Hey folks, I have a question that may have been asked a ton of times, but from what I was able to search at the forum I’m not completely sure about the answer, and this is why I’m creating this thread.
I’m almost pulling the trigger on a Ryzen Pro 5650G to build a home server. I think it has good idle power consumption and it official ECC support. So far, so good. The issue is, Ryzen with ECC is such a nightmare because you’ll see all kinds of reports saying that it works, others saying that it does not work, some say that it partially work, and so on.
There is any documentation or real evidence showing that it indeed works, and which motherboard should I use to have full ECC support?
Right now I have a Synology and a bunch of cloud storage accounts that I would like to aggregate on a local server. I don’t game or have any plans to stream content from the NAS, the ECC part is pretty much for the my photos and videos, nothing more than that.
In general, consumer boards don’t support ECC unless explicitly mentioned in the spec’s, enterprise boards usually do support ECC as standard. Brands to consider include SuperMicro, Asrock Rack and Gigabyte.
Asus have confirmed ECC support on most of their AM5 lineup and some reports suggests that ASRock also have it on at least some of their boards. AM4 is a bit of a different story and I would stay away from Gigabyte by far given their track record and aftermarket support or rather lack of.
Caution! “Supported” does not mean “make use of ECC”!
Almost every board is capable of running ECC memory like a none ECC module.
That is why almost every board “supports” ECC, while at the same time most boards can’t make use of ECC.
Yes, this is my main worry because I see lots of reports of ECC systems where the owner nerver saw an error correction happening, so how can they be sure that it’s working?
I’ve seen very good reports from Asus and Asrock. But still not sure if it’s version specific or even if it’s AM4/AM5.
I’m considering this, but it’s quite expensive. If I ever get to this level, I think I would rather just stay with Intel using W680 as it’s less of a hit and miss like AMD.
Does anyone know if there is any list of components that have been tried and are known to be working with ECC?
W680 is a workstation chipset so it is known to work with ECC, if the cpu you put in support it as well. Historicaly i3’s, pentiums, xeons and celerons used to. Thoe more recent cpu’s I can’t tell but you can verifiy it on Intel ark.
You’ll have to check for the price of the boards, the ones made for workstations are quite expansive.
If ECC is important to you as opposed to a nice-to-have, then I would just go for AM5 EPYC. Here is an example Mobo+CPU+RAM combo courtesy of Newegg, to give you the ballpark numbers:
To the best of my knowledge there is no list. You’ll find several threads here by folks trying to validate EC4 SECDED, though.
Zen 3, 4, and 5 IMCs and AGESA all support ECC. It’s possible it didn’t yield on non-Pro APUs (5600G in this case) but I suspect in most cases it’s just crippled for no reason other than AMD marketing wanting to have it as an upsell. There was one AGESA version shortly after Raphael’s premature launch where AMD broke EC4. Which probably wouldn’t have gotten nearly as much attention if ASRock hadn’t been honest enough to remove ECC support from its mobo specs and then put it back when AMD fixed ECC. IIRC it was like a week or two.
MSI boards and most Asus and Gigabyte boards don’t expose AGESA’s ECC enable in their BIOS menus. Full line AM5 support on ASRock desktop, no need get into ASRock Rack pricing and lack of BIOS updates. I think also full line on ASRrock AM4 desktop but haven’t checked specs on as many of the boards. All the ones I built support ECC, anyways.
I don’t see why some of the posts here are suggesting spending more on EPYC 4004 or 4005 when the corresponding Ryzens support ECC. Same IO die and AGESA. If you’re unwilling to trust consumer platform ECC or do your own validation then I suppose the money for workstation or server sockets and RDIMMs is the remaining option. But, unlike ASRock AM5 desktop and Asus ProArt, I’ve encountered exactly zero third party validations of Threadripper and EPYC EC4 or EC8.
If anything there’s probably less third party validation of W680 boards than AM5. AM4 I haven’t seen get much attention either.
Just a branding thing. It’s more accurate to say that, unlike AMD, Intel cripples EC4 off B and Z boards and lower end processors (235 and up have support, 225 doesn’t). Not like DDR connects to the south bridge, after all.
It’s possible AMD and Intel might differ in how much validation they do of mobos using their chipsets. But I haven’t seen anything. Outside of engineering samples, I suspect the answer’s zero in both cases. Unless maybe a customer worth at least a few million complains.
Thanks for all the suggestion folks. I would not mind to go with an Epyc setup, the only issue is that in Europe it costs at least 3x that price
Do you mind posting the boards that you’ve tried? I’m inclined on going with AM4 given that DDR5 ECC is quite expensive and the server doesn’t need to be that powerful, a 5650G would be plenty for years and years to come.
I verified ECC working (including reporting to the OS) on Ryzen PRO 4650G and Asus ProArt B550-CREATOR on AGESA 1.2.0.E at least. I’ve seen sporadic reports of others verifying other motherboards too with AM4 Ryzen PROs (search for other posts here and on reddit for example).
In general, if you get something like the following from the kernel I would feel reasonably secure about fully functioning ECC support (including reporting as long as you disable “Platform First Error Handling” (PFEH) in the UEFI):
Edit: Note I would not trust dmidecode; as I understand it that only reports the UEFI’s interpretation of what’s in the DIMM SPD.
I haven’t found any other platforms than the AM4 Ryzen PROs that both support ECC and have reasonable idle power draw. (Except perhaps Epyc 8024P, but that’s a awful lot of money to pay for “validated” ECC support and a lot of PCIe lanes. Especially given the atrocious firmware support from the available motherboard manufacturers.)
The Hawk Point Pro APUs are similar but probably even harder to get if you’re not an OEM. Strix Point Pro should be similar and with impossible non-OEM availability for a while once it releases. Since both are DDR5 the non-Pro APUs’ lack of EC4 support’s less of a consideration due to ODEC2 and read CRC.
Arrow Lake’s idle is ~14 W, so might be an acceptable compromise between Zen APUs and desktop Ryzens depending on build objectives.
That’s incorrect and why would you go for a board that’s poorly supported? I mean, we’re on AGESA 1.2.0.3d and it’s still stuck on 1.2.0.1. B650 itself is also on its way out even if the newer version doesn’t offer all that much in terms of new functionality.
Just grab ASUS ROG Strix B850-G Gaming if you need mATX and it’s ~120 EUR cheaper too (with Intel NIC) unless you really need IPMI. That being said, it only has one “desktop” PCIe slots but a bunch of M.2 ones depending on your requirements. If you want more connectivity, ASUS ROG Strix X870-A Gaming WiFi might be of interest.
@lemma
I agree, that seems rather pointless for “home labs” unless you know you need like SEV which possibly is also available on Ryzen 9000-series but I can’t easily verify, a simple dmesg from a fairly recent version of Linux kernel should tell though.
@fenugurod
Given memory frequency scaling with 4 DIMMs I would recommend that you go for a single 32G stick if you think you need at least 32Gb of RAM. It will be slower than using two sticks however given your usecase it’ll be negligble and it’ll be a much nicer upgrade route if you need more RAM further down the road. To be honest, the 16Gb sticks are also so expensive that you might as well go for 32Gb because it gives you a much lower EUR / Gbyte ratio.
MSI’s just started putting out 1.2.0.3e. I can see BIOS update cadences on server boards being more conservative than desktop parts but B650D4U taking 11 months to get EPYC 4004 support and still missing security patches, performance fixes, and compatibility improvements in 1.2.0.2 and 1.2.0.3 that released months ago is IMO absurd. If anything I’d expect server parts to get patched first and get whatever 3e is doing to expand 256 GB support first.
The only EPYC 4004 and 4005 specific feature I’ve been able to identify is server OEMs can turn on TSME if they want. So no differentiation apparent for homelabbers.
My bad. It is listed as DIMM, ECC, Unregistered in Geizhals, I am not well researched on ECC and this was meant as a way to base the expected costs. If you have a similar single- or double stick recommendation, I am all ears.
You should consider the type of ECC setup before choosing a motherboard. ECC UDIMM’s just notify an error and the OS must deal with the outcome. If your not using an ECC aware OS, nothing happens, though you may get better quality RAM by choosing UDIMMs. RDIMM’s (Registered ECC) are the real deal, almost never down, Error Control and Correcting memory but require a different slot so a server or high end workstation motherboard is required.
I’ve been running a AsRockRack B650D4U with a 7800x3d (but will take an Epyc 4000 type) on Win11 and 32g of UDIMM and it’s been up for 1.2yrs, rebooted every 3 or so months for software patching and updates. Love it. Solid as a rock, though there is a slow boot as most servers do check everything and boot an internal BCM before everything else. And the unaware staff hear the 7 high-capacity spinning drives do their POST in sequential order instead of all at once, so they think it has a failing drive that clicks for a whole 40 seconds on boot.
This board was chosen because of Wendel’s review, and has improved over time with bios and driver updates, as most AMD products seem to do. The performance in my case has increased about +6% overall, with spikes of +15% under high load with video transcode and disk access at the same time. That’s a whole jump in processor part number worth.
Get a good case that has plenty of ventilation and fan locations to plan the in-out flow for best performance and longevity. This is smart for any system and the careful planning here can make all the difference in high load environments. For longer life, get quality fans with dual ball bearings or similar high quality solution (Noctua) - many AIO liquid coolers just don’t have the stuff for a 3-5year uptime.
