In 2017 buying old Xeons bought you pretty good performance for the buck, as Intel sold you quad cores with minor single core and multicore uplifts for a lot of money, while at the same time old Westmere or Sandy Bridge Xeons had a pretty good balance of many cores with decent enough clock speeds / single core performance. Even I/O options did not improve by a lot in those years.
In 2022 buying old Xeons means you get stuff from the era where Intel started heavily deprioritizing single core performance over having a lot of cores in Xeons. 14+ lowly clocked Haswell cores still do well in synthetic benchmarks, but I don't think they're worth it for any real world workstation loads. Low clocked Haswell cores are outperformed by such things as the Alder Lake efficiency cores (easily) and even modern Android phones. The latest Zen cores and the Alder Lake performance cores will be worth at least twice the per core performance of one of those individual Haswell cores. Even if you say compile code all day it's probably not worth buying into a 200W behemoth that runs Firefox slower than the Android phone in your pocket.
This blog was written in 2019 and was probably already not great advice then, but it is legit bad advice now.
Nowadays, old server processors have very poor single threaded performance compared to recent desktop grade hardware, and the much larger amount of cores fails to compensate for that in pretty much all workstation loads. Especially considering that you can get a lot of cores in desktop hardware nowadays too. The relative value of desktop hardware vs server hardware has gone up substantially in the past 5 years or so, since Intel started making server chips focused exclusively on the demands of large scale data centers and AMD started normalizing high core counts in desktop chips.
Because cheap second hand server chips are sourced from data centers, that inherently favor high core counts and only baseline single thread performance, I don't see a future where we have cheap server processors ever being a great choice for a workstation again. Quick-ish single thread server chips do exist (e.g. Epyc 72F3), but I think those are too niche to ever see them for a good deal on eBay.
I’m certainly a power user and suspect that additional cores may actually benefit me, but there are some very interesting points about clock speed from you and others.
I wish that was true. Lots of code I deal with bottlenecks on a single thread. Mozilla has abandoned Servo, and we are stuck with browsers poorly parallelizing parsing, painting, execution. Photoshop, aside from a few filters and functions, is also poorly parallelized. Not to mention smaller programs who understandably can’t justify the complexity of multithreading.
I’d trade single tread performance for cores any day.
According to pstree:
3*[Isolated Web Co───28*[{Isolated Web Co}]]
3*[Isolated Web Co───29*[{Isolated Web Co}]]
2*[Isolated Web Co───26*[{Isolated Web Co}]]
Isolated Web Co───27*[{Isolated Web Co}]
Privileged Cont───25*[{Privileged Cont}]
RDD Process───2*[{RDD Process}]
Socket Process───4*[{Socket Process}]
Utility Process───2*[{Utility Process}]
2*[Web Content───9*[{Web Content}]]
Web Content───10*[{Web Content}]
WebExtensions───28*[{WebExtensions}]
chrome-gnome-sh───2*[{chrome-gnome-sh}]
firefox-bin
129\*[{firefox-bin}]Don’t get me wrong, it’s still a huge leap forward to have a single webpage running in its own process, but there’s ample room for improvement.
That’s what GP is referring to.
Single core: 61.1
4 cores: 73.7
Speedometer tests a bunch of ToDO MVC apps in many popular frameworks.
It's very heavy on DOM manipulation, parsing, painting, etc.
It's the hard to parallelize stuff and stuff that's supposed to represent real world usage.JetStream 2 is easier to multithread since it plays with Web Workers and other tricks, but still, results are depressing:
Single core: 54.4
4 cores: 82I'd love for it to be plug and play.
Certain boards (mostly ASRock) will boot with it, but AMD don't support it outside the Pro series, and you won't get full reporting.
I've been down this rabbit hole for a while with NAS builds, and it's simply not worth AMD if you need ECC currently for anything not too high end.
If it doesn't boot, it's very likely that you actually bought registered memory, which is normal in servers, but _not_ supported at all by Ryzen processors, no matter the mainboard.
The Ryzen story is a joke if you are actually in the market for a workstation, meaning you want ECC that works. The only way to get it is to buy integrated OEM workstations with the TR Pro.
PS C:\> wmic CPU get Name
Name
AMD Ryzen 9 3900X 12-Core Processor
PS C:\> wmic Baseboard get Product,Manufacturer
Manufacturer Product
ASRock B450M Pro4
PS C:\> wmic MemoryChip get Manufacturer,PartNumber,BankLabel,Capacity,Speed
BankLabel Capacity Manufacturer PartNumber Speed
P0 CHANNEL A 17179869184 Kingston 9965745-002.A00G 2667
P0 CHANNEL A 17179869184 Kingston 9965745-002.A00G 2667
P0 CHANNEL B 17179869184 Kingston 9965745-002.A00G 2667
P0 CHANNEL B 17179869184 Kingston 9965745-002.A00G 2667
PS C:\> wmic MemPhysical get MemoryErrorCorrection
MemoryErrorCorrection
6 # "6" is Multi-bit ECCSearching for a simple answer to question: "does this CPU + chipset + this particular motherboard actually support ECC DRAM" is infuriating.
Thousands of words on the RGB lighting and the overclocking tweaks, silence on ECC support. It's insane.
Went through four rounds of ECC DRAM (that was on the manufacturer's supported config list) on a Gigabyte board, never got it to POST.
I finally gave up on that one. Bought an Asus X99 WS setup, with Xeon and RAM, from a vendor on eBay.
I can fix simple things, but debugging a configuration in this gamer-saturated market is beyond my current skill level. Once I have at least something working end-to-end, I can bootstrap up to a working machine. But the dead stare of failed POST was a tough one.
It's outdated if you need a high-performance workstation. As a regular desktop PC, it runs just fine. Would work quite well as a NAS CPU, with ECC support and downclocked to some lower P-state.
However, I don't think these were ever available for $50 as author suggests. It all really depends on the price point. At $50, I would buy 1 right now, just to have as a backup/guest PC in the living room.
It took me an obscene amount of time to get macos catalina running on it, however. The latest firmware and/or quadro k5000 just make it permahang.
PS: https://www.reddit.com/r/Amd/comments/lh3m42/demystifying_ry...
https://www.reddit.com/r/Amd/comments/ggmyyg/an_overview_of_...
Ryzen CPUs without an iGPU _will_ support ECC UDIMM, _unless_ the mainboard specs specifically tell you that it won't support ECC, or simply omit mentioning ECC UDIMM at all.
Fwiw, I've enjoyed proper ECC on an ASRock Fatal1ty B450 Gaming-ITX/ac with a Ryzen 5 3600 for more than two years now (using GNU/Linux; I am not sure how Windows would fare).
This said, there are some Xeons which run on moderate to somewhat litle power. You might get one or two used of them on ebay. I consider to do this for my home server some day.
Solid state components are generally very reliable anyway.
Or not.
Yes, it was a great advice in the 2010s -- I had such a Dell, too. It's well documented how IPC only has grown 20-25% from Sandy Bridge to Skylake and then the wait was long, long until a new uarch has emerged in 2021. But towards the end of the decade, Ryzen has disturbed the still waters and the number of cores have, well, risen sharply while keeping a competitive IPC and Intel eventually followed suit. Because of this, the equally long time rule of thumb where most apps don't really use multiple cores well so you don't need them on the desktop is also dead by now.
I wouldn't recommend a Haswell Xeon as this article does today.