“Update bios”, they said. “It will improve your memory situation”, they said.
Well I did and what I got is even worse. In BIOS 7E12v1J, or even sooner because I wasn’t updating for a while, they added a mandatory setting that specifies how long a memory training is valid for. Minimum 1 day, default 30 days, maximum 255 days. I don’t need to ever retrain because I maxed out the RAM from day 1.
This is my last x86 PC to be honest. I can’t tolerate this bullshit anymore and I don’t know how anybody can.
Edit: yes, this is a BIOS of the MSI Tomahawk X670 Wifi. Be on your guard with your motherboards. Those changes are mentioned absolutely nowhere in the BIOS changelogs, so it will take extra time to investigate when those muppets really added this change.
Oooff, rip dude… It’s one aspect I’m not looking forward to, as stories say it’s not always transparent if training, or failing to boot properly…I switch my computer off each night, but if training is very common… Hmm may rethink that strategy… Although, 6 months does not seem terrible; I would just not realise when is training, or when I broke something…
To be the Devil’s advocate, the trainable parameters do drift over time. What’s stable at one point with MCR might no longer be after a while, which would at worst begin to cause silent corruptions. This could also produce support calls, and probably greater grievance.
I turned off MCR after such instabilities produced crashes. Now it always take 5 minutes to boot, but it is rock-stable after doing so.
This new option does not look like a good solution, to a problem that was in all likelihood without one, short of a complete overhaul of the entire memory controller and how trainings are done, I think.
In case anyone else is curious, 7E12v1J looks to be X670E Tomahawk from early July. I haven’t seen similar changes for AM5 boards I’ve built but their most recent releases happen to be June.
Be interesting to see if this is just an MSI thing or if it’s associated with AGESA 1.2.0.3f.
Much the same. With the work machines being Windows a 30 day retraining basically means retrain on patch Tuesday restarts. Which usually happens on its own outside of business hours, so retraining would go unnoticed.
Also the 1DPC training times I get are like 15 seconds, so monthly retaining’s probably not a major impact to most builds.
You guessed the board 100% correct. When I install 1 or 2 DIMM modules it does boot as fast as you claim, but when you max the thing out it takes anywhere up to 20 minutes to train the memory.
I can’t afford to have less than maximum RAM, because more than that I can’t afford to run my SSDs into the ground. On Gentoo, your PORTDIR (default /var/tmp) should be mounted in a ramdisk not only to spare the SSD the extra wear from the build process of your packages, but also because some packages are really bloated so building in RAM speeds things up a lot.
Interesting. With 2DPC on ASRock I usually see a couple minutes, maybe like five if I’m working on an overclock that fails to train and the board drops down to 3600 and retrains.
So far as I know it’s pretty much a desktop thing. All the TR5 boards I can think of are 1SPC and the number of people trying overclock 2DPC SP5 boards is probably close to zero.
For as much of a 2DPC overclock nerd as I am, I haven’t come across a single attempt on Intel. And I’ve googled around on multiple occasions as well as checking in on Intel overclock threads from time to time. Seems the state of DDR5 2DPC gets perceived as an AMD issue because next to nobody thinks LGA1700 or 1851 is worth bothering with.
It is simple. I bought 128GB RAM and I am extremely unlikely to change this RAM configuration until the next socket comes out. Therefore I will never need retraining.
Edit: Or instead of moving to a new socket I’ll take my money to ARM or RISC-V.
Mmm, probably be good to take a look at why the DQS interval oscillator exists.
Also, if memory context restore’s documented outside of AMD’s paywall I’ve been unsuccessful. But I rather suspect it’s retrain using previous results rather than reexploring the training space.
Either way, it’s unclear to me what MSI expects monthly reexploration to accomplish.
Unless changing a memory-related setting, timing, or voltage (including SOC), my ASUS Pro WS WRX90E-SAGE will boot from cold into Windows in less than 30 seconds (just did a reboot to confirm this).
If a memory-related setting, timing, or voltage (including SOC/IMC), are modified, this same board may take up to 3 minutes on first boot following said change.
I’m going to let everyone in on a little secret: DO NOT ALLOW ANY OF YOUR MEMORY TIMINGS SET TO AUTO. ALWAYS HARDCODE A VALUE IN BIOS, EVEN IF IT’S THE SAME AS WHAT IS TYPICALLY SPECIFIED BY AUTO. Doing so will significantly reduce you boot times AND will ensure consistent stable timings set during training.
The only settings that may “drift” over time would be those related to refresh rates, which are temperature dependent. For those that may be incapable of closely managing environmental conditions, such as room temperature, marginal settings may cause instability due to variation in memory IC die temperatures. Performance degradation due to loss of voltage margin over time is entirely different category of concern and cannot be managed without manual intervention.
I don’t have anything helpful to add, but I never had memory training problems. Ever. Not even on first boot. I was all prepped up for a painful experience based on all the hearsay, but if I had not been looking for it, I would not have thought the first boot was any longer than normal with Asrock x870e nova wifi, 9800x3d, and 2x 32 GB 6000. Wasn’t any longer when I enabled EXPO, or the very brief dabbling I did with tightening the timings past that. Except of course for the time it failed to boot. But then even after that, no noticeable memory training.