DDR2 Micron D9GMH vs. D9GKX

[osmiumoc]

Hey you all. I'm a fan of numbers and data and also I want to improve on my DDR2 clocks. So I decided I'll have a take on some memory binning and while I'm at it gather some information about how the favorable Micron chips compare to each other. 

I hope this will not be too redundant with the other threads about DDR2 already posted. What I plan to do is plotting some graphs for voltage/frequency scaling and potentially down the road some cold scaling as well. 

I will use my EP45-Extreme as a constant base with a wolfdale (E8XXX) CPU. Divider 1:2. Generic sticks with no heatsink will be fitted with an OCZ-Reaper sink.
I plan to start with single sticks at 1.90V and see how high they clock on CL5-5-5-15 44 2T with quick S-Pi 1M runs as 'validation'. As soon as they fail I will up the voltage by one step (0.02V) and go again. All the way up to 2.3V. This will hopefully give me a nice resolution on the frequency/voltage relation for these modules. 
I'm very interested in a potential difference between the two different part numbers as well as the spread in general on these.

My focus is on GMHs and GKXs. I spend some time hunting a few down and I will lift heatsinks to check when in doubt. I will edit my results into this first post and if its ok I'll reserve 2nd post for stuff like pictures / oddities I encountered and to share part numbers etc..

Click to enlarge.

So these were my first GKX batch, all 4 were generic blank sticks pulled from an office PC. No binning by Micron other then the 800 CL5 spec. 
The oddball A003 shows a strange frequency wall that suddenly falls (see the 2.22V spike, +24MHz just by setting 2.22 instead of 2.20V. I confirmed that with multiple runs). My current speculation is that it actually does scale up similar to the other ones but the board does not set optimal RAM termination voltage or DRAM ref voltage. I left these at auto and for this test I do not bother to tweak my board to each individual stick.






I'll link some validation screenshots here (usually only bother when I reach a new highest clock):
A001
i.ibb.co/BV2YNVq/GKX-A001-2-3-V-682.jpg
valid.x86.fr/s1h271

A002
valid.x86.fr/hq7lg0

B001
valid.x86.fr/ugy0cn

[osmiumoc]

Just for reference, this is the heatsink I will be using on the blank / generic sticks:
   

27°C @2.3V. 22°C ambient and the sensor was stuck into the thermal pad on the module. I replaced the stock TIM with its horrible glue for some decent pads.

Underneath this thing is the stick on the left:
 
Blank Micron-labeled sticks. On the right are the donor OCZ sticks stripped from the heatsink. 

It worked well already, and they run decent with 650MHz at 2.2V:



These G.Skill F2-6400 2GBHZ have a good chance for GMHs:


I do struggle a bit with the Ballistix-Tracers, they have only Crucial numbers on them. Size-wise they could be microns, but they are spaced a bit odd due to the LED logic.


Maybe the secret to the Ballistix stuff is not the Micron-Chips but rather these super special caps (yes the stick is working, yes this came from the factory like that) :


Update: Found a few more GKXs


And I took a closer look at these Mushkins:

6400 4-5-4-11
According to the ramlist these have a chance for GMHs but I'm not sure if these really are GMHs. The pcb looks the part, the modules physical measurements fit but they suck. Best they could do at 2.2V was 465MHz. They also struggle to boot their profile. But they do 400MHz 5-5-5-18 fine at 1.8V. 


Update: More potential GMHs.

[osmiumoc]

Update 2: I can share a few PNs from experience that will most likely have GKX on them.

Micron Number: MT16HTF12864AY-800D4 (100% GKX from 18/18 sticks so far) and MT16HTF12864AY-80ED4 (100% GKX from 5/5 sticks)
The first part of the PN defines the physical layout important (!!!) is the second part. Things like 800 or 80 indicate the 800MHz bin for GMHs or GKXs.

Crucial Number: CT12864AA80E.16FD (also had the MT16HTF12864AY-80ED4 Micron label on it) these are value rams without heatsink and came with 100% GKX from 4/4 sticks

HP number: 404574-888 CT: R96214LTFVD323 These are also MT16HTF12864AY-800D4. 100% GKX in 10/10 sticks. From my findings, just the HP number alone does not seem to guarantee anything.

They all come without heatsinks and seem to be OEM parts from e.g. HP PCs. Which in this case is a good thing, my first GMH kit as you can see above is absolutly burned out and not really usable while these OEM sticks seem to be in top shape. Probably never seen more then 1.8V their whole life.



If there aint no maniac like me buying them all, you can find them on ebay or with electronics reseller stores for ~ 4-6$ per stick.



For GMHs:

Micron PN: MT16HTF12864AY-667D4 <- Again its the last part that is important. 667 is the speedbin for GMHs and D4 is the marker for GMH ICs. 

The other string for these GMHs: 2Rx8 PC2-5300U-555-12-E0 <- That E0 is important. I other endings seem to be different chips.

BEWARE of sellers sending you different sticks! I had it 2/3 times now that they listed the GMH part numbers but send me single ranks with a completly different PN. Most e-waste recyclers don't pay much attention or simply don't care. They just look at capacity and 1GB is 1GB for them, no matter the rank or ICs.
So far I heard back from one of these cases, they acknowledged that they made a mistake and offered to send me the correct PN right away.

Photo:

[Aleslammer]

Did a little looking for some of the PNs above, not a lot of luck. Used the HP PN had a large selection on Ebay but most were not using Micron ICs. Good post with some great information.

[osmiumoc]

Jul 22, 2020 8:43:26 GMT -5 Aleslammer said:

Did a little looking for some of the PNs above, not a lot of luck. Used the HP PN had a large selection on Ebay but most were not using Micron ICs. Good post with some great information.

Oh thanks for noting that, I thought the HP partnumber was specific enough. I added the CT: XXX number on the HP sticker too. Maybe that helps. 

I'll add more numbers for GMH soon too. But I have to wait for them to arrive. 

The reason you don't see them on ebay is because I literally bought everything I could find worldwide. Beware of some trickery tho, I had one seller who photographed GMHs and even listed the correct PN but send me a completly different PN with single rank sticks and totally different ICs.

[alpi]

Nice Gkx You've found ! 1300 cl5 @ 2.2v is pretty strong !
Cool game finding nice D9's nowadays ! Keep up, I wish You good luck !


I have 4 sticks of BL12864AA1065.16FD5 Crucial Ballistix.
I have limited info on ddr2 types but If You find this sticks, You should buy them I think !

[osmiumoc]

Jul 24, 2020 16:09:30 GMT -5 alpi said:

Nice Gkx You've found ! 1300 cl5 @ 2.2v is pretty strong !
Cool game finding nice D9's nowadays ! Keep up, I wish You good luck !


I have 4 sticks of BL12864AA1065.16FD5 Crucial Ballistix.
I have limited info on ddr2 types but If You find this sticks, You should buy them I think !

Yeah these first 4 GKX are decent. I think getting them as OEM parts is a good idea these days as most OC-labled RAM from Crucial etc.. is heavily used. My GMH kits that I received seem to be burned out 80% of the time. Most of them struggle to hit DDR2 1066 CL5 @2.1V. But then there are a few of them who just fly and do the same or better as the GKX.

Testing is very time intensive. I spend about 1-1,5h per stick so I will continue this sporadically to avoid getting tired.

Updated the third GMH stick, this one is an OEM stick from crucial like my GKX. It seems to be on par. So as a very early assumption from running just these 7 sticks so far: GMH or GKX are very similar within 1.90V to 2.30V. They scale up very well with voltage but there are no surprises. If a stick is bad @1.90V it will also be bad @2.30V. The higher the voltage the more all sticks start to equal out e.g. differences in possible clock is higher at lower voltage and with higher voltage the gap between the different sticks gets a bit smaller. 

Since I had some luck with flashing the OCZ SLI-Ready RAM to a different profile (it did not work before my flash, after flashing it to a Super*Talent GMH profile they run) I decided to look under the heatsink and found GMHs on them. Sadly only one stick of this kit is good, the other one is as bad as B002. 



Same goes for the Mushkin stick. It has GMH but is even worse then B002 (2.3V and can't hit DDR2 1200). I think its not that the GMH bin is bad on those. They are just done after years of use. The sellers I got them from ran them until recently and they had 2.3V and 2.25V profiles as default on them. My board tried to run them at 2.3V as default so I think they probably were seeing this voltage their whole life.

Flashing different SPDs on the bad Mushkin stick did not change its performance. Just on the OCZ kit it enabled me to run them at all. 

Also, I collected some funny photos that I quickly took when something went wrong during windows boot:
    

Even after all these unstable boots, errors and crashes this windows install is managing to hold on. Only thing that slightly worries me is that sfc /scannow instantly crashes, but as long as SPI runs its fine by me. Oh and I had to reinstall memset a couple of times. Seems to be the most likely program to corrupt when you set unstable changes.


You guys think the bot would allow a screen like this?

[Mr.Scott]

Aug 3, 2020 12:45:22 GMT -5 osmiumoc said:

You guys think the bot would allow a screen like this?

LOL
Not a shot in hell.

[osmiumoc]

Brief intermission about thermals:

I taped a K-type with thermal tape to a blank D9GKX stick and ran it at 2.32V @625mhz for a couple 32m runs.




As you can see, with a fan on top of it you can safely run blank sticks too. Just need some good airflow.

Which is good news for me, because I don't have to mount a heatsink each time.

[obscureparadox]

If you're looking for raw clocks then you have to go ballistix unfortunately. That being said, it might be interesting to see what you get from GKX sticks on other PCBs to possibly use as donors onto a Ballistix PCB 🤔🤔

[osmiumoc]

Oct 19, 2020 20:57:53 GMT -5 obscureparadox said:

If you're looking for raw clocks then you have to go ballistix unfortunately. That being said, it might be interesting to see what you get from GKX sticks on other PCBs to possibly use as donors onto a Ballistix PCB 🤔🤔

That is something I thought about in my head. Moving modules from one stick to another. These DDR2 BGAs are simpler to solder as compared to something like a mainboard chipset. I have a hot-air station, maybe I could do it. Will need some testing tho. 

I do have a few good brainpower pcbs with some bad D9 chips on them. Interesting thought. 

[osmiumoc]

Found a small curiosity. Single Rank / one sided D9GMH. Probably nothing special, but it's the first time I saw these single sided sticks with D9GMH. 

Micron PN: MT8HTF6464AY-53ED7 
Puerto Rico Fab
PC2-4200U-444-12-ZZ speedbin

[osmiumoc]

I'm in the process of binning through my GMHs. From the non-OEM ones a shocking amount of garbage piled up. 12 Sticks tested, 2 dead and 8 need 2.6V+ for 650-680 MHz CL5. Don't know if they always were like this or if time and previous owners degraded them. These were my G.Skills, Teamgroup Xtreems and ADATA Vitesta Extremes.

There is however some light. I finally found a partner for stick B001.

hwbot.org/submission/4967554_

valid.x86.fr/fuyhmm

It validates 690+ and runs 675-680 on 2.30V. If they work together I may finally have a very good dual channel GMH kit. Both are G.Skill PHU2-BHZ.

For GMHs I now still have my Cellshocks and one more G.Skill PHU2-BHZ left to test on top of the OEM blank sticks.

[Mr.Scott]

I have a set of HZ's also. They do not clock as high as your's but they have been very reliable for me. They run in everything.

[freeagent]

I had a lot of GMH, they all died. I think I have a set of GKX, still lives.. Does 600 5-5-5..