Magic number "3800Mhz" For Ryzen 3000 Series

[Krit]

Have a question what in most of the time is the biggest bottleneck in times when there is no way to hit that magic number 3800Mhz 1:1 when overclocking Micron E-Die, Samsung B-die or Hynix CJR.

1. It's cpu memory controler
2. It's motherboard
3. Bios doesn't allow
4. It's memory inself

 

[Calmmo]

None of the above, CPU not being able to get IF stable at 1900 is the biggest hurdle, then maybe next cheap motherboards that cant handle fast ram.

 

[Krit]

None of the above, CPU not being able to get IF stable at 1900 is the biggest hurdle, then maybe next cheap motherboards that cant handle fast ram.

So it's not a cpu memory controller just a bad cpu bin right

 

[Calmmo]

Kinda, mine can't do 1900 IF for example no matter what, but i can get the ram to run without 1:1 much higher.

 

[Zach_01]

From my understanding, you cannot pin point it to a single reason. I believe it’s a combination of things. It’s not about high/low tier of CPUs, boards either. DRAM sticks capability can play a role but it’s far from being the only one.

I do believe tho that most times it’s more of a combination of DRAM sticks and board, and less the CPU’s UMC.
Board memory channel tracing and topology can make difference.

 

[TheLostSwede]

Considering that there have been some memory world record overclocking on Ryzen 3000, it's not the memory controller that's ever likely to be the issue.
My understanding is that if you get a chip now, you might even be able to get 3866 1:1 with a little bit of luck. The 5000 series is obviously expected to do 4000 1:1 based on the leaked AMD slide.
That said, AMD's memory controller has some quirks that takes a bit of getting used to, so if you're coming from Intel, it's not going to be the exactly same things to tweak to get the best results.
As you can see from my system specs, I'm running four sticks of fairly affordable RAM at 3800 1:1, which is both an overclock and tightened latencies compared to the specs these modules were sold as for Intel systems.

 

[Deleted member 193596]

it is pretty much the substrate itself and not the IMC or motherboard.

having this much traffic through the PCB of the CPU itself is more likely the limiting factor.

 

[Makaveli]

Considering that there have been some memory world record overclocking on Ryzen 3000, it's not the memory controller that's ever likely to be the issue.
My understanding is that if you get a chip now, you might even be able to get 3866 1:1 with a little bit of luck. The 5000 series is obviously expected to do 4000 1:1 based on the leaked AMD slide.
That said, AMD's memory controller has some quirks that takes a bit of getting used to, so if you're coming from Intel, it's not going to be the exactly same things to tweak to get the best results.
As you can see from my system specs, I'm running four sticks of fairly affordable RAM at 3800 1:1, which is both an overclock and tightened latencies compared to the specs these modules were sold as for Intel systems.

I'm looking forward to see the testing of the 5000 series and optimal IF settings.

Once I drop a Zen 3 chip in my system I will need to see if I can hit 4000 speed with my current ram. I was planning on going 4x8 with another set of the same sticks next year. But need to wait to see how this will play out.

Also related asus just released the recommended bios for Zen 3

Version 2802 2020/10/2320.01 MBytes
PRIME X570-PRO BIOS 2802
"1. Updated AMD AM4 AGESA V2 PI 1.1.0.0 Patch B
2. Improve system performance and stability
3. Improved DRAM compatibility"

I've updated and this bios is excellent I got a nice boost in my scores with PBO Fmax on.

Also wonder is this bios will make 3800 easier with those using Zen 2?

 

[Taraquin]

IF/fclk is the hurdle for me. 99% stable but random reboots now and then.

 

[Zach_01]

And forgot to response in this yesterday...

So it's not a cpu memory controller just a bad cpu bin right

Not exactly. As I said it’s not about high/low tier CPUs (Higher tier has better binned chiplets).
I have maybe the worst binned 3600 out there with bronze rating, no OC capability and lousy boosting. Yet I’m running 1900MHz all 3 memory subsystems (UCLK = FCLK = MEMCLK) without cranking up high voltage to SoC, IF and UMC.

I’m not aware tho how the CPU substrate can be across the binned chiplets and CPUs in general. Probably has no correlation.
Rumors allegedly claim a better substrate also for 5000 series, other than better bins.

 

[R0H1T]

It's a combination of things, stable memory timings & indeed speed depends on a lot of factors including but not limited to OS & the software you run. Which is to say that your OC may work in Windows but maybe not on Linux & yeah I know the Memtest or Prime95 argument, except that doesn't work for all possible software/hardware combinations out there not in my experience at least.

 

[Taraquin]

And forgot to response in this yesterday...

Not exactly. As I said it’s not about high/low tier CPUs (Higher tier has better binned chiplets).
I have maybe the worst binned 3600 out there with bronze rating, no OC capability and lousy boosting. Yet I’m running 1900MHz all 3 memory subsystems (UCLK = FCLK = MEMCLK) without cranking up high voltage to SoC, IF and UMC.

I’m not aware tho how the CPU substrate can be across the binned chiplets and CPUs in general. Probably has no correlation.
Rumors allegedly claim a better substrate also for 5000 series, other than better bins.

IF fclk is on the I/O-die I think. You won the I/O-die lottery and lost the core-die lottery
My 3600 is both poorly binned and I/O-die is not top notch either

 

[Zach_01]

IF fclk is on the I/O-die I think. You won the I/O-die lottery and lost the core-die lottery
My 3600 is both poorly binned and I/O-die is not top notch either

Yes IF is located on the I/O Die with the UMC (UnifiedMemoryController) but don’t forget that IF is connecting the I/O Die with the CCD(s) and the CCXs between them.
So it runs through the CPU substrate. So the substrate it self also matters a lot.
The I/O Die is on very mature node now, the 14nm of GlobalFountries and I hardly believe that the Dies have too much difference between them in binning. Could be tho...

 

[mxking035]

I’m not aware tho how the CPU substrate can be across the binned chiplets and CPUs in general. Probably has no correlation.

Most likely.
My cpu can finish cinebench with 4.575/4.475 with 1.368v just tried it. But my problem is IF max out at 1866.

My previous stick (corsair) and now (g.skill) can both run 3800mhz decoupled IF or 1800 IF.
But cannot boot with 1900 flck not matter what recommended voltages, soc, cldo_...

It's easy to spot ram/mobo is the culprit by running decoupled mode.