Friday, October 16th 2020

AMD Ryzen 5000 Series Features Three Synchronized Memory Clock Domains

A leaked presentation slide by AMD for its Ryzen 5000 series "Zen 3" processors reveals details of the processor's memory interface. Much like the Ryzen 3000 series "Matisse," the Ryzen 5000 series "Vermeer" is a multi-chip module of up to 16 CPU cores spread across two 8-core CPU dies, and a unified I/O die that handles the processor's memory-, PCIe, and SoC interfaces. There are three configurable clock domains that ensure the CPU cores are fed with data at the right speed, and to ensure that the MCM design doesn't pose bottlenecks to the memory performance.

The first domain is fclk or Infinity Fabric clock. Each of the two CCDs (8-core CPU dies) has just one CCX (CPU core complex) with 8 cores, and hence the CCD's internal Infinity Fabric cedes relevance to the IFOP (Infinity Fabric over Package) interconnect that binds the two CCDs and the cIOD (client I/O controller die) together. The next frequency is uclk, or the internal frequency of the dual-channel DDR4 memory controller contained in the cIOD. And lastly, the mclk, or memory clock is the industry-standard DRAM frequency.
The three clock domains, according to AMD, are in a 1:1:1 synchrony, i.e. DRAM SDR clock is the same as uclk and fclk (for example, DDR4-3600 would see the memory real clock, uclk, and fclk run at 1800 MHz. AMD also appears to have increased the fclk limit to 2000 MHz with "Zen 3," so you can run memory clock at up to DDR4-4000 without having to engage the mclk:fclk 1:2 divider, which means DDR4-4000 on "Vermeer" will fetch as much of a real-world performance dividend as DDR4-3800 did for "Matisse." With AMD Ryzen, memory overclocking has been an effective way to improve overall system performance, due to the synchrony between mclk and fclk. and so, noticeable performance improvements can be had up to DDR4-4000, beyond which, you can still engage the divider to push mclk further up.
Sources: VideoCardz, WCCFTech, Tecnopat
Add your own comment

51 Comments on AMD Ryzen 5000 Series Features Three Synchronized Memory Clock Domains

#26
lynx29
TheLostSwede
Keep in mind that at least the current Ryzen 3000-series memory controller doesn't seem to like odd CAS latencies, so 15 might be a no go.
Zen 3 is a whole new ball game with ram though so I am not worried about it. also since im aiming for 4000 on my 32 gig kit... hopefully i can get 16-18-18... we will see..
Posted on Reply
#27
Makaveli
Zach_01
The principal is exactly the same with ZEN2, and if true, the only thing changing is the 3800MHz -->> 4000MHz (2000MHz) upper limit for 1:1:1 synchronization.

And I believe that DRAMcalc will probably need an update for the new 5000. But dont expect it to be updated upon Nov5th. These things need time...
It will I was checking today and any attempt to set 4000 with my B-die gives me this error.

Posted on Reply
#28
lynx29
Makaveli
It will I was checking today and any attempt to set 4000 with my B-die gives me this error.


i think what im going to do is calculate safe at 3800 cas 16... then use the safe alts on that safe calc, then change the 3800 to 4000 and leave everything else same, and use max voltage column for all voltages... and change IF to 1:1 - we will see how that goes, prob will crash... in which case i will try to raise some other things but keep the cas 16 at 4000. if i have to go to cas 17 at 4000 im just gonna stay at 3600 cas 14.
Posted on Reply
#29
Zach_01
It may support RAM speeds up to 4000MHz with proper sticks, but we really dont know if the difference of the CPU make difference in settings.
ZEN3 may need other set of settings.
We cant really know anything about it.

Who can tell, and be 100% sure, that ZEN3 can tolerate the same amount of voltages on the memory subsystem?
Dont be eager to try... a few weeks (hopefully) wont kill you. But you could kill your brand new CPU...
Posted on Reply
#31
mtcn77
Makaveli
It will I was checking today and any attempt to set 4000 with my B-die gives me this error.


Perhaps, your dimms are inserted incorrectly?
Posted on Reply
#32
Makaveli
mtcn77
Perhaps, your dimms are inserted incorrectly?
That is not the issue simply dropping the speed to 3800 from 4000 will load the profile.
Posted on Reply
#33
TheLostSwede
lynx29
Zen 3 is a whole new ball game with ram though so I am not worried about it. also since im aiming for 4000 on my 32 gig kit... hopefully i can get 16-18-18... we will see..
How so? I was told they've only done minor changes to memory controller.
Posted on Reply
#34
Chrispy_
TheLostSwede
How so? I was told they've only done minor changes to memory controller.
Not even minor changes. Lisa Su is on record confirming that they're using the exact same I/O die as for Zen2 and the memory controller is on that IO die.

I think the only confusion I remember on this is that Mark Papermaster dodged the question somewhat when interviewed by Anandtech and mentioned incremental advancements in the I/O die. He was referring to the switch from GloFo 14nm to 12nm which dropped idle power use slightly,.
Posted on Reply
#35
mtcn77
Chrispy_
Not even minor changes.
Is it, though?
btarunr
Each of the two CCDs (8-core CPU dies) has just one CCX (CPU core complex) with 8 cores, and hence the CCD's internal Infinity Fabric cedes
That is a major change, no?
Posted on Reply
#36
basco
Mr.Makaveli i think you need to put in dram pcb revision.

from changelog:
DRAM PCB Revision - select PCB RAM for more accurate calculation of timings. In most cases it is recommended to use "A0" for better compatibility.
Posted on Reply
#37
Chrispy_
mtcn77
Is it, though?
Yes. Google is your friend :)
Anandtech and THG both officially confirmed by AMD from different sources.

In case you're wondering where any of the confusion comes from regarding the I/O die manufacturing process, all Zen2 CPUs use an I/O die fabbed on GloFo 12nm. The X570 chipset is fabbed on GloFo 14nm. You might be asking yourself what the X570 has to do with anything?

Well, they're the same thing; The X570 chipset is just a Mattise I/O die made on an older, cheaper process node. and that's why even though a Zen2 I/O die as part of a CPU uses around 12W, the X570 chipset's older 14nm process means it can draw up to 20W and necessitates active cooling.
Posted on Reply
#38
INSTG8R
Vanguard Beta Tester
Chrispy_
Yes. Google is your friend :)
Anandtech and THG both officially confirmed by AMD from different sources.

In case you're wondering where any of the confusion comes from regarding the I/O die manufacturing process, all Zen2 CPUs use an I/O die fabbed on GloFo 12nm. The X570 chipset is fabbed on GloFo 14nm. You might be asking yourself what the X570 has to do with anything?

Well, they're the same thing; The X570 chipset is just a Mattise I/O die made on an older, cheaper process node. and that's why even though a Zen2 I/O die as part of a CPU uses around 12W, the X570 chipset's older 14nm process means it can draw up to 20W and necessitates active cooling.
Well I learned something., I knew it wa GoFlo but thought X570 was also 12nm.
Posted on Reply
#39
TheLostSwede
MxPhenom 216
1.5v what in the actual f****

So with Ryzen, is 2 sticks better than 4? I know with Intel systems there isn't really an issue with 4 sticks and running high speeds and lower latency.
Supposedly four is better than two, unlike Intel.
Not a whole lot of proof out there, but I have no problem running four sticks at higher clocks and lower latencies than they're rated at.
This is the only test I've seen and it looks like the Ryzen 3000's memory controller likes four sticks of RAM and it seems to be the same across games and applications.
I guess more RAM might also help to a degree in these tests. This is also quite old and more recent AGESA/UEFI releases should have further improved things.
www.tomshardware.com/reviews/amd-ryzen-3000-best-memory-timings,6310-2.html
AusWolf
Just had a quick look... The price difference between 3600 and 4000 MHz RAM is massive (at least here in the UK). Will have to wait for benchmarks to see if it's worth it.
With a bit of luck, you can get some 3733MHz modules and run them at 4000MHz, at least if it works anything like my 3600MHz that I run at 3800MHz and tighter latencies.
Again, keep in mind that most memory is designed and tested for Intel or Intel and AMD's memory controller doesn't behave in the same way.
Posted on Reply
#40
Chrispy_
INSTG8R
Well I learned something., I knew it was GoFlo but thought X570 was also 12nm.
It might be at this point. AMD reserves the right to change the product without notification, and that confirmation was 15 months ago.
Posted on Reply
#41
INSTG8R
Vanguard Beta Tester
Chrispy_
It might be at this point. AMD reserves the right to change the product without notification, and that confirmation was 15 months ago.
Well I certainly have a personal beef with it. Despite having active cooling it‘s still the hottest thing in my PC after my GPU under load...
Posted on Reply
#42
mtcn77
Chrispy_
Well, they're the same thing; The X570 chipset is just a Mattise I/O die made on an older, cheaper process node. and that's why even though a Zen2 I/O die as part of a CPU uses around 12W, the X570 chipset's older 14nm process means it can draw up to 20W and necessitates active cooling.
Well, I was looking at the core complex. I wouldn't mind a similar matisse chip, all the better. It seems obsolescence to me to change it every so often, not getting payed enough for that. One less timer is a good thing, imo. Make them stand still without the controller chip, I don't care. They don't sell it on retail.
Posted on Reply
#43
Chrispy_
TheLostSwede
With a bit of luck, you can get some 3733MHz modules and run them at 4000MHz, at least if it works anything like my 3600MHz that I run at 3800MHz and tighter latencies.
Again, keep in mind that most memory is designed and tested for Intel or Intel and AMD's memory controller doesn't behave in the same way.
I'm kind of hoping this is the case too. I have a bunch of 3600 available and absolutely zero sticks of faster RAM. I'm mostly interested in FCLK gains and don't fancy paying through the nose for what is likely to be only a small ~5 improvement over budget-grade 3600. I think the best RAM I have is Hynix C-die that'll do 3600 C16, any ideas what clocks that'll reach if I slacken the timings?
Posted on Reply
#44
TheLostSwede
Chrispy_
I'm kind of hoping this is the case too. I have a bunch of 3600 available and absolutely zero sticks of faster RAM. I'm mostly interested in FCLK gains and don't fancy paying through the nose for what is likely to be only a small ~5 improvement over budget-grade 3600. I think the best RAM I have is Hynix C-die that'll do 3600 C16, any ideas what clocks that'll reach if I slacken the timings?
My Patriot Viper Steel modules are Hynix C-dies and as you can see in my system spec, they run just fine way out of spec.
They're supposed to be 3600MHz 17-19-19-39, but they run fine at 3800MHz 16-19-16-36.
Posted on Reply
#45
Chrispy_
TheLostSwede
My Patriot Viper Steel modules are Hynix C-dies and as you can see in my system spec, they run just fine way out of spec.
They're supposed to be 3600MHz 17-19-19-39, but they run fine at 3800MHz 16-19-16-36.
This stuff?
Posted on Reply
#46
tabascosauz
TheLostSwede
Supposedly four is better than two, unlike Intel.
Not a whole lot of proof out there, but I have no problem running four sticks at higher clocks and lower latencies than they're rated at.
This is the only test I've seen and it looks like the Ryzen 3000's memory controller likes four sticks of RAM and it seems to be the same across games and applications.
I guess more RAM might also help to a degree in these tests. This is also quite old and more recent AGESA/UEFI releases should have further improved things.
www.tomshardware.com/reviews/amd-ryzen-3000-best-memory-timings,6310-2.html


With a bit of luck, you can get some 3733MHz modules and run them at 4000MHz, at least if it works anything like my 3600MHz that I run at 3800MHz and tighter latencies.
Again, keep in mind that most memory is designed and tested for Intel or Intel and AMD's memory controller doesn't behave in the same way.
Distinction there is that you're running 4 single rank sticks, which for most intents and purposes is about equal (in "difficulty" for the IMC to sustain) to 2 dual rank sticks like mine, as they're both 2 ranks per channel and the same density.

I think you'll find that 4 x dual rank sticks with any 8Gb die (B-die, CJR, Rev.E) totaling 64GB total might be a rather different, difficult and underwhelming experience.

Looks like AMD's Achilles heel (IF) won't be changing too much this generation. The UMC remains as strong if not stronger tham Matisse (already had a comparable or better memory controller than Intel), but the +200MHz is probably coming from more optimized IF traces on the substrate and the CCX layout reshuffle moving uncore components physically closer to the I/O die.
Posted on Reply
#47
TheLostSwede
Chrispy_
This stuff?

Nah, those are the Dark or whatever they call them. I think the Viper Steel is one step up from those. That said, it seems like your modules have the same timings.
tabascosauz
Distinction there is that you're running 4 single rank sticks, which for most intents and purposes is about equal (in "difficulty" for the IMC to sustain) to 2 dual rank sticks like mine, as they're both 2 ranks per channel and the same density.

I think you'll find that 4 x dual rank sticks with any 8Gb die (B-die, CJR, Rev.E) totaling 64GB total might be a rather different, difficult and underwhelming experience.

Looks like AMD's Achilles heel (IF) won't be changing too much this generation. The UMC remains as strong if not stronger tham Matisse (already had a comparable or better memory controller than Intel), but the +200MHz is probably coming from more optimized IF traces on the substrate and the CCX layout reshuffle moving uncore components physically closer to the I/O die.
So is the guy running the benchmarks at Tom's. I didn't see the guy I replied to asking if AMD could run four dual rank modules, only two or four modules vs. Intel two and four modules and whether you would gain any performance from running four vs two modules.

There seem to be a few people here are considering four dual rank modules, not sure if anyone ever ended up going for it though.

Supposedly later Ryzen 3000 CPUs can do 3866MHz IF already. Not a huge improvement, but an improvement nonetheless and I guess if AMD is bold enough to say 4000MHz will work, then maybe we'll see some chips doing a bit better than that.
Posted on Reply
#48
AusWolf
TheLostSwede
Supposedly four is better than two, unlike Intel.
Not a whole lot of proof out there, but I have no problem running four sticks at higher clocks and lower latencies than they're rated at.
This is the only test I've seen and it looks like the Ryzen 3000's memory controller likes four sticks of RAM and it seems to be the same across games and applications.
I guess more RAM might also help to a degree in these tests. This is also quite old and more recent AGESA/UEFI releases should have further improved things.
www.tomshardware.com/reviews/amd-ryzen-3000-best-memory-timings,6310-2.html


With a bit of luck, you can get some 3733MHz modules and run them at 4000MHz, at least if it works anything like my 3600MHz that I run at 3800MHz and tighter latencies.
Again, keep in mind that most memory is designed and tested for Intel or Intel and AMD's memory controller doesn't behave in the same way.
I don't want a bit of luck. I prefer being certain. :pimp:
Posted on Reply
#49
TheLostSwede
AusWolf
I don't want a bit of luck. I prefer being certain. :pimp:
Well, when you overclock, it all comes down to luck, or you pay more up front...
Posted on Reply
#50
Woomack
If Ryzen 5000 will work about the same as Ryzen 4000 then you can expect DDR4-4400/IF 2200MHz on pretty much every chip. My ASUS B550/X570 mobos set 2200MHz IF at auto when I pass DDR4-4400 on RAM. Behavior is the same as on Ryzen 3000 with mem at DDR4-3600 and 1800MHz IF. Max on my Ryzen 4650G is 4600 1:1 boot and 4533 1:1 100% stable.
Here are some quick tests 1:1 up to 4533 and async up to 5200 -> www.overclockers.com/forums/showthread.php/797596-G-Skill-Trident-Z-Neo-2x8GB-DDR4-3600-CL14-F4-3600C14D-16GTZNB-w-Ryzen-4650G
Posted on Reply
Add your own comment