I’ve been soaking myself with DDR5 OC information these last couple of days, since I’ve been very off in the overclocking scene for a while.
I’m completely new to AM5 game. Never tweaked an AMD CPU before, I’m going for 9800X3D + Gigabyte B650E AORUS STEALTH ICE + 32Gx4 Klevv Cras V Hynix 6000 CL30.
Ok, I’ve been reading AM5 doesn’t like 4 sticks. Let alone dual rank. And even worse with high sizes. Very hard on the memory controller, yadda yadda. People run “very low” speeds to make them work. Like sub 5200.
Now, there’s these Hynix A dies that apparently are very good with speed. They can handle +6000 speeds ““easily””. And timing pretty nice, apparently very common to have CL30 out of the box.
So, my doubt is, why don’t just run 2:1 (uclk = mclk / 2)? Hynix A/M dies are happy with the speed, memory controller is happy with the lower frequency, and the infinity fabric can be sync-ed!
So, makes sense to:
ddr 7200
mclk 3600
uclk 1800
fclk 1800
There must be something I’m not understanding, but afaik 1800 uclk should be VERY lightweight for any AM5 memory controller. Maybe timings to make it stable would become atrocious? But we are talking about 6000 CL30 dual rank chips. Getting 7200 doesn’t seem crazy to me, even if we need to put JEDEC timings. Maybe mobos are the problem? But lately I’ve heard they take these speeds nicely.
I’m here all about to learn, hope to see your opinions on the matter!
DDR 7200 MT/s (transfer rate)
mclk 3600 MHz (clock of the memory, its half the DDR rate because DDR memory doubles the effective speed)
uclk 1800 MHz (clock speed of the integrated memory controller - in CPU. In gear 2:1, it runs at half the mclk)
fclk 1800 MHz (Infinity Fabric clock, the thing that connect the cores and memory)
Doesn’t appear to be needed. UCLK = MCLK works fine to 3.0 GHz with a 2.0 GHz FCLK, sometimes up to 3.2 or 3.3 with the same 3:2 ratio to FCLK.
People’ve been doing 4x32 M-/A-die at 6000 since early Zen 4 and my sense of the overlock reports I’ve come across is it’s gotten (somewhat) easier and more likely to remain stable across AGESA updates over time. Seen a few 2DPC 2R builds get stuck at like 3600 or 4200 but most seem to hit 4800+ without difficulty. Even with B-die.
X870(E) tweakery’s supposed to help some, unclear how much but I wouldn’t guess more than +200 MT/s for quad DIMM. AGESA and TSMC 6N maturity have probably also contributed. 2DPC’s ultimately a signal integrity limit, though, as is 2R, and currently their combined limit seems to be 5600-6000. Maybe 6200 with a lot of tweakery if you’re lucky.
If multiplexed rank (C)UDIMMs happen they might lift things some.
You will not get 4 sticks of DDR5 past DDR5 3600. not 6000, not 5600, not 5200, and probably not 4800. You will get 3600 speeds.
If you want DDR5 at 6000 and CL30 you will need 2 sticks of RAM. And probably a max of 64GB as the 48GB sticks don’t want to hit 6000 whatevers either.
Many folks have reached 5000MHz+ using 4 sticks. It’s not easy, but not impossible.
Also, 2x48GB is totally possible to run at 6000MHz or more with just XMP. 48GB sticks clock better than the 32GB ones for some reason.
WIth a quick search I could even find 2x48GB 6400MHz listed in the QVL of some AM5 mobos.
Nope. I have 6000 cl30 2x48 running stable. Can do 6600 too but it’s slower. Perhaps that’s true of the micron sticks but Hynix 48GB can do high 6000s low 7000s depending on the CPU.
Prolly easier if you do 2DPC 1R instead of 2R. Not a common choice, so little in the way of build reports, but MSI specs both 1DPC 2R and 2DPC 1R at 6400 on their X870(E) boards.
I had no trouble booting 2DPC 2R B-die at 5200. Would be willing to try for 5600 with M-/A-die.
Right cheerio, could you pop round then and explain to the DIMMs in the 4x48 build I did a year ago that they’ve been malfunctioning by operating without errors at 4800 this whole time?
We’ll need you to stop by the 2x48 build I did a couple months ago as well.
I think 2100, but never really stability tested that. 2066 is stable for sure. But I can’t do 1:1 mclk over 6200, that’s why it’s slower. But I run 6000 now since I need to max out VSOC at or close to 1.3 for 6200 and it’s barely faster than 6000.
The two are interchangeable in some non-spatial aspects. The differences in physical topology mean signal integrity differences, though my guess would be 2DPC 1R and 1DPC 2R tend to behave more like each other electromagnetically on the bus than 1DPC 1R in a dual socket channel or 2DPC 2R.
I’ve also found 1DPC fairly responsive to top intake cooling airflow if there’s dual tower air on the CPU with DIMM overhang. In comparison, couldn’t measure any difference to 2DPC. Probably not a significant distinction with AIO and DIMM downflow builds, and maybe not with setback dual tower or single tower, but I haven’t come across comparative measurements.
I think most DIMM manufacturers make matching RGB blanks for 1DPC to look like 2DPC. But I’d expect blanks to have airflow disadvantages similar to 2DPC. Doesn’t seem to be data on how that affects tRFCn and tREFI and I haven’t had luck finding tRFC+tREFI scaling on actual workloads. Also don’t know of RGB on and off comparative thermal data for 1DPC 2R + blank versus 2DPC 1R.
Yeah, we’re shuffling DIMMs again so I was benching DDR4-3400 CL20 XMP against DDR4-3600 CL18 XMP for Zen 3 yesterday. For many benches the difference is small enough it’s not worth all the reruns needed to build empirical probability distributions and estimate the likelihood of 3600 outperforming. I’d expect the relative difference between 6000 and 6200 on the same kit to be smaller.
yeah, still a lot of headaches running 4 DIMMs at high speeds and you can still see skepticism there from me as well. 2 DIMMS, runs just fine. 4, at high speeds, I am skeptical about.
