Question 6400MT/s RAM too fast for AMD 7xxx?

[lakedude]

With all this talk of CPUs having trouble keeping up and getting damaged from both Intel and AMD I'm starting to worry about my kid's rig.

I bought 6400MT/s RAM* and enabled the XMP profile. The system has been rock solid but I'm seeing reports of 6000 being the sweet spot and of AMD CPUs burning up.

I don't really want to revert to JEDEC.

Shoud I manually set the speed to say 6000 and enable some voltage or power limits?

* G.SKILL Trident Z5 RGB Series (Intel XMP 3.0) DDR5 RAM 32GB (2x16GB) 6400MT/s CL32-39-39-102 1.40V Desktop Computer Memory

 

[Tech Junky]

I've always just let the system manage the speed. My current AMD setup has ram that has 3 speed profiles it will auto select from. I think with the speed and capacity these days it doesn't make a huge difference as it did back in the day say 20 years ago.

 

[Markfw]

With all this talk of CPUs having trouble keeping up and getting damaged from both Intel and AMD I'm starting to worry about my kid's rig.

I bought 6400MT/s RAM* and enabled the XMP profile. The system has been rock solid but I'm seeing reports of 6000 being the sweet spot and of AMD CPUs burning up.

I don't really want to revert to JEDEC.

Shoud I manually set the speed to say 6000 and enable some voltage or power limits?

* G.SKILL Trident Z5 RGB Series (Intel XMP 3.0) DDR5 RAM 32GB (2x16GB) 6400MT/s CL32-39-39-102 1.40V Desktop Computer Memory

I have some of that. I don't think it set it to 6400 before I gave up, but I could be wrong. But everything else is at 6000 or so.

 

[cusideabelincoln]

When enabling DDR speeds beyond 6000 motherboards (at least my ASrock board) will put the memory controller in 2:1 mode. You have to change it manually to get the memory controller running in 1:1 with your DRAM sticks. This probably explains why your system has been rock stable. It might also be using a higher than necessary SOC voltage with your XMP kit, which is the thing to worry about concerning longevity. So manually setting SOC voltage would be a bit safer, and setting 6000 with 1:1 mode would give you more performance than 6400 in 2:1 mode.

When I enabled 6400 MTs for the DRAM with XMP timings and setting 1:1 mode, Windows immediately corrupted itself upon first bootup. 6200 worked fine at the same SOC voltage though. But I'm just going to leave it at 6000 with some manual tuned timings and will push it further come winter time.

 

[gdansk]

Windows immediately corrupted itself upon first bootup

It doesn't really matter and I'm nitpicking but "Windows corrupted itself" sounds wrong to me. Memory that wasn't functioning properly caused Windows to become corrupt.

 

[Tup3x]

I've always just let the system manage the speed. My current AMD setup has ram that has 3 speed profiles it will auto select from. I think with the speed and capacity these days it doesn't make a huge difference as it did back in the day say 20 years ago.

What? That might somewhat work if you only buy JEDEC spec sticks. Most with XMP profiles will default to dog slow speeds if you don't use the profile. Mine would run at 2133 MT/s instead of 3600 MT/s.

 

[Tech Junky]

XMP

That's Intel for you. They aim for the safest default unless you enable the option. EXPO is the AMD version. And I'm on DDR 5 which behaves a bit different.

 

[Tup3x]

That's Intel for you. They aim for the safest default unless you enable the option. EXPO is the AMD version. And I'm on DDR 5 which behaves a bit different.

XMP doesn't really have anything to do with it. One example:

https://www.kingston.com/datasheets/KF560C30BBEA-8.pdf

It will run at 4800 MT/s by default. In XMP/EXPO kits JEDEC profile is always very conservative. I don't see how that kit for example wouldn't run at any other speed than 4800 MT/s when XMP/EXPO is not in use (well, it could maybe run even slower...).

 

[igor_kavinski]

It will run at 4800 MT/s by default.

May also depend on mobo. I've tried two different kits (one G.Skill and one Teamgroup) and both seem to fallback to DDR5-3600 on my ASROCK Z790 Sonic.

 

[IEC]

No reason why you can't run the 6400 CL32 at 6000 CL30 - that's the same 10ns first word latency at those clocks/CL. 6000 CL30 is easily attainable on AM5 and you don't need elevated voltages. Even 1.2V vSOC is overkill at those speeds.

6400 CL32 is typically Hynix dies so depending on your motherboard they may have memory presets for Hynix (e.g. 6000 CL30, 6200 CL30, 6400 CL32 etc). But given past issues with ASUS on AMD with exploding X3D fiasco (admittedly long since fixed), and Intel in general with degrading Raptor Lake CPUs (allegedly fixed with microcode 0x129), I would avoid blindly applying mobo presets without manual review as their voltages and current limits are less than ideal at best and downright dangerous at worst.

 

[Tech Junky]

Another key issue is using 4 slots vs 2.

 

[lakedude]

1:1 mode would give you more performance than 6400 in 2:1 mode

Oh my, thank you!

Was I effectively running 3200?

I put it back to JEDEC which was 4800 and got better benchmark score.

Then I tried 6000 manually choosing 1:1 and the benchmark went down??

Put it back to JEDEC for now.

 

[lakedude]

6400 CL32 is typically Hynix

Yep, Hynix.

 

[cusideabelincoln]

Oh my, thank you!

Was I effectively running 3200?

I put it back to JEDEC which was 4800 and got better benchmark score.

Then I tried 6000 manually choosing 1:1 and the benchmark went down??

Put it back to JEDEC for now.

The memory sticks would be running at 6400 MT/s (3200 MHz real clock) and the memory controller would be running at 1600 MHz in the 2:1 mode - remember the real clock speeds are half of what DDR advertises. Depending on how your BIOS handles EXPO you could have had half the memory bandwidth.

The benchmark score depends on the benchmark. Cinebench is not good for memory testing. Check out OCCT as they have some quick memory latency and bandwidth tests and it's a free download. Just run it multiple times and make sure nothing in Windows is happening in the background - like Steam updating a game.

 

[igor_kavinski]

Check out OCCT as they have some quick memory latency and bandwidth tests and it's a free download. Just run it multiple times and make sure nothing in Windows is happening in the background - like Steam updating a game.

+1

On my 128 thread 8-channel Epyc, I got worryingly bad bandwidth with MaxxMem2. Tried OCCT and it showed the correct expected bandwidth figure. It was a serious relief because seeing the MaxxMem2 reported figure gave me such a sinking feeling after sinking over $1000 into the server combo.