The reason of 6000 sweet spot is that, by default, the memory controller runs at 1/2 clock of the ram clock when above 6000.
If you can lock the 1/1 uclock ratio in bios and get stable at 6400. 6400 is your sweet spot for your chip.

 

I do believe that "North Bridge Clock" shows you are running 1:1. You can double-check using other utilities like ZenTimings or Ryzen Master. If you post the ZenTimings shot, that gives the full picture of your RAM config.

 

MCLK:UCLK = 1:1
So I should probably leave well enough alone - unless of course there are are some specific timings I might tweak?
This system is pretty much dedicated to MSFS, which runs entirely on the Vcache chiplet (via Process Lasso Affinity setting.)
So far it's very stable.

 

Should be able to run significantly tighter secondary and tertiary timings while still keeping 6400MT/s. That profile is running 1:1 UCLK:MCLK, so if it's stable, you know the IMC can handle it. Increasing FCLK, depending on what it's at now, and tightening timings would certainly improve memory performance. However, it's not going to be a huge difference, especially for games that limit themselves to the vcache CCX.

 

Perfect example of why neophytes like me shouldn't mess with things they know very little about:

I bumped FCLK up to 2200 and the computer wouldn't POST.

Flashed the most current BIOS and reset everything. Lesson learned.

 

I bumped FCLK up to 2200 and the computer wouldn't POST.

Essentially every Raphael part that isn't defective will do 2000 FCLK. ~90% of them will do 2100 FCLK. More than half will do 2133 FCLK. Past that things start to fall off pretty quickly; probably less than one-in-fifty will do 2200+ FCLK with any kind of stability and most won't POST.

MCLK:UCLK = 1:1
So I should probably leave well enough alone - unless of course there are are some specific timings I might tweak?
This system is pretty much dedicated to MSFS, which runs entirely on the Vcache chiplet (via Process Lasso Affinity setting.)
So far it's very stable.

View attachment 2671980

There is a lot of room for improvement here, but it will take quite a bit of effort to tune everything optimally.

Some of the easier changes would be to reduce tRAS to 64, tRC to 102, tRFC to 560, and increase tREFI to maximum. FCLK at 2133 should also be tested.

Many other timings can be reduced significantly, but will require far more testing to validate. Note that GDM being enabled makes tRCD of odd numbers pointless, so you're actually running 40 there.

Also, some of the termination resistances are excessive (RttWr doesn't need to bne 240), as are some voltages (VDDIO will almost certainly work at 1.3v), but nothing should be dangerous.

 

Essentially every Raphael part that isn't defective will do 2000 FCLK. ~90% of them will do 2100 FCLK. More than half will do 2133 FCLK. Past that things start to fall off pretty quickly; probably less than one-in-fifty will do 2200+ FCLK with any kind of stability and most won't POST.

That explains the reason for the POST failure.

There is a lot of room for improvement here, but it will take quite a bit of effort to tune everything optimally.

Some of the easier changes would be to reduce tRAS to 64, tRC to 102, tRFC to 560, and increase tREFI to maximum. FCLK at 2133 should also be tested.

Many other timings can be reduced significantly, but will require far more testing to validate. Note that GDM being enabled makes tRCD of odd numbers pointless, so you're actually running 40 there.

Also, some of the termination resistances are excessive (RttWr doesn't need to bne 240), as are some voltages (VDDIO will almost certainly work at 1.3v), but nothing should be dangerous.

I'm a tweaker at heart, so I'll try your suggestions and report back. What's the best test software to use in order to see the incremental improvements? Is Aida64 good enough?

Thanks for your help. I'll keep an eye on DIMM temps as I make these changes. My case has three side fans that duct air over the motherboard. Hopefull that will be enough. Right now both DIMMs are running around 50°C under load, so I've got got some thermal headroom hear (based on what I've read.)

 

First change was to set Infinity Fabric Frequency to 2133. Booted OK, but got a 'Memory Allocation Timeout' when running Aida64 benchmark.
I set it back to 2000. Tested again, and got the same error. I think it might be a Windows problem, not a memory problem, since the Aida64 stress test ran fine for 30 min.
So I created an SPD profile using your recommended settings and bumped FCLK up to 2100. Stress test running fine as I type this.
I'm going to bump FCLK back up to 2133 with the new SPD settings and test again.

I set tRFC to 560 as recommended. (Was 943 in XMP1)
I noticed two other similar timings, which I left the same as the XMP1 settings.

tRFC2 = 512
tRFCsb = 416

Should I tighten those up?

 

tRFC2 = 512
tRFCsb = 416

Should I tighten those up?

You can tighten them proportionally, but they are used so rarely that there is unlikely to be a meaningful performance impact from doing so.

 

Here's my result with the Custom SPD timings and FCLK=2133 compared with the baseline XMP-1 and FCLK=2000

XMP-1 SPD timings, FCLK = 2000

Custom SPD timings, FCLK = 2133. (I tried FCLK = 2200, but the computer wouldn’t POST. I had to reset the CMOS and reload the BIOS settings.



Aida64_Mem_XMP5_FCLK 2133_test3829×797 72.6 KB

• 
  
• 
  

I had to test system stability, and so far the system is 100% stable. And temps are as good as I could wish for.

The Aida64 stress test loads all CPU cores @ 100%, and loads RAM at 100% as well. This one ran for an hour. RAM temps (40 and 44°C) are very good, as is CPU temp of 70°C (thanks to the PBO negative offsets.) It shows 64°C because I stopped the test just before I grabbed the screenshot. The RAM temps were a bit higher too as the test ran, but never exceeded 50°C.



Thanks again for the great help.

 

Hey could you share your zentiming screenshot please? I want to copy your OC setting to mine to test it out. @Joe Burwocky