X399 MB ECC Confirmation/Clarification Thread

Joined
Apr 7, 2012
Messages
55
I am creating this thread to welcome sources of confirmation to clarify or confirm the degree to which each X399 MB has support for ECC. I imagine this is a subject of great importance in this community, since Threadrippers are bound to end up in a plenty of workstations.

I've been trying to find a support email address for MSI that doesn't require a ticket or support ticket, and it seems that the [H] community rep is no longer around, so if anybody can accomplish that please do so.

I am forwarding this thread to the Gigabyte [H] community rep, so let his PM box rest.

I have yet to see anything regarding Load Reduced ECC RAM so that seems to be out of the question for the launch MBs.

First up is Asrock:
They have 2 boards coming up for launch, and they do specify it supports ECC UDIMMs but do not state that they operate in non-ECC mode only. There is mention of Data Poisoning settings in the user manual, so it sounds like a yes.
X399 Taichi
Fatal1ty X399 Professional Gaming

Second up is ASUS:
ASUS put up their pages for their MBs now. Only the Zenith Extreme has any mention of ECC support, and the Prime obviously doesn't. They have only one mention in the Zenith's manual about ECC, and they do not state if it is EUDIMM support and no ECC mode, or what have you.
ROG ZENITH EXTREME
Prime X399-A

Third up is Gigabyte:
While they have had support for the AM4 Ryzens, and they were looking to be a safe bet from this display, and the product landing page doesn't have any exclusion clauses, but the specifications page states Support for ECC Un-buffered DIMM 1Rx8/2Rx8 memory modules (operate in non-ECC mode)
X399 Aorus Gaming 7
(current status: supported)
Update 8/10/17 9:00am American CST: Gigabyte has updated the specification's page, and now shows ECC support.
The Gigabyte [H]REP has also confirmed ECC mode.
X399 DESIGNARE EX states support in the product specifications. When it was introduced it did not state anything.

Fourth up is MSI:
The only entry on their website at this time is the X399 Gaming Pro Carbon AC and it states "Supports ECC UDIMM." It doesn't mention anything about non-ECC mode only or anything.
*Update- got some ECC confirmation so it should be good. This a great about face from their AM4 support.
*Update (10/20/2017) - With the introduction of the SLI Plus MSI has removed mention of ECC altogether on both of their x399 options.
X399 Gaming Pro Carbon AC
My Pro Carbon ECC confirmation link
X399 SLI PLUS (Current status: unconfirmed)

Edit for 8/13/2017: You can simply test when you get the board in windows with the following commands in power shell:
wmic MEMORYCHIP get DataWidth,TotalWidth
If the TotalWidth value is larger than the DataWidth value then ECC should be enabled. 72bit indicates ECC function.

A more detailed description of functions can be gotten from the following
wmic memphysical get memoryerrorcorrection
The values should be:
  • 0 (0x0) Reserved

  • 1 (0x1) Other

  • 2 (0x2) Unknown

  • 3 (0x3) None

  • 4 (0x4) Parity

  • 5 (0x5) Single-bit ECC

  • 6 (0x6) Multi-bit ECC

  • 7 (0x7) CRC
 
Last edited:
I'm rolling the dice and trying my registered modules on the asrock when it arrives. Worst case I'll yank some unbuffered out of another machine to finish the build and go order some more. Still hate having to source unbuffered, its too niche.

LR dimms do require some extra lifting by the motherboard OEM, but for X99 at least registered is all about the cpu IMC. (granted often you end up with stupid boards not handling the ECC or but it still boots if they at least put the cpu ids for the xeons)
 
LRDimm is reserved for Epyc per AMD.

Awesome thread! Would those who are using ECC memory with Threadripper let the rest of us know what works? It seems all the ECC memory I've found are registered and not unbuffered...for example, I went on the crucial website and it doesn't appear they make unbuffered ECC DIMMs. Thanks!
 
Here is another VLP EUDIMM for those wanting capacity. I originally tried to order these, but they ended up being backordered until late September, and now it seems the price increased by $100.

I already have 4 of the ct16g4xfd824a in my MSI board, I'm just waiting on a waterblock for now, then I'll verify functionality. I'm starting to hate this waiting game.
 
Hi, about to pull the trigger on my system...it sounds like the Aorus Gaming 7 is the only one with confirmed ECC mode operation at this point? Thanks.
 
Hi, about to pull the trigger on my system...it sounds like the Aorus Gaming 7 is the only one with confirmed ECC mode operation at this point? Thanks.
So, did you get the aurus board? Can you confirm functioning ECC?

Thanks
 
I'm rolling the dice and trying my registered modules on the asrock when it arrives. Worst case I'll yank some unbuffered out of another machine to finish the build and go order some more. Still hate having to source unbuffered, its too niche.

LR dimms do require some extra lifting by the motherboard OEM, but for X99 at least registered is all about the cpu IMC. (granted often you end up with stupid boards not handling the ECC or but it still boots if they at least put the cpu ids for the xeons)

So what happened? Did LRdimm or udimm ecc work for the asrock taichi?
 
Hmm. All the unbuffered ECC dimms I tagged above are out of stock now. I hope the memory situation gets better around Black Friday..
 
Is there a link to keep an easy track of progress on this?

I'm hoping for the same thing. Best resource I've found is Level 1 Tech. Just doesn't seem like many are looking at TR as a poor man's EPYC which is what I would like to do and hence the ECC memory and IOMMU.
 
It's a shame as I was looking forward to building a VM server with Threadripper. I hope it all gets sorted in the next couple of months.
 
Just doesn't seem like many are looking at TR as a poor man's EPYC which is what I would like to do and hence the ECC memory and IOMMU.

We're around. A proper Supermicro TR board with ECC and IOMMU done right would sell like hotcakes.
 
Reporting in with my findings. I had been through hell getting my setup to even boot with the MSI x399 Pro Carbon since there was only one bios that would load any boot manager (7B09v14) and one BIOS (7B09v13) forced me to use the BIOS flashback function. Non of them worked with the SATA controllers set to RAID. Which forced me to do a from scratch approach and migrate data over through Backup images in Macrium.

I did spot Data-poisoning functions in the BIOS though so :athumbsup:

Here are my Windows results for ECC Confirmation:
DataWidth TotalWidth
64 128
64 128
64 128
64 128
and

MemoryErrorCorrection
6
This part indicates that multi-bit error correction gets set without any user input on the v14 BIOS.

I find it weird that TotalWidth gets indicated as 128, but I guess the dual memory controllers on the socket have something to do with that. I'll do a BSD test later and see if I can get better details. I'll update this post when I do that since it serves as the confirmation post.

So the x399 Pro Carbon passes for ECC support. The current BIOS selection just eats a bag of dicks for my setup.

This is the ECC RAM I'm using since it was the only one that was going to have any stock in time. So if anyone is considering compatibility the v14 BIOS is the best bet I'm afraid.
 
Last edited:
This part indicates that multi-bit error correction gets set without any user input on the v14 BIOS.

So the x399 Pro Carbon passes for ECC support. The current BIOS selection just eats a bag of dicks for my setup.

Thanks, this is great to know when our local Intel shills try to FUD the ECC topic.
 
Hi, I bit the bullet and purchased an Aorus 7 MB an a 1950X. Powershell shows DataWidth as 64, TotalWidth as 128, and MemoryErrorCorrection as 6. What is the difference between single-bit (5) or multi-bit correction (6)? The memory I am using is 4 sticks of CT16G4WFD824A.

I have Windows Server installed with numerous VM's and everything is running fine, although in Device Manager, there are a few unknown devices and I am not sure what those are. One I found is AMD GPIO and I can't seem to find a driver for that. I have installed the X399 chipset drivers from the AMD website, but this device still shows up as unknown. One thing I am noticing is that I cannot do nested VM's (eg., host Hyper-V on top of Hyper-V). Any ideas here? Also, I have not tried to do SR-IOV.
 
Last edited:
Hi, I bit the bullet and purchased an Aorus 7 MB an a 1950X. Powershell shows DataWidth as 64, TotalWidth as 128, and MemoryErrorCorrection as 6. What is the difference between single-bit (5) or multi-bit correction (6)? The memory I am using is 4 sticks of CT16G4WFD824A.

I have Windows Server installed with numerous VM's and everything is running fine, although in Device Manager, there are a few unknown devices and I am not sure what those are. One I found is AMD GPIO and I can't seem to find a driver for that. I have installed the X399 chipset drivers from the AMD website, but this device still shows up as unknown. One thing I am noticing is that I cannot do nested VM's (eg., host Hyper-V on top of Hyper-V). Any ideas here? Also, I have not tried to do SR-IOV.
For a good explanation on Multi-bit error correction see this article, plus this one for a ridiculous amount of depth.

Basically it is the highest form of HW data correction available, and it is the most desirable function to run at no real cost to speeds at all. The level of accuracy it provides really fortifies an ECC RAM setup.

You want to see it supported and running. You can run with confidence even during some pretty nasty solar flares with an extremely low chance of data loss occurring.
 
For a good explanation on Multi-bit error correction see this article, plus this one for a ridiculous amount of depth.

Basically it is the highest form of HW data correction available, and it is the most desirable function to run at no real cost to speeds at all. The level of accuracy it provides really fortifies an ECC RAM setup.

You want to see it supported and running. You can run with confidence even during some pretty nasty solar flares with an extremely low chance of data loss occurring.

Thank you. System is running fine and very stable. I see that the memory is running at 2T command rate. Does anyone know what timings these sticks support at 1T? I looked on the Crucial website and contacted support, but did not get any useful info. The sticks are CT16G4WFD824A.
 
Next confirmation: ECC works on Gigabyte's X399 DESIGNARE EX board just fine.

I tested with 4x 16GB 2Rx8 modules (CT16G4WFD824A) and 8x 8GB HPE 2Rx8 unbuffered ECC modules.
In fact I was able to boot the machine with 4x 16GB + 4x 8GB resulting in 96GB ECC RAM :)

Linux 4.14 kernel boot messages with 8x 8GB modules:

Code:
kernel: EDAC MC: Ver: 3.0.0
kernel: EDAC amd64: Node 0: DRAM ECC enabled.
kernel: EDAC amd64: F17h detected (node 0).
kernel: EDAC amd64: MC: 0: 16383MB 1: 16383MB
kernel: EDAC amd64: MC: 2: 16383MB 3: 16383MB
kernel: EDAC amd64: MC: 4:     0MB 5:     0MB
kernel: EDAC amd64: MC: 6:     0MB 7:     0MB
kernel: EDAC amd64: MC: 0: 16383MB 1: 16383MB
kernel: EDAC amd64: MC: 2: 16383MB 3: 16383MB
kernel: EDAC amd64: MC: 4:     0MB 5:     0MB
kernel: EDAC amd64: MC: 6:     0MB 7:     0MB
kernel: EDAC amd64: using x8 syndromes.
kernel: EDAC amd64: MCT channel count: 2
kernel: EDAC MC0: Giving out device to module amd64_edac controller F17h: DEV 0000:00:18.3 (INTERRUPT)
kernel: EDAC amd64: Node 1: DRAM ECC enabled.
kernel: EDAC amd64: F17h detected (node 1).
kernel: EDAC amd64: MC: 0: 16383MB 1: 16383MB
kernel: EDAC amd64: MC: 2: 16383MB 3: 16383MB
kernel: EDAC amd64: MC: 4:     0MB 5:     0MB
kernel: EDAC amd64: MC: 6:     0MB 7:     0MB
kernel: EDAC amd64: MC: 0: 16383MB 1: 16383MB
kernel: EDAC amd64: MC: 2: 16383MB 3: 16383MB
kernel: EDAC amd64: MC: 4:     0MB 5:     0MB
kernel: EDAC amd64: MC: 6:     0MB 7:     0MB
kernel: EDAC amd64: using x8 syndromes.
kernel: EDAC amd64: MCT channel count: 2

First bits of "dmidecode -t memory" output:
Code:
Handle 0x002E, DMI type 16, 23 bytes
Physical Memory Array
        Location: System Board Or Motherboard
        Use: System Memory
        Error Correction Type: Multi-bit ECC
        Maximum Capacity: 512 GB
        Error Information Handle: 0x002D
        Number Of Devices: 8

Handle 0x0035, DMI type 17, 40 bytes
Memory Device
        Array Handle: 0x002E
        Error Information Handle: 0x0034
        Total Width: 128 bits
        Data Width: 64 bits
        Size: 8192 MB
        Form Factor: DIMM
        Set: None
        Locator: DIMM 0
        Bank Locator: P0 CHANNEL A
        Type: DDR4
        Type Detail: Synchronous Unbuffered (Unregistered)
        Speed: 2133 MT/s
        Manufacturer: Samsung
        Serial Number: 33129D7A
        Asset Tag: Not Specified
        Part Number: M391A1G43DB0-CPB  
        Rank: 2
        Configured Clock Speed: 2133 MT/s
        Minimum Voltage: 1.2 V
        Maximum Voltage: 1.2 V
        Configured Voltage: 1.2 V

The DESIGNARE EX board was chosen thanks to this thread :)
 
I forgot to update this back then, registered failed very early in the post cycle so I gave up on it. Asrock X399 (either model, they are the same board).

I was using kingston unbuffered ECC for awhile 4x16GB (KVR24E dual rank) and no problems at 2400 and ECC confirmed enabled, linux and W10. I didn't get out a heat gun to force errors though, fuck that. Used that for the first couple months while other people beta tested bios updates for me.

More recently I am using single rank non-ecc 8Gb B-die (best stuff) gskill 4x8GB runs 3200 14-14-14 fine. The retail XMP rating on it is 3600 16-16-16 but using intel profiles on amd is typically a pointless endeavor.

I just got my hands on some unbuffered ECC built with the same amd-loved B-die (also 8GB single rank) and will see where it goes. One nice thing about overclocking ECC is you get an easy early warning of stability problems :)
 
For reference, samsung B-die ECC, pay extremely close attention to the part numbers because the fucking resellers don't.

8GB single rank:
M391A1K43BB1 (CRC, CPB, etc are speed grade, but we're overclocking so whatever) actually made by samsung. However, you won't find this. Seriously, you will not find this. There are multiple ebay sellers and websites that claim to have this, but they don't. Most will confirm this when you actually try to order it, your order will get cancelled, they will offer to substitute or put you on backorder for months etc. However some will even ship you the older 4Gbit die stuff with similar part numbers (only 2 letters different) and hope you don't notice, ask me how I know. Once again, you will not find this.
F24EA8GS super talent with samsung chips - I actually got these, yep its samsung. The pcb looks like the reference design too.

16GB dual rank
M391A2K43BB1 you can find this, but not cheap
F24EB16GS super talent with samsung chips


So good news/bad news time:
I booted up at 3200 14-14-14 1.35v quad channel just fine, kinda by accident. I assumed when you swap dimms the motherboard would notice and bring up the bios, nope lol! It actually kept the tweaked xmp settings from my previous dimms and applied them. I figured what the hell and started up a nightly gaming session for a lazy stress test. I have various loggers load up on boot anyways.
No crashes, got a good 6 hours of fun in. Unfortunately WHEA cache mismatch errors, ID 18 type 9, yup damn thats ECC alright. Under load around an error every 15-30mins average, though clumped together usually. I do have 1 dimm around 5~8 degrees hotter than the rest, its in the absolute back slot right next to VRMs and gpu though.

Overclocking ECC sure is nice though, none of this mystery crashing shit or wondering if any of your stuff is getting corrupted and saved to disk. Going to run at 2400 for a couple days just to make sure its not junk and then ramp up through bins to find what the real limit is, probably test 2933 next.

I also decided to get some dual rank to try, I have a couple systems that can use ddr4 unbuffered ecc and probably going to build another ryzen and raven ridge soon too, lots of projects. (rr mining platform base might be interesting!)

PS samsung actually has officially 3200-rated (jedec yet?) 8Gbit b-dies listed as mass production from late last year. They need to put 9 of them on a dimm and sell it! http://www.samsung.com/semiconductor/dram/ddr4/
 
F24EA8GS: I beat them up good at 2400/stock a couple days, now have them running at 2933 16-16-16 1.35v quad channel 1dpc (+whatever asrock X399 P2.0 bios motherboard default sub timings for 2933, aka lazy) for two days no errors yet. Will beat them up hard this weekend, then try another stab at going higher/lower timings if they pass.
If you price the binned B-die right now getting these for ~$100 each seems pretty good.

F24EB16GS: should arrive next week, $180 each. Don't know what to expect, but dual rank b-die so hey. Either way I can rearrange my dimms between half a dozen systems that can use ddr4 unbuffered ecc and find a good home for it all.


The ram calculators and such are zero help with these, they all depend on extrapolating/tweaking known XMP settings. If you plug in the extra slow 2400 ecc jedec to start with their formulas go crazy.

Lazy shopper method for non-ecc: fast gskill kits with mostly matched timings that don't go on sale, if it is on sale or 16-18-18 bullshit its hynix crap. The resellers are paying attention to this detail, annoying. Also microcenter carries exactly 0 of them! Corsair ram, any kind, is not worth bothering with because they change chips with every revision, don't document it and rarely use samsung. I always thought they were overrated as hell for most things anyways.
 
I grabbed some F24EA8GS for the reasons Aluminum listed: single rank B-die, decent price, etc..
As expected, there were zero problems with four sticks on the initial boot of an x399 Taichi.

The sticks have no heatspreader and a basic green PCB; bonus. The actual DRAM chips on my sticks are Samsung K4A8G085WB-BCPB, binned DDR4-2133.

Note that Super Talent's specs make no guarantee that the chips will be b-die, just that they will be Samsung. Looking at Samsung's DDR4 part list, there are some C-die chips that might work, but currently there are no Samsung OEM DDR4 ECC UDIMMS using C-die.

These aren't miracle sticks, but I am able to run four of them at significantly faster than advertised speeds. After some extended testing, my sticks are stable at a true latency of ~10.50ns. Of course, YMMV. For reference, the advertised timings of 2400C17 have a TL of ~14.17ns, while 3200C14 kits have a TL of 8.75ns.

My current timings are 2666-14-14-14-28-1T @ 1.2V. I used The Stilt's HQ B-die - 3200MHz "Safe" timings as a base.

DIMM info: Thaiphoon Burner
Note that the DRAM chips in the report do not match the actual chips on the sticks.
super_talent_f24ea8gs_tb.png

Extended timings @ 1.2V:
super_talent_f24ea8gs_rtc.png
 
I don't think it is possible to make a 8GB single rank dimm with C-die(?) or E-die chips. Super talent does advertise the 1Gbit x 8 per chip x 9 (ECC) which can only be B-die as far as I can tell.

Samsung does have E-die ECC, but it is dual rank on 8GB modules (http://www.samsung.com/semiconductor/dram/module/M391A1G43EB1-CPB/) 512x18x9. These also have part numbers very close to the direct from samsung b-die modules, which one vendor tried to send as a substitute without telling me. (fucking fleabay)

I bought some SR 8GB and DR 16GB, been playing around with different configurations on a pair of threadrippers (asrock x399) and some am4 boards. Raven ridge (ryzen + vega) also appears to work with ECC which is kinda neat because now it is the cheapest ECC compute gpu ever released. Tested 1 board with the 2400G so far (asus x370) no errors yet but appears to be enabled.
I have been testing them at 3200C16 and 2933, I went straight to 1.35V as every other high speed b-die kit starts there anyways.

I think thaiphoon just has a database, I'm not totally sure but I think you're not able to read actual chip ids just from SPD data. SPDs can be totally zeroed out and programmed to whatever by the OEM.
 
Interesting, those look like they came out a year after B die. No really obvious difference in the spec sheets that jumped out at me.

The 3200 rated chips don't have a public datasheet on their site, but they all appear to be b-die: K4A8G085WB-BIWE seems to be the desirable one for larger unbuffered dimms.
 
what are some quality 4GB DDR4 unbuffered ECC sticks I should be looking for? and do you have any (re)sellers you recommend? thanks

You could try Super Talent F24EA4GS for 4GB sticks of single-rank Samsung E-die.

Superbiiz is probably the best place to get these UDIMMS. They have an ebay store under the name imicros that often has better prices than their main site.

Link: https://www.ebay.com/itm/192399927517
 
As an eBay Associate, HardForum may earn from qualifying purchases.
I've bought most of mine from superbiiz as well, fair warning that both their website and ebay listings are not reliable indicators of actually having parts in stock. More than once I have waited a week or two for them to get inventory in from the supplier, even now I am waiting for a set of f24veb16gs to ship (low profile dual rank b-die 16GB).

The ram has all been good though, it all runs at least 2933 16-16-16 1.35V on various ryzen systems without errors. I'm kinda busy doing other computer things but hope to eventually push some up to 3200 and/or down to C14.
 
Slight update to this, it appears that memory.net has regular stock of the samsung dual rank 16GB (M391A2K43BB1-CRC) for $200 which is reasonable in today's insane market.

I picked up 4 samsungs coming in today, will see how they compare to the super talent, I'm hoping better odds of hitting 3200 off the bat. If they consistiently can theres really no reason for people to get gskill b-die kits (~$450+ for 2x16) for amd anymore.

Getting a little tired of superbiiz in-stock games, they don't have realtime inventory (either on their website or ebay listings) as I bought 4 of the low profile 16GB (f24veb16gs) but delayed a few times and only got 3 of the 4 so far. Also ended up cancelling an ebay order during one of those 20% discount days, otherwise I would've gotten 2x16 for $300 :(

Also for months now no one actually has the samsung brand single rank 8GB b-die dimms at normalish prices ($100) and even some of the listings on ebay are bullshiters that will send you different parts (same listing I ordered from before is there with no change). I'm convinced this part is no longer in production.

Seriously though, why can't someone just fucking order the 3200 chips direct from samsung (mass production, and 1.2V too) and put 9/18 on a goddamn reference pcb and sell it?
Or gskill could take notice and put out ecc versions for amd, but not likely - imagine if intel didn't intentionally turn lga 2066 into a neutered dumpster fire, this would probably be a thing by now. Overclocking ecc is much better than the 'run it overnight and pray' with regular stuff.
 
Aluminum You seem to be the only person on earth who has been testing f24eb16gs. I need to know whether this stuff is going to work well or not before I blow $700 on it!

I've found it on ebay for $180 which is a steal in today's market. But I'm concerned about dual rank making it hard to clock it up for Threadripper. My 1950X is screaming out for more bandwidth. I am running 2x16GB dual channel at 2133 and the cinebench score is lacking (under 3k)
 
Well, I thought better of it and went with 4 8GB sticks. I can at least console myself with what appears to be by far the speediest server-grade workstation money can buy today. I'll try to push for 3200Mhz for sure.

Hopefully in the distant future when prices have come down I could try to bring this machine to 96GB by getting some 16G sticks. With the tax it was closer to $800 for them. It's just too much. 2 years ago I got 64GB of DDR4 (non-ecc of course) for $360...

Still... with "only" 32GB, that means my workloads can only afford one gigabyte per thread. Sigh.
 
Well, I thought better of it and went with 4 8GB sticks. I can at least console myself with what appears to be by far the speediest server-grade workstation money can buy today. I'll try to push for 3200Mhz for sure.

Hopefully in the distant future when prices have come down I could try to bring this machine to 96GB by getting some 16G sticks. With the tax it was closer to $800 for them. It's just too much. 2 years ago I got 64GB of DDR4 (non-ecc of course) for $360...

Still... with "only" 32GB, that means my workloads can only afford one gigabyte per thread. Sigh.

Hi, running 4 x CT16G4WFD824A - micron 16 GB ECC UDIMM in a TR 1950X system. Does anyone have OC experience with this memory they can share? Thank you!
 
Hi, running 4 x CT16G4WFD824A - micron 16 GB ECC UDIMM in a TR 1950X system. Does anyone have OC experience with this memory they can share? Thank you!
Don't know about those, but my VLP ct16g4xfd824a sticks hit the following @ 1.27V.
stable n' nuff.png


I could go with more aggressive timings, but then that would be more voltage on a 24/7 system.
 
Last edited:
Back
Top