memtest86+ 6.10 Released With UEFI Secure Boot Signing, Headless EFI

[piorunz]

i recall asrock being mentioned as (secretly) supporting ECC during the early days of ryzen

prime is lower tier than TUF, or maybe they're about the same, so i guess that's a good sign that it's not limited to high end boards



what timings are those 3600 sticks? do you know which dies? does overvolting also increase instability? there are probably a lot more moving parts still like the SoC voltage, board topology, etc

i always assumed ECC meant JEDEC (either that or i'm unaware that JEDEC went over 3200, but i dont care about 1.2v, give me 1.35v at tighter timings)

(as an aside i've gone way deep into tweaking timings+volts+ohms trying to make 3200c16 hynix-M stable with 2600x, not that there were known memory errors or app crashes, but the entire system would reset, eventually figured out a combo of settings that was almost always stable, with occasional amdgpu pcie errors in the journal and a reset once a month or two... recently switched to 3600c16 micron with 5600 PBO same b450 mobo, no more journal errors, stable ever since, dont feel like adding the old ram to test if the ram was fine while the old cpu wasnt... oh well, learned a lot and always prefer messing with hardware that regular people can access/afford or that isnt constantly reviewed like samsung)

My sticks are 3200 MT/s JEDEC. I overclocked them.

This is what they have for sale in my country currently:
https://www.kingston.com/unitedkingd...ffered-dimm

My BOM number is 9965745-020.A00G which seem to correspond to (not for sale anymore):
Part Number: KSM32ED8/32ME
Product Description: 32GB DDR4-3200 ECC Unbuffered DIMM CL22 2Rx8 1.2V Micron E (Server Premier)
JEDEC Specification: 32GB 2Rx8 PC4-3200AA-EE1-12

With experience from my homeserver on PRIME B350-PLUS, which supported 2666 MT/s ECCs beautifully, I went ahead and upgraded my workstation PC to AMD Ryzen 7 5800X + PRIME B550-PLUS, with that ECC RAM added later. Never had any issues whatsoever, even after overclocking from 3200 to 3600 MT/s. Everything just works. If you want peace of mind, do something like I did.
Even on the first gen Ryzen 1600X I bought, I had no RAM issues, because my first RAM was Ryzen-certified G.Skill 3200 MT/s (normal non-ECC though). I had that CPU until 5800X with ECC upgrade. Happy days.

[kn00tcn]

i recall asrock being mentioned as (secretly) supporting ECC during the early days of ryzen

prime is lower tier than TUF, or maybe they're about the same, so i guess that's a good sign that it's not limited to high end boards

[*]Kingston Server Premier unbuffered ECC DDR4 overclocked from 3600 MT/s to 3800 MT/s
Testing ECC correction is possible when you deliberately make the memory unstable, by changing voltage or other settings.

what timings are those 3600 sticks? do you know which dies? does overvolting also increase instability? there are probably a lot more moving parts still like the SoC voltage, board topology, etc

i always assumed ECC meant JEDEC (either that or i'm unaware that JEDEC went over 3200, but i dont care about 1.2v, give me 1.35v at tighter timings)

(as an aside i've gone way deep into tweaking timings+volts+ohms trying to make 3200c16 hynix-M stable with 2600x, not that there were known memory errors or app crashes, but the entire system would reset, eventually figured out a combo of settings that was almost always stable, with occasional amdgpu pcie errors in the journal and a reset once a month or two... recently switched to 3600c16 micron with 5600 PBO same b450 mobo, no more journal errors, stable ever since, dont feel like adding the old ram to test if the ram was fine while the old cpu wasnt... oh well, learned a lot and always prefer messing with hardware that regular people can access/afford or that isnt constantly reviewed like samsung)

[piorunz]

Ah yeah, interesting idea for performing such tests. Thanks for mentioning it, I had forgotten about that unofficial UDIMM ECC support on many Zen models
Using ECC UDIMMs with Ryzen PRO, or non-APU Ryzen, on a highly configurable consumer MB is indeed more affordable than HEDT / server equipment. I do have a Ryzen 7 PRO 5750GE on a B550-based MB advertised to support ECC UDIMMs.
Did you benchmark the latency of your ECC memory ? On the one hand, the CL22 figure for DDR4-3200 ECC initially struck me as pretty high, but on the other hand, it's not worse than some cheap DDR4-3200 (JEDEC 3200) non-ECC sticks. Some non-ECC sticks are sold as CL14 DDR-3200 (XMP) or even CL14 DDR4-3600 (XMP).

Yes, it's very interesting option we have nowadays! Reliable computing without buying Xeon/Epyc and spending millions. Support for ECC UDIMM is official in Ryzen processors, but unofficial in third party motherboards, because they didn't care to certify this feature in consumer boards. But some vendors did make sure it's working in some of their boards.

I did some benchmark at the time when I was upgrading the sticks, but changes were insignificant to remember. I replaced existing 16GB I had at a time (3200 MT/s and 14CL I think it was), with new ECC 64GB 3600 MT/s but CL22, ended up overclocking that ECC sticks to 3600 MT/s, because that's Ryzen processor sweet spot. To sum it up, I may be losing some frames in games or seconds in my data crunching (for work) due to increased latency, but reliability of data and results was crucial, and it has been achieved.

Disclaimer: I lost tons of hours fixing RAM issues in the past, due to using non-ECC sticks and non-redundant, non-checksum filesystems, I learned my sh*t the hard way Now, ECC only.

[debrouxl]

I am talking from experience with not with very expensive Supermicro boards and Xeon processors, but cheap and accessible unbuffered ECC RAM on AMD Ryzen platform. All you need to do is buy Ryzen processor (all are ECC supported) and ASUS motherboard (with ECC support). Other motherboard vendors may not care about ECC support, or even disable it, but ASUS is making sure it's working. I build home PCs only with ECC RAM now. I have now currently:
[...]
Overclocking, changing voltage is possible on these boards.
Testing ECC correction is possible when you deliberately make the memory unstable, by changing voltage or other settings.

Ah yeah, interesting idea for performing such tests. Thanks for mentioning it, I had forgotten about that unofficial UDIMM ECC support on many Zen models
Using ECC UDIMMs with Ryzen PRO, or non-APU Ryzen, on a highly configurable consumer MB is indeed more affordable than HEDT / server equipment. I do have a Ryzen 7 PRO 5750GE on a B550-based MB advertised to support ECC UDIMMs.
Did you benchmark the latency of your ECC memory ? On the one hand, the CL22 figure for DDR4-3200 ECC initially struck me as pretty high, but on the other hand, it's not worse than some cheap DDR4-3200 (JEDEC 3200) non-ECC sticks. Some non-ECC sticks are sold as CL14 DDR-3200 (XMP) or even CL14 DDR4-3600 (XMP).

[enlar]

I am talking from experience with not with very expensive Supermicro boards and Xeon processors, but cheap and accessible unbuffered ECC RAM on AMD Ryzen platform. All you need to do is buy Ryzen processor (all are ECC supported) and ASUS motherboard (with ECC support). Other motherboard vendors may not care about ECC support, or even disable it, but ASUS is making sure it's working. I build home PCs only with ECC RAM now. I have now currently:

Just a note, All Ryzen processors EXCEPT those with integrated graphics like 2200G...5700G.

[piorunz]

You're welcome.

I don't have access to recent high-end professional workstations equipped with ECC memory. However, in the UI in the (UEFI) BIOS of the Dell & HP server models I have access to, I'm not used to see menus which would make it possible to fiddle with the voltage of the registered ECC memory, whereas I routinely see them on consumer MBs equipped with non-ECC UDIMMs, even some old models.

I am talking from experience with not with very expensive Supermicro boards and Xeon processors, but cheap and accessible unbuffered ECC RAM on AMD Ryzen platform. All you need to do is buy Ryzen processor (all are ECC supported) and ASUS motherboard (with ECC support). Other motherboard vendors may not care about ECC support, or even disable it, but ASUS is making sure it's working. I build home PCs only with ECC RAM now. I have now currently:

• ASUS PRIME B550-PLUS
• CPU: AMD Ryzen 7 5800X (16) @ 3.800GHz
• Kingston Server Premier unbuffered ECC DDR4 overclocked from 3200 MT/s to 3600 MT/s

and older:

• ASUS PRIME B350-PLUS
• CPU: AMD Ryzen 7 1700 (16) @ 3.000GHz
• Kingston Server Premier unbuffered ECC DDR4 2666 MT/s

$ sudo inxi --memory --memory-modules --machine --cpu
Machine: Type: Desktop Mobo: ASUSTeK model: PRIME B550-PLUS
Memory: RAM:total: 62.71 GiB used: 15.34 GiB (24.5%)
Array-1:capacity: 128 GiB slots: 4 EC: Multi-bit ECC
Device-1: DIMM_A2 type: DDR4 size: 32 GiB speed: 3600 MT/s
Device-2: DIMM_B2 type: DDR4 size: 32 GiB speed: 3600 MT/s
CPU: Info: 8-core model: AMD Ryzen 7 5800X bits: 64 type: MT MCP cache:L2: 4 MiB

Machine: Type: Desktop Mobo: ASUSTeK model: PRIME B350-PLUS
Memory: RAM:total: 62.78 GiB used: 14.8 GiB (23.6%)
Array-1:capacity: 128 GiB slots: 4 EC: Multi-bit ECC
Device-1: DIMM_A1 size: 16 GiB speed: 2666 MT/s
Device-2: DIMM_A2 size: 16 GiB speed: 2666 MT/s
Device-3: DIMM_B1 size: 16 GiB speed: 2666 MT/s
Device-4: DIMM_B2 size: 16 GiB speed: 2666 MT/s
CPU: Info: 8-Core model: AMD Ryzen 7 1700 bits: 64 type: MT MCP L2 cache: 4 MiB

$ sudo dmesg | grep EDAC
[ 0.461229] EDAC MC: Ver: 3.0.0
[ 7.169338] EDAC amd64: MCT channel count: 2
[ 7.169395] EDAC MC0: Giving out device to module amd64_edac controller F19h_M20h: DEV 0000:00:18.3 (INTERRUPT)
[ 7.169397] EDAC amd64: F19h_M20h detected (node 0).
[ 7.169400] EDAC MC: UMC0 chip selects:
[ 7.169401] EDAC amd64: MC: 0: 0MB 1: 0MB
[ 7.169402] EDAC amd64: MC: 2: 16384MB 3: 16384MB
[ 7.169405] EDAC MC: UMC1 chip selects:
[ 7.169405] EDAC amd64: MC: 0: 0MB 1: 0MB
[ 7.169406] EDAC amd64: MC: 2: 16384MB 3: 16384MB
[ 7.169407] EDAC amd64: using x16 syndromes.
[ 7.169414] EDAC PCI0: Giving out device to module amd64_edac controller EDAC PCI controller: DEV 0000:00:18.0 (POLLED)
[ 7.169415] AMD64 EDAC driver v3.5.0

Overclocking, changing voltage is possible on these boards.
Testing ECC correction is possible when you deliberately make the memory unstable, by changing voltage or other settings.

[debrouxl]

Thanks for your valuable reply.

Just a note for a developer and anyone interested - any stick can be made faulty for debugging purposes, all you need to do is to lower the voltage until it fails in memtest, while not preventing booting.

You're welcome.

I don't have access to recent high-end professional workstations equipped with ECC memory. However, in the UI in the (UEFI) BIOS of the Dell & HP server models I have access to, I'm not used to see menus which would make it possible to fiddle with the voltage of the registered ECC memory, whereas I routinely see them on consumer MBs equipped with non-ECC UDIMMs, even some old models.

[piorunz]

Nope, ECC corrected error reporting and actual live RAM settings are features Passmark's memtest86 supports and memtest86+ does currently not (*). There's got to be an advantage to commercial software which (well, hopefully) enables its producers to make a living - though native support for x86_64, UEFI, Secure Boot capability, USB 1/2/3 controllers, SPD decoding for DDR4+ (on most platforms, including EPP / XMP profiles, code checked against a sizable portion of the CPUID RAM database), or live CPU temperature reporting for recent models, among major user-visible features, are no longer exclusive to Passmark's proprietary memtest86

Every model of memory controller requires specific code paths, be it for its EDAC or its IMC part, and testing for this capability requires having access to ECC RAM sticks which are faulty enough to cause errors, but not too faulty to prevent the computer from considering the stick altogether... Such sticks seem to be hard to find, unlike faulty non-ECC RAM sticks.

*: the tracking discussion for ECC corrected error reporting in memtest86+ is https://github.com/memtest86plus/mem...discussions/92 ; Sam posted in https://github.com/memtest86plus/mem...iscussions/248 that he intends to work on a single model first, if he can get a test stick. The tracking discussion for live RAM settings is https://github.com/memtest86plus/mem...iscussions/165 .
Users can help by voting for their favorite models and donating suitably faulty RAM sticks

Thanks for your valuable reply.

Just a note for a developer and anyone interested - any stick can be made faulty for debugging purposes, all you need to do is to lower the voltage until it fails in memtest, while not preventing booting.

[debrouxl]

Nope, ECC corrected error reporting and actual live RAM settings are features Passmark's memtest86 supports and memtest86+ does currently not (*). There's got to be an advantage to commercial software which (well, hopefully) enables its producers to make a living - though native support for x86_64, UEFI, Secure Boot capability, USB 1/2/3 controllers, SPD decoding for DDR4+ (on most platforms, including EPP / XMP profiles, code checked against a sizable portion of the CPUID RAM database), or live CPU temperature reporting for recent models, among major user-visible features, are no longer exclusive to Passmark's proprietary memtest86

Every model of memory controller requires specific code paths, be it for its EDAC or its IMC part, and testing for this capability requires having access to ECC RAM sticks which are faulty enough to cause errors, but not too faulty to prevent the computer from considering the stick altogether... Such sticks seem to be hard to find, unlike faulty non-ECC RAM sticks.

*: the tracking discussion for ECC corrected error reporting in memtest86+ is https://github.com/memtest86plus/mem...discussions/92 ; Sam posted in https://github.com/memtest86plus/mem...iscussions/248 that he intends to work on a single model first, if he can get a test stick. The tracking discussion for live RAM settings is https://github.com/memtest86plus/mem...iscussions/165 .
Users can help by voting for their favorite models and donating suitably faulty RAM sticks

[piorunz]

Great news! Using it now, v6.10. Works beautifully.
I wonder if it will support detection of ECC corrections?