AM5 DDR5 6000 EXPO Adventures - Failure to warm reboot

[macintoshme]

I purchased an AM5 setup a while ago and experienced some RAM issues. 1min+ (post-initial memory training) cold boots and later I realized the system could not warm-reboot, at all. I figured some of the ram issues I had been having were teething issues of the platform as many of the reports I had heard of stability issues. I religiously just updated my firmware every few weeks or so like I had with X370 waiting for a fix. However last week I saw some posts about the new AGESA supporting much faster RAM than I was working with.

I reached out to ASUS finding it likely it was a board firmware issue. I had already done some single stick testing and found that the behavior was present on all slots with both sticks. Asus went back and forth with me, eventually challenging me to try a different ram kit to resolve my issue before they would issue an RMA. I did so, buying another set of the same supported kit. The issues continued (no surprise as the kit had passed memtest86+ many times on EXPO speeds). They started moving ahead with an RMA that I wasn't very confident would solve my problem.

Preparing to be without my hardware and for interim testing, I picked up the cheapest x670 board that had my RAM listed on the QVL.

 

Right after prepping my board for RMA, I moved the parts over to the new board (which is still on an oldish bios), but was able to reproduce the issue on the first warm boot. Same as the Creator, the MSI board would boot cold, but a reboot would just hang at the memory light.

I have now swapped:

Motherboard

RAM

PSU

I have come to the conclusion that this is a poor memory controller on my chip. I have reached out for an RMA from AMD. I am curious to know how this will go. The ability to run the Infinity fabric at the optimal speed they suggested strongly influenced my purchase. I might be able to apply additional voltage to the SoC to help the controller, but I was hoping for a better out-of-the-box experience for a highly binned and expensive part. I was able to run a previous generation at the suggested speed for (3900x) with just as much RAM, shouldn't this be stable? I would however love community opinions on this.


Original Parts:
 

Ryzen 9 7950X3D

GSkill Trident Z5 Neo 6000 CL30 (F5-6000J3040G32GX2-TZ5NR)
ProArt X670E-Creator

SuperFlower Platinum 1000w

(There are more, but they are not needed to reproduce).

Interim parts:

MSI Pro X670-P

GSkill Trident Z5 Neo 6000 CL30 (F5-6000J3040G32GX2-TZ5NR)

EVGA Platinum 750w


I wanted to document this on a better format than Reddit where I saw some similar threads. I am extremely open to any suggestions and am happy to answer any additional questions.
 

Edited August 5, 2023 by macintoshme
grammar and spelling

[RONOTHAN##]

First thing I'd be thinking is Memory Context Restore is broken with your CPU. Most boards should have that on by default, and until the latest 1.0.0.7b AGESA update it wasn't 100% reliable on a lot of setups. It's still not perfect, but for a lot of people it's fixed. 

 

15 minutes ago, macintoshme said:

The ability to run the Infinity fabric at the optimal speed they suggested strongly influenced my purchase.

Just to get this out of the way, that's not how it works on Ryzen 7000. They permanently desynchronized the FCLK on Ryzen 7000, so no matter if you're running DDR5 5200, 6400, or even DDR5 8000 on the newest AGESA, the FCLK will run at 2000MHz unless manually tuned. The reason that 6000 CL30 "sweet spot" was picked was because it was just the maximum frequency that would work on 99% of CPUs up until very recently, wasn't actually a sweet spot. Now it is more of a sweet spot because it's the highest frequency that will work in 1:1 mode reliably on 99% of CPUs. 

[macintoshme]

8 hours ago, RONOTHAN## said:

First thing I'd be thinking is Memory Context Restore is broken with your CPU. Most boards should have that on by default, and until the latest 1.0.0.7b AGESA update it wasn't 100% reliable on a lot of setups. It's still not perfect, but for a lot of people it's fixed. 

 

Just to get this out of the way, that's not how it works on Ryzen 7000. They permanently desynchronized the FCLK on Ryzen 7000, so no matter if you're running DDR5 5200, 6400, or even DDR5 8000 on the newest AGESA, the FCLK will run at 2000MHz unless manually tuned. The reason that 6000 CL30 "sweet spot" was picked was because it was just the maximum frequency that would work on 99% of CPUs up until very recently, wasn't actually a sweet spot. Now it is more of a sweet spot because it's the highest frequency that will work in 1:1 mode reliably on 99% of CPUs. 

I've toggled MCR in the past with no relief, the Creator board defaults to on. I've tried the 1.0.0.7b (1516) bios on the main board and it continues to have issues. I'll try it on the other board tomorrow.

I'll have to read a little more about it. I do assume that 2GHz /6000 was selected due to it being an even ratio? It saddens me that my chip clocked in at the very bottom of the controller lottery.

[RONOTHAN##]

11 minutes ago, macintoshme said:

I do assume that 2GHz /6000 was selected due to it being an even ratio?

Nope, 2GHz is just what was deemed what will work on all CPUs. 6000MT/s with 2200MHz FCLK will be faster than 6000MT/s with 2000MHz FCLK, 6200MT/s with 2000MHz FCLK will be faster than 6000MT/s with 2000MHz FCLK assuming both are running in 1:1 mode. It just kinda worked out that way. 

[macintoshme]

AMD accepted an RMA on my CPU and the new chip did not have the issue.