Discussion DDR5: Experiences with enabling Memory Context Restore

[mikeymikec]

I thought I'd start this thread as I don't believe a dedicated one exists, I'm interested (and I think others would be too) in what kind of results people are getting with enabling this setting.

Board: Asus TUF Gaming B650-PLUS, BIOS 1811
CPU: Ryzen 7800X3D
RAM: Kingston Fury Renegade 6000MHz 32GB dual-channel
Enabling Memory Context Restore result: Fail

I've been running this system for a month (maybe a bit more), more or less default BIOS settings (no OC for example, DOCP enabled), 100% stability. I believe I ran memtest86 7.4 overnight when I originally built it (I definitely ran it for at least 4 passes at any rate). As I dual-boot Win11 and Linux Mint 21.x, being able to reboot quicker is definitely handy (it would get rid of approx. 25 second delay in POST) so as I felt that I had reached a decent milestone stability-wise, I enabled this setting. Win1 BSOD'd within a minute, IRQL_NOT_LESS_OR_EQUAL, failed to do a memory dump and the busy cursor was still on the screen.

I've set it back to auto (the default).

 

[William Gaatjes]

I did a short glance instead of reading :

DDR5 Memory Specification Released: Setting the Stage for DDR5-6400 And Beyond

www.anandtech.com

DDR5 Memory Standard: An introduction to the next generation of DRAM module technology - Kingston Technology

DDR5 is the 5th generation of Double Data Rate Synchronous Dynamic Random Access Memory, aka DDR5 SDRAM, which is available in Q4 2021. It brings better performance, stability and efficiency compared DDR4.

www.kingston.com

And if i am not mistaken, this "Memory context restore" is about reusing previously required initialisation values. And that seems to me not a very good idea.

Without knowing enough of DDR5 or DDR4 in detail, my experience with high speed electronics is that when looking at this issue, you always want the system to train the memory at power up.
Subjects like temperature differences at startup, aging of the silicon, voltage differences from the supply at startup, transmission line impedance matching, clock timing, PLL calibration for the clock, page refresh counters : Are some of the aspects that needs to be taken into account for a high speed system.
It is about training and calibration to make sure the data is reliable , for example signal integrity.

I found this interesting post at overclock.net :

What is *Memory training*

As the title suggests I want to know what memory training is, how it is done, why is it done, and anything else on the subject. Thanks in advance,

www.overclock.net

And found these links in the post to be very interesting :

https://www.micron.com/-/media/client/global/documents/products/technical-note/dram/tn46_08.pdf

DDR4 SDRAM - Initialization, Training and Calibration - systemverilog.io

A detailed tutorial on DDR4 SDRAM Initialization, Training and Calibration. Exploring topics such as Read/Write Training, ZQ Calibration, Vref Training, Read Centering, Write Centering, Write Leveling and Periodic Calibration.

www.systemverilog.io

Long story short :
Just let the system do the intialisation properly at every powerup. The computer is supposed to be reliable.

As a sidenote :
For more information about routing high speed pcb copper traces :

Types of Transmission Lines in High-Speed PCB Layout

To ensure good signal integrity, circuit board designers must fully understand the principles and applications of different types of transmission lines.

resources.pcb.cadence.com

 

[CP5670]

I don't use MCR but have read that it works if you also enable Power Down at the same time. Otherwise you get crashes. I only reboot occasionally for driver/windows updates so it's not worth the hassle.

 

[Tech Junky]

I find it to be a bandaid for beta released boards. I don't have it enabled and can reboot in 10-15 seconds.

 

[William Gaatjes]

I find it to be a bandaid for beta released boards. I don't have it enabled and can reboot in 10-15 seconds.

Why does everybody get so frustrated about having to reboot and it takes a few extra seconds to reboot or start up ?
Do people overclock that much that the system is constantly in error because of bit errors causing undefinable behavior ?
Causing system hangups and reboots or needing to reboot ?

If a computer is not hacked or purposely given a virus by some @sshole who has physical acces or who has network acces, it can be a perfectly stable system. Windows or linux, unix derivatives or any other Operating System.

 

[Tech Junky]

Why does everybody get so frustrated about having to reboot and it takes a few extra seconds to reboot or start up ?

Don't ask me.

I think the initial problem was training DDR 5 on the first boot taking in excess of 1 minute. The more ram in a system the longer it would take.

 

[repoman0]

X670e Taichi with 64GB DDR5-6000. Manually tightened timings, default voltages with the exception of 1.35V on memory. No issues at all with context restore enabled.

A PC should be stable with uptime approaching or even exceeding a year, IMO. That means a year of aging, temp differences etc with no memory retraining. Why should I have to retrain every reboot then with uptime measured in days?

 

[William Gaatjes]

Don't ask me.

I think the initial problem was training DDR 5 on the first boot taking in excess of 1 minute. The more ram in a system the longer it would take.

Aha, that timeperiod i was not aware of. That 1 minute training kind of sounds like what was mentioned before : Beta released boards, may that be software or hardware. Either that or users mishandling the boards, the cpu and the ddr5 sticks and there is ESD damage.
ESD damage may prevent proper termination or creating an excess of input impedance.
If the signal integrity (think ringing , overshoot. undershoot) for the DQ signals that is desired cannot be reached with standard clocks as defined by the specifications, then there is a serious issue with getting the data reliable from cpu to ram and vice versa.
And that is just the data lines and adresslines, clock signal.
That could cause for the system to remain in an endless training and calibration loop.
Like : Give the command to do the calibration .
Send some data and read it back : If error , do training and calibration again. GO back to step 1. And an endless rinse and repeat.
Perhaps no timeout and no message that a certain memory module has a problem : No sufficient diagnostic messages.
That and ESD damage would cause such long training times.
Or that the software in the BIOS as fallback scenario, keeps lowering the clockspeed until training and calibration succeeds.
Also in an seemingly endless loop.


For example, see this image with some tech words that are very important to know when it comes to signal integrity.
Impedance mismatch would cause excessive ringing. And then a 0 or 1 may be seen on a glitch caused by ringing and not on the desired edge or moment in time.
Think subjects like DDR (Dual data rate), QDR(Quad data rate), timing, phase locked loop. Everything gets critical.

 

[igor_kavinski]

Could just be an ASUS issue where they are ignoring their cheaper mobos:

https://www.reddit.com/r/ASUS/comments/11prlo4

 

[Hotrod2go]

MCR was a new thing for me with DDR5 & apparently if you want to avoid BSOD when trying to get into Windows, power down mode has to be enabled at the same time if MCR is enabled. Both of them work in tandem it seems with this type of memory architecture. Disabling one & leaving the other on can cause all kinds of stability issues so I just leave them on auto & go about my business OC the ram & experimenting with different agesa versions. No matter what version of agesa I"ve used, the MCR & PD thing is the same behaviour.
My advice is just leave them on auto unless your into competitive benchmarking & looking for that extra 1 -2 ns in latency reduction to order to win contests is vital to you. In day to day usage of the system you won't notice the effects on latency if you can boot the system successfully & use it for your apps or programs anyway.

 

[mikeymikec]

@Hotrod2go

What is "Windows power down mode"?

 

[igor_kavinski]

What is "Windows power down mode"?

It might be called Memory power down mode: https://www.overclock.net/threads/memory-power-down-mode-must-be-on.1802656/post-29115490

 

[Hotrod2go]

@Hotrod2go

What is "Windows power down mode"?

No, I said "... windows, power down mode..." - see the comer there, they are not related. But igor_kavinski in his post above cleared it up for you anyway.

 

[mikeymikec]

No, I said "... windows, power down mode..." - see the comer there, they are not related. But igor_kavinski in his post above cleared it up for you anyway.

My mistake, I realised it after igor's post. In hindsight I should have gone back and done a 'nevermind' edit.

 

[spacejamz]

Don't ask me.

I think the initial problem was training DDR 5 on the first boot taking in excess of 1 minute. The more ram in a system the longer it would take.

I returned some laptop DDR5 because I didn't know this...I thought the ram was defective because the system didn't boot immediately after installing it. When the same thing happened with the new RAM, I figured something was up and found out that the new RAM was being 'trained'.

 

[igor_kavinski]

When the same thing happened with the new RAM, I figured something was up and found out that the new RAM was being 'trained'.

Once trained, does the laptop now boot up quick like a DDR4 laptop?

 

[spacejamz]

Once trained, does the laptop now boot up quick like a DDR4 laptop?

Yes...it is only the first time booting with the new DDR5 that takes a few minutes...I installed 64GB on Dell G16-5530 and it took around 2-3 minutes for this process to complete.

Now the laptop boots very quickly (I also put a Samsung 2TB 970 SSD for my boot drive)...

This is on the Crucial website about training DDR5 memory:

DDR5 systems may require training

A small number of DDR5 systems and motherboards require a period of "training" newly installed memory to work with the system and allow access to the UEFI, or otherwise to complete the system's power-on self-test (POST).

www.crucial.com

 

[historicalaverage]

I thought I'd start this thread as I don't believe a dedicated one exists, I'm interested (and I think others would be too) in what kind of results people are getting with enabling this setting.

I've set it back to auto (the default).

If you’re like me with an MSI motherboard and “intel-optimized” teamgroup RAM googling around for how to deal with overwhelming POST times, I wanted to share with you the process I went through:

enable XMP >> 129 seconds to post but RAM runs at advertised speed (6000mhz in my case)
I have a Ryzen chipset so disable XMP enable EXPO >> 131 seconds to post
With XMP off, set frequency to 6000 manuallly >> 126 seconds to post
✔️Enable memory context restore with Power Down >> Win! 14 seconds to post and at full speed.
Simply don’t enable memory context restore >> nope haha I did
Learn how to overclock/add some volts for stability >> not today 😌

 

[mikeymikec]

I built a PC for a customer with almost the same board as mine (theirs is ASUS TUF GAMING B650M-PLUS WIFI II), 32GB DDR5 5600, the training time was a lot less (maybe 10 seconds), so I think that while the RAM capacity is relevant, I think the timing that is being attempted is more pertinent.

My system (specs in OP) without DOCP enabled POSTS more or less instantly like any pre-DDR5 system. My system with DOCP enabled takes a good 25-30 seconds to POST. Both PCs on the latest BIOS.

Another system I built with the same board as mine which wasn't really performance oriented at all only had 8GB DDR5 and more or less posted instantly.

 

[igor_kavinski]

I built a PC for a customer with almost the same board as mine (theirs is ASUS TUF GAMING B650M-PLUS WIFI II), 32GB DDR5 5600, the training time was a lot less (maybe 10 seconds)

It could also be because the mobo may have had an updated BIOS from the factory so it is better able to deal with DDR5 training in lesser time.

When I originally built my Z790/12700K PC, I tried memory overclocking and it was pretty frustrating, with long training times and occasionally locking up and requiring CMOS reset. I tried a year later with updated BIOS and this time didn't have to mess with the CMOS jumper. The training time also seems to have reduced by about 20 seconds at least.

 

[mikeymikec]

It could also be because the mobo may have had an updated BIOS from the factory so it is better able to deal with DDR5 training in lesser time.

Same BIOS version.

When I originally built my Z790/12700K PC, I tried memory overclocking and it was pretty frustrating, with long training times and occasionally locking up and requiring CMOS reset. I tried a year later with updated BIOS and this time didn't have to mess with the CMOS jumper. The training time also seems to have reduced by about 20 seconds at least.

I tried enabling DOCP again on my board after the BIOS update, still had similarly long training times.

 

[igor_kavinski]

I tried enabling DOCP again on my board after the BIOS update, still had similarly long training times.

DOCP is probably helpful only if you are trying to tweak the timings. I run my 12700K at DDR5-7000 CL34 (with other timings set to Auto) with a kit rated for 7600 CL34 and XMP off. The 7000 MT/s wall seems to be a mobo limit for me. I think I have to wait like 10 secs for the mobo logo to appear and then 5 secs more to get to Windows desktop.

 

[mikeymikec]

@igor_kavinski
I assumed DOCP was the quick way to achieve the timings that the modules are advertised to do (DDR5-6000 with better latency than the average part at that frequency).

 

[igor_kavinski]

I think the timing that is being attempted is more pertinent.

You are probably right. Anytime I try to set the timing lower than CL34, it will take about a minute before the message appears that PC failed to boot and then I go into BIOS with RAM set to DDR5-3600.

 

[igor_kavinski]

@igor_kavinski
I assumed DOCP was the quick way to achieve the timings that the modules are advertised to do (DDR5-6000 with better latency than the average part at that frequency).

AMD EXPO vs DOCP

AMD EXPO is the newest memory overclocking technology but how does it differ to the likes of XMP and DOCP?

www.corsair.com

DOCP uses the XMP protocol information to automatically set the data rate and timings on AMD motherboards and allows you to run your DDR4 RAM at up to 3200MHz.

I suppose ASUS developed this to help XMP only kits work better with AMD mobos. If your kit has dual profiles (both EXPO and XMP), just use the EXPO one for the rated speed. Anything lower, you would need to try manual tweaking with maybe 0.05V higher DIMM voltage (so if it's 1.4V, try 1.45V for stability. not sure how much extra volts for the memory controller on AMD). Reducing tREFI timing to something like 6000 also seems to help.

It doesn't hurt to try. Worst thing would be, you have to pull out the PSU power cable from the PSU backside for a few seconds and reset the CMOS jumper.