SK Hynix A-die / M-die - How to pick amid the XMP / EXPO marketing confusion (Baselining)

[Encrypted11]

I've noticed the narrative on HWZ change in regard to RAM OC.

In the past, it was

4800MHz, 6800MHz, 8000MHz are all the same.

Now

Fast RAM more better.

But a lot of us are still confused on which pairs are better.
Per Toppc from MSI HW/R&D: Every vendor sets its screening criteria for XMP/EXPO. Simply, no standardization... which leads to a MHz arms race similar to GPU and MB inferno; TGP/TDP arms race. Brand marketing is not an objective measurement of baseline quality.

Does RAM vendor A at 6000 C32 1.4V mean it's better than vendor B 6000 C30 1.4V?

This thread is about pre-purchase RAM IC baselining, amid HWZers deliberating over Vendor A vs. B, Kit A vs. B.

 

[Encrypted11]

The module house contestants will be 1 from Taiwan, 1 from China, 1 from Korea.

Among the big 4 desktop board vendors (Asus/Gigabyte/MSI/ASRock), MSI is the only one among the 4 that has given the consumer a little peek at the underwear of the DRAM Module Houses... true JEDEC speeds of the RAMs you're buying... whether they are A-die downbin to 4800 etc.

As a reality check, SK Hynix IC binning is very strict these days just like Intel. An i9 will have a better ASIC quality than i7 than i5. Likewise for JEDEC 5600 vs 4800 class V Hynix. Expecting a 4800 JEDEC IC to work better than 5600 JEDEC is no different from expecting i7 to be faster than an i9.

 

[Encrypted11]

Now let the data speak.

Data Source: https://www.msi.com/Motherboard/MEG-Z890-UNIFY-X/support#mem

G.SKILL (Taiwan) - SPD 5600

Spoiler

G.SKILL (Taiwan) - SPD 4800

Spoiler

Lexar (Acquired by ShenZhen Longsys, China) - SPD 5600

Spoiler

Lexar (Acquired by ShenZhen Longsys, China) - SPD 4800

Spoiler

KLEVV (Korea owned) - SPD 5600

Spoiler

KLEVV (Korea owned) - SPD 4800

Spoiler

 

[Encrypted11]

Summary
Cross check your shortlisted RAMs against MSI Z890 Unify X's QVL. They are openly showing you an open-secret in the DRAM Module house industry in a tabular form.

It is conclusive that an XMP/EXPO stick made out of 5600 JEDEC Hynix are better than 4800 JEDEC ones. Don't be fooled just by XMP numbers incl vDIMM. Screen JEDEC rating and shortlist your kits than follow influencers or netizens selling referral links they recommend. All (well most) vendors sell you 4800 V (downbins) unless you are paying an exorbitant price per DRAM capacity. It's business.

Interestingly, KLEVV naked black DIMMs are JEDEC 5600 while XMP sticks incl RGB are 4800 V (downbin). LEDs and aluminum are not free.

If you find a 4800 V based kit with a very tight XMP profile (but higher priced), you may want to opt for the worst 5600GB based stick and input THAT XMP profile from the former. It will be Plug and Play.

Repeat: Expecting a 4800 JEDEC IC to work better than 5600 JEDEC is no different from expecting i7 to be faster than an i9.

So the conclusion? SK Hynix naked DIMMs #1 and KLEVV naked DIMMs #2, before considering reference PCB or custom PCB based module house kits affecting baseline compatibility since Intel/AMD reference platform/board testing are based on reference DIMMs.

 

[Encrypted11]

Interesting reads:
https://www.chiphell.com/thread-2562013-1-1.html
https://www.hardwareluxx.de/communi...-ram-oc-thread.1306827/page-454#post-30289091

 

[stanlawj]

Now let the data speak.

Isn't this superseded by the clock-driver technology in CUDIMM RAM?
Just buy CUDIMMs and be done with it.

What is CUDIMM?

https://www.kingston.com/en/blog/system-builder/what-is-a-cudimm-csodimm

As DDR5 standard speeds get faster and faster for next-generation desktop and laptop computers, stability and data integrity become critical issues to mitigate. Increasing speeds means electronic noise and jitter become problematic, potentially resulting in errors, lost data or system freezes.

To address this challenge, the body for maintaining industry standards in memory, JEDEC, specified a new requirement in late 2023 for DDR5 Unbuffered DIMMs (Dual Inline Memory Module) and DDR5 SODIMMs (Small Outline Dual Inline Memory Module).

Beginning with the introduction of 6400MT/s DDR5, a Client Clock Driver (CKD) component must be added to the memory module to redrive the clock signal, which minimises noise and jitter and improves overall signal integrity.

New memory classification​

To differentiate from legacy DDR5 modules that don’t feature a clock driver, JEDEC requires the addition of a “C” for “Clocked” to indicate the inclusion of this component, hence the new module types CUDIMM and CSODIMM. This is an important distinction, since CUDIMMs and CSODIMMs use the same sockets as existing DDR5 UDIMMs and SODIMMs. To prevent compatibility issues, the chipset and motherboard manufacturers will roll out BIOS updates to legacy motherboards that disable the clock driver if a CUDIMM/CSODIMM is installed. When a CUDIMM or CSODIMM is installed on a legacy system, it will automatically operate at the supported speed of that chipset and processor.

 

[Encrypted11]

Isn't this superseded by the clock-driver technology in CUDIMM RAM?
Just buy CUDIMMs and be done with it.

So assume a year down the road, if JEDEC DDR5-7200 appears, requires CKD as well. Module houses now make
-Module House A: DDR5-10000 XMP (SPD 6400)
-Module House B: DDR5-9866 (SPD 7200)

How do you tell which one is innatively a better product without peeking into SPD info? Digging up the module datasheets or MB QVLs are needed, still baselining.

Nothing changed.
6400MT/s JEDEC is another tick, 6400 MT/s sticks, same logic applies.

Today,
Intel, only Lunar Lake and Arrow Lake memory controller (based on Intel Ark) supports CUDIMMs.

AMD's memory controller doesn't support CKD yet. Ryzen 7000 IMC is missing a circuit to run in bypass mode. Due to the workaround and technical reasons, POR speeds are far below JEDEC on Ryzen 9000. (Toppc -MSI HW/R&D)

Not all DDR5 systems can run CUDIMMs with updated BIOSes on older platforms not phased out. Not everyone will foot the bill on cutting-edge node chips these days.

This thread is about potentially spending less for more.

 

[wwenze]

Just call JEDEC-6400 to avoid confusion

Remember JEDEC standard / settings use lower voltage and higher timings

 

[wwenze]

Unfortunately I only need one screenshot to prove the opposite

Which die would you rather buy, the 4800 by SK Hynix or the 5600 by Spectek

 

[Encrypted11]

Unfortunately I only need one screenshot to prove

Which die would you rather buy, the 4800 by SK Hynix or the 5600 by Spectek

Spectek, because Micron employee discount

 

[Encrypted11]

Interesting post on baseline quality from Chiphell, where occasionally users working for a hardware OEM/ODM comes into a thread to ssrr.

24/48GB MGBD vs MHBD on silkscreen.

Likewise for JEDEC 4800 vs 5600 Hynix, they are are definitely not the same grade. The silkscreen text is different for a reason.

 

[wwenze]

MGBD = 5600
MHBD = 6400

 

[naked123]

i know G.Skill Trident Z5 6400MT/s CL32 2x16GB kit is SK Hynix M die

 

[Yongkit]

Spectek, because Micron employee discount

https://www.micron.com/content/dam/micron/global/public/spectek/buyers-guide/spectekbuyersguide.pdf

This guide on spectek have different grading too.

Mine crucial pro (micron) was based 5600Mhz

This video show micron is capable of running 6600mhz 1:1 when oc from 5600mhz Stick

 

[wwenze]

I just skimmed thru the video, looks like he went from 5600MT/s to 6600MT/s without modifying any other setting?

Well, yea. That's what happens when you're using JEDEC voltages. You're not going to get 9000MT/s without increasing the voltage to >1.4V. On the other hand, if you grab any random JEDEC module, you are not guaranteed to get these speeds even when you up the voltage.

When we buy a XMP-9000MT/s stick, we are buying the guarantee and testing of the seller company. They tested the stick to work at those speeds. Do you know if your JEDEC-5600 module is going to not-die after testing at 1.45V? Do you know if the PCB itself is designed for 9000MT/s? Do you know what is the yield of those companies testing at XMP-9000? If you have these internal yield data, then by all means, share it, show it, spread the advice. But I see the posts so far and... meh.

This whole thread is like trying to tell people that a good rally car driver is likely also a good F1 driver. But we have test results that directly measure the drivers' F1 performance, so if you're in an F1 competition why be bothered by the rally car performance?

 

[Encrypted11]

One thing on XMP voltages though, primary reason why these XMP sticks usually have a preset between 1.35V-1.45V on vDIMM is not really to tell you how low in voltage it can run in 7000MT/s++ speeds and why it is not 1.75V XMP instead.

It is more related to the cooling capability of the factory heat sink.

 

[matique]

Interesting post on baseline quality from Chiphell, where occasionally users working for a hardware OEM/ODM comes into a thread to ssrr.

24/48GB MGBD vs MHBD on silkscreen.

Likewise for JEDEC 4800 vs 5600 Hynix, they are are definitely not the same grade. The silkscreen text is different for a reason.

WRT MGBD vs MHBD, Gskill used to main MHBD in their 8000+ kits, whilst teamgroup mains MGBD in their 8000+ kits. Eventually, Gskill started dumping their MHBD and now their high end kits are also MGBD. Newest revision in town is the 16gb AHBD, found in the latest gskill 6000c26 kits. It clocks really well, without needing for high voltages.

 

[matique]

I just skimmed thru the video, looks like he went from 5600MT/s to 6600MT/s without modifying any other setting?

Well, yea. That's what happens when you're using JEDEC voltages. You're not going to get 9000MT/s without increasing the voltage to >1.4V. On the other hand, if you grab any random JEDEC module, you are not guaranteed to get these speeds even when you up the voltage.

When we buy a XMP-9000MT/s stick, we are buying the guarantee and testing of the seller company. They tested the stick to work at those speeds. Do you know if your JEDEC-5600 module is going to not-die after testing at 1.45V? Do you know if the PCB itself is designed for 9000MT/s? Do you know what is the yield of those companies testing at XMP-9000? If you have these internal yield data, then by all means, share it, show it, spread the advice. But I see the posts so far and... meh.

This whole thread is like trying to tell people that a good rally car driver is likely also a good F1 driver. But we have test results that directly measure the drivers' F1 performance, so if you're in an F1 competition why be bothered by the rally car performance?

I think it's difficult for these mobo/ram companies to guarantee a certain speed at a certain voltage, as there's far too many variables to consider. XMP rating is lowest common denominator that most boards could do. Some CPUs have good mem controllers, some boards have better tracing, some ram chips require less voltages etc.

There's OC specific kits that do not exist on any mobo QVL list. I know teamgroup and oloy have kits that have specifically binned mem chips, purchased only via word of mouth.