ASRock X670E Steel Legend Motherboard Needs Hundreds of Seconds at First Boot or Clear CMOS to Train Memory

[btarunr]

At this point, we don't know if this is a limitation at AMD's level or ASRock's, but someone with access to a retail ASRock X670E Steel Legend motherboard, with all its packaged paraphernalia in place, spotted an interesting sticker covering the board's four DDR5 DIMM slots. The sticker has some info on the ideal DIMM slot selection for dual-channel memory (4x sub-channels); but what catches our eye is a table which states just how long the motherboard will take to train the memory the first time it's booted up, or after a clear-CMOS operation (where your BIOS settings are erased).

The table says that a typical setup with two 16 GB modules (read: two single-rank modules in a 1 DIMM per channel/1DPC configuration), takes 100 seconds to train (or until first boot). Two 32 GB modules (typically a pair of dual-rank modules in 1DPC configuration) take 200 seconds, as do four 16 GB modules (four single-rank modules in a 2DPC configuration). The least optimal config, four dual-rank modules in a 2DPC configuration, takes a whopping 400 seconds (almost 7 minutes) to train. That's 100 to 400 seconds of a black screen, or no display signal, enough to unnerve anyone and assume something is DOA.



Update Sep 2nd: The source behind this story confirmed that this is an ASRock-level issue, and that it's been "fixed" with the latest BIOS.

Update Sep 8th: This has been fixed according to ASRock.



Here's the kicker—since UEFI BIOS updates typically clear CMOS, you'll have yourselves some nerve-racking hundred(s) of seconds until the display lights up, letting you know that the BIOS update went through. Interestingly, we haven't yet seen anything to suggest that memory overclock (which involves dozens of reboots and re-training of memory), takes hundreds of seconds—not unless you clear CMOS for some reason.

View at TechPowerUp Main Site | Source

 

[Bloax]

The Fury X will be an overclocker's dream, and so will AM5!

 

[Dyatlov A]

How about 2x8GB? Less memory faster system

 

[Nanochip]

If this is true generally for most or all am5 motherboards, I’ll be avoiding first gen am5 like the plague until these issues are resolved.

I used to be an early adopter but these days I’m loathe to early adopt new platforms anyways, and let the early adopters aka beta testers give feedback to the motherboard companies so that the bugs are worked out. I also don’t run beta OSes on my phones or tablets anymore either.

I just want stability and good usability.

 

[btarunr]

How about 2x8GB? Less memory faster system

I think 100 secs, since those are going to be a pair of single-rank modules in 1DPC config.

 

[ARF]

Sorry but what does "train" memory mean?

Also, what is 1 DIMM per channel? One DIMM of DDR5 has two channels, two DIMMs is effectively 4x32-bit

 

[ir_cow]

Usually after 30 seconds I consider the training failed. Its going to be painful to overclock memory on AM5 if this is the case for all MBs.

 

[phanbuey]

400 seconds? what is it doing for that long exactly

 

[btarunr]

Sorry but what does "train" memory mean?

Also, what is 1 DIMM per channel? One DIMM of DDR5 has two channels, two DIMMs is effectively 4x32-bit

A DDR5 module has 1 channel, and 2 sub-channels. Too many people think DDR5 is "dual-channel on a stick" or that Alder Lake and Raphael are "quad-channel" processors. They're not.

 

[ARF]

400 seconds? what is it doing for that long exactly

My best guess is some weird self test...

 

[mechtech]

If this is to improve performance stability and reliability by automatically letting the cpu/ram/mobo do it’s thing then that’s cool.

 

[ARF]

A DDR5 module has 1 channel, and 2 sub-channels. Too many people think DDR5 is "dual-channel on a stick" or that Alder Lake and Raphael are "quad-channel" processors. They're not.

Well, I though it has been officially stated as such. CPU-Z even changed its tab in order to facilitate the change:

Previously, it always showed Dual instead of 2 x 64-bit.

 

[btarunr]

Well, I though it has been officially stated as such. CPU-Z even changed its tab in order to facilitate the change:

View attachment 260253

Previously, it always showed Dual instead of 2 x 64-bit.

Exactly. DDR4 has 1 channel, no sub-channels, and a 64-bit wide path (72-bit with ECC). DDR5 has 1 channel and two sub-channels, with two 32-bit paths (40-bit with ECC, each). In our interactions with AMD since Rembrandt (their first CPU with DDR5), they refer to DDR5 as being 1 channel per DIMM, and two sub-channels per channel.

 

[Octavean]

And if you have 64GB,…..?

 

[Dyatlov A]

Usually after 30 seconds I consider the training failed. Its going to be painful to overclock memory on AM5 if this is the case for all MBs.

uhh, than i ll stay with Alder Lake, i hate waiting for something to happen.

 

[Nanochip]

uhh, than i ll stay with Alder Lake, i hate waiting for something to happen.

My z690 trains new memory kits in under 30 seconds. I can’t imagine waiting 400 seconds. No way.

 

[evernessince]

If this is true generally for most or all am5 motherboards, I’ll be avoiding first gen am5 like the plague until these issues are resolved.

I used to be an early adopter but these days I’m loathe to early adopt new platforms anyways, and let the early adopters aka beta testers give feedback to the motherboard companies so that the bugs are worked out. I also don’t run beta OSes on my phones or tablets anymore either.

I just want stability and good usability.

It's inconvenient but given that it only happens on the first boot it's really a non-issue.

If the longer training time is being used to better tune the memory that's a superior approach IMO. First boot being longer in exchange for superior memory performance is more than a worthwhile trade-off.

 

[ir_cow]

Also it says first boot. I assume that means every time you disconnect the power or clear cmos. Normal users are just going to have to wait a long time once. Anyone planning on overclocking memory, making any memory related changes in the BIOS could be waiting a while.

At least on Intel ADL, the more values you manually type in related the memory, the faster it boots. Could be true here as well. We don't know yet.

 

[Nanochip]

Sorry but what does "train" memory mean?

Also, what is 1 DIMM per channel? One DIMM of DDR5 has two channels, two DIMMs is effectively 4x32-bit

It's inconvenient but given that it only happens on the first boot it's really a non-issue.

If the longer training time is being used to better tune the memory that's a superior approach IMO. First boot being longer in exchange for superior memory performance is more than a worthwhile trade-off.

But when you update the bios or you need to clear CMOS, memory will have to be retrained. Also, you don’t know what other conditions (such as applying an overclock or changing a voltage setting or any other setting in the bios) might trigger memory retraining. 400s is ridiculous.

 

[AM4isGOD]

Is this with all X670E or just this board?

 

[ir_cow]

I wonder is AMD EXPO will fix this...or that is the real reason AMD is pushing this new memory profile because XMP doesn't work well with AM5.

 

[TheLostSwede]

This is a bug that should already have been fixed, at least by AMD and ASRock's competitors.

 

[Nanochip]

This is a bug that should already have been fixed, at least by AMD and ASRock's competitors.

I hope so because a production BIOS should not take 100-400 seconds to train memory. That’s unacceptable.

 

[Dyatlov A]

Why this training needed at all? I dont remenber in Pentium 3 and 4 era for any training. The system eithet started after new memory settings immediately or could not start.

 

[evernessince]

But when you update the bios or you need to clear CMOS, memory will have to be retrained. Also, you don’t know what other conditions (such as applying an overclock or changing a voltage setting or any other setting in the bios) might trigger memory retraining. 400s is ridiculous.

Only time memory training is needed is first boot, CMOS clear, and BIOS update.

In the span of 5-7 years 400 seconds is nothing, especially if it's providing a superior experience to shorter training times.

Why this training needed at all? I dont remenber in Pentium 3 and 4 era for any training. The system eithet started after new memory settings immediately or could not start.

To either set timings and test them or ensure that user set timings function. Without memory training the system would simply fail to boot in instances where the user set bad timings or an XMP / EXPO profile doesn't work as you pointed out. That is not an acceptable option as people who aren't computer whizes will think the PC is broken and aren't aware that they could remove the CMOS battery. Mind you it's inconvenient to have to remove the battery in the first place. Memory training mostly gets rid of having to do that.