roost of Enterprise computing with AD.
Computers these days are also ridiculously powerful and the only reason to get, say, a Threadripper system is if you need to be able to compile 200 GB worth of source code within a single hour. Mind you, a 7900X can do it in an hour and a half for a third of the price.
RAM, 64 GB is the bees knees, higher is possible to reach but again, requires Threadripper to truly utilize it. 128 GB on AM5 is possible but you will run at 3600 MHz instead of 6000 MHz.
Graphics Card, if gunning for a programmer machine I would get a 6700 XT, if gunning for a high end Machine Learning platform then you want the 4090, despite the abysmal Linux support Nvidia keeps providing*.
With this in mind here is my own personal preference for the workstation core, I did not include a cooler ($50+), case($100+) or PSU($150+):
*Nvidia Support is fine, the cards work just fine minus a few interesting features, but the integration is far from excellent and far below the level of support AMD and Intel provides.
Is ECC something you need? Because that’s easily another $1000+ for proper platform support, and all it does is provide insurance to something that only happens like… Once a year, tops?
I can understand it if you are doing something like, say, bank trading. If a data integrity error doesn’t cost you lives or large sums of money, don’t really bother. Not worth it.
A year and a half of my professional career has been spent tracking down software issues that were caused by the lack of ECC memory.
The follow up on one of the events resulted in them firing my boss’s boss, and the bosses above him for 5 levels. In that case I was the in house evaluator of a third party software package. The developers has a team of over 60 people attempting to diagnose and reproduce the errors that I was getting. I had been attempting for 3 months to be allocated different ram, as had my boss, but the ram allocation I was using had to be moved with me from computer to computer as the computer lab we were allocated only had bare computers, without ram or hard drives. I was a contractor, and there was a ram tester that employees were allowed to use, but as a contractor I was denied. I basically begged for my ram to be tested for several hours next to the guy manning the box until a high level employee tested it with his employee number, and it was found to be faulty.
In another case 30 developers were working to track down a software bug for about 4 months. The computers were using vmware with high availability. Someone accidentally powered off the problem computer, and no one noticed until the bug we were diagnosing stopped occurring, and we found that the computer that it was running on was different. Upon examining the hardware before powering it back on, I noticed that the DIMMs had 8 chips each instead of 9. We replaced the ram with ECC ram, and had constant failures of a DIMM, no matter the piece of ram in in the slot. The problem was solved with a replacement motherboard.
I do not see how ECC would have improved this situation since the RAM stick itself was faulty. My take is “Don’t use faulty RAM and make sure to test it first for anything important” - Something you should always do and most Linux live distros have a memory tester available - rather than ECC being a necessity here. The only thing ECC would have done is report the RAM stick as faulty sooner than it was. Also, moving RAM around like that is just… Why? Wasn’t it cheaper to just buy the RAM sticks and insert them to begin with? Don’t people realise the friction caused by constant insertion and reinsertion eventually wears the contact out, just like it did with NES and SNES cartridges?
Again - Motherboard failure. ECC helped with diagnostics, but did it really prevent any error occuring here?
What you are telling me is that hardware failures exist and ECC RAM is very good at detecting hardware failures quickly. Is this then worth $1000 dollars extra per machine (motherboard, CPU and RAM combo)? If you already have the motherboard and CPU, sure, at that point the $20 extra for ECC RAM is worth it. But if you are on a Ryzen or Core platform, an equivalent Threadripper / Xeon / EPYC will cost you $500-$1000 more.
ECC was designed to detect bitflips and correct them. Bitflips happen once every three weeks of constant usage and 95% of the time that happens in some unimportant part of the memory, causing a pixel of a video frame to become black or flipping a website loaded into memory that gets wiped the next time you reload a website. If you are really, really unlucky, it could theoretically change a dot to a zero making you pay $10000 instead of $10.00. This is unlikely to happen though.
When data must be preserved ECC is the obvious choice. When data is NOT critical, like family photos, your warez stash or a steam folder, then ECC is nice but not necessary.
That said, I do think that ECC should be mandatory on all RAM sticks chipped from now, period. I also do think that ECC is a feature that should permeate everything, and the fact that we are still arguing about it in 2023 is just bonkers. But hey, that’s capitalism for ya, if ya can cut a corner and fleece the customer some more, many will.
The first case occurred in 1996 on a PowerPC platform. While the internet was a thing by then, I didn’t know of a software package that I could run to test my ram. The ram tester that they used cost about the same as a new car.
The second case they were dealing with transient data, but the client facing application kept crashing.
Speed, reliability, good price, choose two. I will usually choose reliability and good price for my server builds, as I don’t need speed. And then buy the previous generation parts for a computer that is not in desktop iuse. If it is available for 20% more price on whatever system I am purchasing I will always choose something with ECC. However it is often not an option on certain devices (laptops) without paying several times more, then I choose.
ECC uses 9 bits per byte, on a 64 bit data bus, it needs 72 bits. The CPU socket and ram can both support ECC, but if the motherboard did not put the wires in their motherboard, the feature will not function.
This is why you have to look for motherboards who specifically support ECC.
