DDR5 CUDIMM Explained & Benched

[N/A]

So close to DDR6 now. Who would want to buy this. Does it at least lower the voltage required to 1.25.

 

[Wirko]

I feel like CKD's will end up integrated on motherboards eventually as well midway between CPU and system memory. I imagine possibly on the rear rather than front because it could help with RF given CPU and memory are on the front side alternating a CKD chip on the rear side in between is a bit akin to like twisted wires for Ethernet or audio to help improve signal integrity. Essentially a CKD is weight average improvement of signal integrity to boost up MT/s.

The CKD acts on the clock signal, which is not bi-directional but only ever travels from the IMC to the DRAM, at a speed of ~200 mm/ns, or ~50 mm per clock period.

CUDIMMs are not currently supported on AMD platforms. They work, as you see in the thread, but in bypass mode - this means the CKD chip which is what amplifies the signal and makes the CUDIMM what it is remains disabled.

It still does something - perhaps some amplification of the clock signal (red lines in the diagram, which shows that the word "bypass" shouldn't be taken literally). However, it does not regenerate the clock signal.

 

[bug]

It still does something - perhaps some amplification of the clock signal (red lines in the diagram, which shows that the word "bypass" shouldn't be taken literally). However, it does not regenerate the clock signal.

Forgive my ignorance, but what's the difference?

 

[Aquinus]

no need. Rocking solid DDR 4 and 3200


View attachment 392641

I mean, if we're in a race to the bottom, my tower still has a 3930k and DDR3-2133 in it. I'll eventually upgrade.

 

[_roman_]

With my DDR5 6000 CL30 kit I see this with expo

5600 MT/s Hynix a-die - just shared for information purposes. I mentioned earlier 100ns is far too high.

Daily stable / gentoo linux compiler stable

I really have no knowledge about overclocking. I do not tune it to dead.

Other settings are windows 11 pro 24h2 stable. But will crash the gentoo compiler. = unstable.
I tried to figure out a year ago how to get all those windows test software and how to use it. Useless in my point of view.
When my gentoo crashes during compiling. And it will - I know its unstable. I can play for 12 hours a day in windows 11 pro without any crash with those "crash" settings.

AM4 also had something around 55 - 60ns afaik for my Ryzen 5800X ~ 3 years ago. (all screenshots already deleted - value to the right - on the first row.)

I only change one setting and test it for two weeks. If something wrong I revert via uefi profile on my flash drive for uefi updates. (asus and msi have a feature to save all uefi=bios settings to a usb pen drive)
e.g TCL 30 = won*t work - i revert = does not boot - needs more voltage most likely. I do not want to go voerr 1.35V DC yet.

No, there is nothing on the 64 (plus ECC) data lines between the IMC and the DRAM chips, except in certain types of server memory:

FBDIMM DDR2 (Fully Buffered)
LRDIMM DDR3/DDR4 (Load Reduced)
MRDIMM DDR5 (Multiplexed Rank)

May I ask. Are you 100% sure?
It says "redriver" / clock improve circuitry. That means it some sort of logic stone which introduces extra clock cycles. Which introduces latency when you have similar stone without that circuitry. I talk about extra clock cycles.

I do not think DDR2 / DDR3 / DDR4 had those redriver circuits. I think we are discussing CUDIMM no UDIMMS.

In my point of view the article lacks information for myself

How about two full schematic for regular 5600MT/s udimm vs Cudimm one. If possible with full datasheet for all parts. I really doubt the basics have changed in past 25 years how logic stones operate.
without that I can not agree or disagree on "opinions". Which is very bad as the schematic and datasheet for every part in the schematic for both udimm and cudimm module would make it 100% clear.


A simplified block diagramm where most of the information is left out - does not answer it. Datasheets answer it 100% + with corresponding schematic which is a proper schematic.

 

[Makaveli]

5600 MT/s Hynix a-die - just shared for information purposes. I mentioned earlier 100ns is far too high.

Daily stable / gentoo linux compiler stable

I really have no knowledge about overclocking. I do not tune it to dead.

Other settings are windows 11 pro 24h2 stable. But will crash the gentoo compiler. = unstable.
I tried to figure out a year ago how to get all those windows test software and how to use it. Useless in my point of view.
When my gentoo crashes during compiling. And it will - I know its unstable. I can play for 12 hours a day in windows 11 pro without any crash with those "crash" settings.

AM4 also had something around 55 - 60ns afaik for my Ryzen 5800X ~ 3 years ago. (all screenshots already deleted - value to the right - on the first row.)

View attachment 392745

AM4 uses DDR4 so cannot compare those to DDR5.

Your numbers look good here. With your tREFI at 65535 how are your memory temps. That settings is very temperature sensitive I believe anything thing over 60c you may start to see errors. You are using SR sticks and not Dual Rank so you can push them alittle abit more.

 

[AnarchoPrimitiv]

Would it be possible to add a third memory kit to the performance graphs? One somewhere in the middle of those two to give us an idea how the performance scales, maybe a 7200 or 8000 kit?

 

[ir_cow]

Is there anyway to check what standard jedec timings are programmed into the SPD??

Here is the JEDEC SPD as requested. Thaiphoon Burner just gives me a error saying i'm missing drivers are some crap and needs Windows 10. So AIDA64 it is!

This comes from Micron memory and it looks incomplete, though this is the same from ADATA, Team Group and Corsair in the JEDEC part.

Would it be possible to add a third memory kit to the performance graphs? One somewhere in the middle of those two to give us an idea how the performance scales, maybe a 7200 or 8000 kit?

No. This is not a scaling articling. A CUDIMM memory review will be a better way to look at scaling on Ultra 200 series.

 

[Sol_Badguy]

The main purpose of the article to to explain a bit about how CKD works and the benefits. It is truly amazing I can take a sub $200 B860 chipset motherboard and run 8400 MT/s with ease. Before CUDIMM, this required a ITX motherboard or the much more expensive 2-DIMM slot Intel Z790 motherboards like EVGA Dark, ASUS Apex or Gigabyte Tayhon.

That's the thing you used a $700 mobo. I'm not saying you can't do it with a much cheaper one but if it's not included (as in actually tested) in the article then the article doesn't truly advertise that benefit. Also a 48GB 8400 MT/s kit would cost more than a budget mobo, why would someone pair expensive RAM with a cheap mobo???

Good to see solid gaming improvements with the CUDIMMs. The original TPU review of Arrow Lake had it using only DDR5 6000C36 which seems to have crippled its gaming performance.

You can say that again. Raptor Lake, Ryzen 7000 & 9000 are running overclocked RAM but Arrow Lake underclocked.

I think people here are confusing what this article is about and what they want it to be.

It is NOT an Intel Ultra 200 Memory Scaling Article.

It IS an article about CUDIMM Technology and the potential benefits it brings to Intel Ultra 200 series currently and future platforms that will support it.

True. Perhaps unfulfilled promises caused strong reactions on this one.

Intel Core Ultra 9 285K Review

Finally! Intel's new Arrow Lake architecture is launched. The new CPUs are full of design changes, like removal of Hyper-Threading, new Lion Cove P-Cores, an improved Thread Director and more. In our review we got surprising results that were both impressive and disappointing.

www.techpowerup.com

G.SKILL sent us their upcoming F5-8800C4255H24GX2-TZ5CK, a 2x 24 GB DDR5-8800 CUDIMM memory kit—very impressive.
We will be using this in an upcoming Arrow Lake Memory Scaling article. For now, see below for some quick results.

The crowd is getting restless.

It's a technology showcase article, not a technical article. It's something aimed at letting people know this technology exists and how it can benefit them. It is of no use to enthusiasts or anyone with a technical level of understanding.

It's not for cubicle workers running office workloads on OEM prebuilt PCs either.
Hence the very strong criticism of comparing a champion CUDIMM kit against a JEDEC (read trash) kit for Arrow Lake. Especially with nothing else in between.

or "why would you test anything other then 4k ultra on the 4090 nobody will use it with any other settings in real life!"

Actually for some very demanding games with all the bells & whistles turned on, upscaling from 1440p is better. Meaning the game is running internally at 1440p and then upscaling, so the native 1440p results become relevant, the performance is going to be slightly worse than native 1440p because of the overhead but still those will be the most relevant results from the suite of tests.

Except an OEM prebuilt is likely to have the JEDEC 6400 installed.

Yeah but not with a $700 mobo and a K CPU. And it definitely won't be used for gaming with a 4090.
So why make this article sound like it caters to everyone: OEM people, DIY people, overclockers/benchers, casuals etc.?

Did you actively searched for the worst kit available on the market??

I keep imagining it has a green PCB and no heatsinks. Ugh can't unsee it now, it makes my skin crawl, LMAO!

 

[Dr. Dro]

The CKD acts on the clock signal, which is not bi-directional but only ever travels from the IMC to the DRAM, at a speed of ~200 mm/ns, or ~50 mm per clock period.


It still does something - perhaps some amplification of the clock signal (red lines in the diagram, which shows that the word "bypass" shouldn't be taken literally). However, it does not regenerate the clock signal.
View attachment 392730

Interesting. Signal amplification will likely still help a little with squeezing the extra few MHz out of it, but without the redriver operational, you won't really get the full benefits of it.

I've been waiting for the Trident Z5 CK 9600 to become available here in Brazil for some time, but thanks to the current administration's hostility to imports and the relative flop of the Core Ultra processor, no shops seem to be carrying it. Things look bad on the Intel camp in my corner of the world.

 

[Chrispy_]

Great article and thanks for posting it.

View attachment 392627

this latency on the 6400 kit at 100.20ns is brutal though no memory sub timings were tweaked for this at all can probably get that much lower.

Presumably DDR5-6400 CL32 brings latency down by a good ~40ns from those loosey-goosey CL52 timings?

 

[Makaveli]

Presumably DDR5-6400 CL32 brings latency down by a good ~40ns from those loosey-goosey CL52 timings?

Yes I would assume so.

 

[ir_cow]

Presumably DDR5-6400 CL32 brings latency down by a good ~40ns from those loosey-goosey CL52 timings?

Kinda. Using Intel Arrow Lake , the latency will always be high at stock unless you change the Cache Ring and D2D ratio. Once you change the DRAM from 6400 CL52 1.1v its no longer in JEDEC specifications though.

 

[tfp]

And considering how easy it is to change the D2D clock rates with out changing voltages it seems odd that Intel picked such a low number. Cache Ring doesn't seem to move much with out playing with voltages.

 

[Braegnok]

Kinda. Using Intel Arrow Lake , the latency will always be high at stock unless you change the Cache Ring and D2D ratio. Once you change the DRAM from 6400 CL52 1.1v its no longer in JEDEC specifications though.

Seems that some folks are not grasping the meaning of JEDEC DDR5-6400 specifications.

I have the Crucial 32GB DDR5-6400MTs JEDEC CUDIMM kit, and in default settings it runs at 52-52-52-103 @ 1.1v,... that is the specification.

Crucial 32GB Kit (16GBx2) DDR5-6400 CUDIMM | CT2K16G64C52CU5 | Crucial.com

Buy Crucial 32GB Kit (16GBx2) DDR5-6400 CUDIMM CT2K16G64C52CU5. FREE US Delivery, guaranteed 100% compatibility when ordering using our online tools.

www.crucial.com

The Crucial kit has the new Micron D8DKT ICs,..

The Crucial 6400 CUDIMM kit also works very well in bypass mode,.. with the right combo the kit will do 6800 CL32 1:1,... but that's not JEDEC spec.

 

[Makaveli]

Seems that some folks are not grasping the meaning of JEDEC DDR5-6400 specifications.

I have the Crucial 32GB DDR5-6400MTs JEDEC CUDIMM kit, and in default settings it runs at 52-52-52-103 @ 1.1v,... that is the specification.

Crucial 32GB Kit (16GBx2) DDR5-6400 CUDIMM | CT2K16G64C52CU5 | Crucial.com

Buy Crucial 32GB Kit (16GBx2) DDR5-6400 CUDIMM CT2K16G64C52CU5. FREE US Delivery, guaranteed 100% compatibility when ordering using our online tools.

www.crucial.com

View attachment 392851

The Crucial kit has the new Micron D8DKT ICs,..

View attachment 392854

The Crucial 6400 CUDIMM kit also works very well in bypass mode,.. with the right combo the kit will do 6800CL32 1:1 @ 1.7v,... but that's not JEDEC spec.

View attachment 392855

I think everyone know what jedec specs are just nobody actually runs their memory at jedec out in the wild.

So useful numbers in a review but real world not so much.

 

[ir_cow]

I think everyone know what jedec specs are just nobody actually runs their memory at jedec out in the wild.

So useful numbers in a review but real world not so much.

Nearly all pre-builts use JEDEC. It's also very common to see people think that XMP/EXPO will enable itself or filling all the ram slots gives the best performance.

I'm currently running 5600 MT/s in my workstation / gaming computer because I need higher capacity over bandwidth.

 

[Makaveli]

Nearly all pre-builts use JEDEC. It's also very common to see people think that XMP/EXPO will enable itself or filling all the ram slots gives the best performance.

I'm currently running 5600 MT/s in my workstation / gaming computer because I need higher capacity over bandwidth.

You are right and will depend on the user and use case.

I would hope vendors selling machines would use XMP/EXPO but the only prebuild machines I touch these days would be the ones that I order for work coming from Lenevo, so laptops and TR workstations.

understand your example about 4 slots vs 2 and having to run lower speeds to keep it stable. RAM capacity is the reason I went with 32x2 sticks for AM5 instead of 16x2 slightly faster sticks.

 

[xBruce88x]

Hopefully AMD's new "AI" cpus rumored to be coming to desktop will support this. APU users will love the bandwidth for sure!

 

[mechtech]

Seems that some folks are not grasping the meaning of JEDEC DDR5-6400 specifications.

I have the Crucial 32GB DDR5-6400MTs JEDEC CUDIMM kit, and in default settings it runs at 52-52-52-103 @ 1.1v,... that is the specification.

Crucial 32GB Kit (16GBx2) DDR5-6400 CUDIMM | CT2K16G64C52CU5 | Crucial.com

Buy Crucial 32GB Kit (16GBx2) DDR5-6400 CUDIMM CT2K16G64C52CU5. FREE US Delivery, guaranteed 100% compatibility when ordering using our online tools.

www.crucial.com

View attachment 392851

The Crucial kit has the new Micron D8DKT ICs,..

View attachment 392854

The Crucial 6400 CUDIMM kit also works very well in bypass mode,.. with the right combo the kit will do 6800 CL32 1:1,... but that's not JEDEC spec.

View attachment 392855

what will it do at 1.100v?

Here is the JEDEC SPD as requested. Thaiphoon Burner just gives me a error saying i'm missing drivers are some crap and needs Windows 10. So AIDA64 it is!

This comes from Micron memory and it looks incomplete, though this is the same from ADATA, Team Group and Corsair in the JEDEC part.

View attachment 392755


No. This is not a scaling articling. A CUDIMM memory review will be a better way to look at scaling on Ultra 200 series.

thanks
so its a 6400 jedec stick............again still weird that the ddr4-3200 is 22-22-22.........one would have thought that ddr5-6400 would have been 44-44-44

Dammit AMD pull your balls together and fix the damn latency lol..

I don't even want to show you what mine is right now.. you would all laugh..

But it "feels" no different than before. The bump to 2R is what I can "feel" in game..

Ugh. Dammit.

My ddr4-3200 is stock 22-22-22
laugh all you want its stable at least lol

 

[Braegnok]

what will it do at 1.100v?

Sorry I only tested default settings for about 30-min, I'm guessing your not going to see any thing performance wise much higher then 6400 CL52 running 1.100v.

The Micron ICs do not scale with voltage, or overclock like Hynix,... if your overclocking or running XMP, or EXPO your better off buying 6400 CL32 1.35v kit.

https://www.newegg.com/team-32gb-ddr5-6400/p/N82E16820985153

 

[remixedcat]

is this gonna have a CAMM form factor or be soldered?

 

[T1beriu]

Nearly all pre-builts use JEDEC.

I thought you used regular DDR5 at 6400 CL52. Now I see that was CUDIMM as well. I apologise for my previous comment.

 

[ir_cow]

I thought you used regular DDR5 at 6400 CL52. Now I see that was CUDIMM as well. I apologise for my previous comment.

I didn't make it clear enough in the article. No worries.

 

[Makaveli]

265k isn't great either but memory latency with 8200 CUDIMM, Ring at 4Ghz, D2D at 3.2Ghz, and NGU at 3.2Ghz is similar to Makaveli's 9800X3D latency.

AID64 isn't as accurate with Hypervisor enabled, I might try booting in safe mode just to see if it changes.

View attachment 392655

After booting to safemode, looks like CPU clock speed was down impacting Cache scores but memory latency dropped as well, nice work Windows.

View attachment 392661

There are some extra tricks people are using to get lower numbers like

Going into the bios and changing Core tunning from Auto to legacy which took me from 71 to 68.

Then of course the Safe mode option

Both of those combined now has me at sub 65ns.

But these are not real world as I don't use my system in safe mode or in legacy mode.

So tweaked at 71ns is what I actually see in my day to day use.