Jump to content
Khenglish

[HARDWARE MOD] GTX980M Hynix to Samsung memory swap

Recommended Posts

So I didn't like that the memory on my 980m only clocked to 6.4 GHz after raising the voltage to 1.48V from 1.35V, and wanted my memory to run even faster. I knew someone with a spare 970, so we made a deal where I buy the card, and if it still worked after I switched all the memory chips, he'd buy it back (for reduced amount if it could no longer do 7GHz, but at least 6GHz). Long story short, he bought the card back and I got faster memory.

 

MSI 970 4GB Lightning original memory: Samsung K4G41325FC-HC28 (7GHz rating, 8GHz max overclock)
MSI 980m 4GB original memory: Hynix H5GQ4H24MFR-T2C (6 GHz rating, 6.4GHz max overclock)

 

Both cards are GM204 chips. The 980m has one less CUDA core block enabled than the 970, but it has the full 256-bit memory interface and L2 cache with no 3.5GB issues, while the 970 is 224-bit with 1/8th of the L2 cache disabled. Both cards are 4GB with 8 memory chips.

 

I highly suspected this memory swap would work because video cards read literally nothing from a memory chip. There is no asking for what the chip is or even the capacity. They write data to it and hope they can read it back. Memory manufacturer information read by programs like GPU-z isn't even read from the memory. It's set by an on-board resistor. I also had changed multiple memory chips in the past, so was fairly confident I could physically do the job.

 

I started with just one chip switched from both cards. This meant both cards were running a mix of memory from different manufacturers and of different speed ratings, but same internal DRAM array configuration. Both cards worked. Here is a picture of the 980m with one chip switched over:

 

JXVyaWy.jpg

 

Now how did the cards react? The 980m behaved no differently. No change in max overclock. The 970 though... I expected it to be slower... but...

 

970 with 1 Hynix chip, 7 Samsung (originally 8 Samsung)
7GHz = Artifacts like a crashed NES even at desktop
6GHz = Artifacts like a crashed NES even at desktop
5GHz = Artifacts like a crashed NES even at desktop
2GHz = Fully Stable, 2d and 3d

 

I didn't try 3GHz or 4GHz, but yeah, HUGE clock decrease. I shrugged though and kept switching all the memory figuring that as long as it worked at any speed, I could figure out the issue later. With switching more chips through 7/8 switched there was no change in max memory clocks.

 

What was really fun was when I had 7/8 chips done. My GDDR5 stencil got stuck and ripped 3 pads off the final Samsung chip. Needless to say there was a very long swearing spree. Looking up the datasheet I found that 2 pads were GND, and a 3rd was some active low reset. Hoping that the reset was unused, I checked the 970's side of the pad and found it was hardwired to GND. This meant the signal was unused. I also got a solder ball on a sliver of one of the GND pads that was left, so I was effectively only missing a single GND connection.

 

I put the mangled 8th chip in the 980m and it worked. Net gain after all of this... 25 MHz max overclock. Something was obviously missing. I figured I would switch the memory manufacturer resistor, hoping that would do something. I saw that Clyde found this resistor on a k5000m, and switching it to the Hynix value from Samsung had no effect for him. He found that for Hynix on the k5000m the value was 35k Ohms, and for Samsung 45k Ohms. I searched the ENTIRE card and never found a single 35k Ohm resistor. Meanwhile the 970 also worked with all 8 chips swapped, at a paltry 2.1 GHz.

 

Then I got lucky. Someone with a Clevo 980m killed his card when trying to change resistor values to raise his memory voltage. His card had Samsung memory. He sent his card to me to fix, and after doing so I spent hours comparing every single resistor on our boards looking for a variation. Outside of VRM resistors there was just a single difference:

 

AIU6Ph3.jpg

 

On his card (his is shown here) the boxed resistor was 20k Ohms. On mine it was 15k Ohms. I scraped my resistor with a straight edge razor (I could not find a single unused 20k resistor on any of my dead boards) raising it to 19.2k, hoping it was close enough.

 

And it was! Prior to this I also raised the memory voltage a little more from 1.48V to 1.53V. My max stable clocks prior to the ID resistor change were 6552 MHz. They are now 6930 MHz. 378 Mhz improvement.

 

Here's a 3dm11 run at 7.5 GHz (not stable, but still ran)
http://www.3dmark.com/3dm11/10673982

 

Now what about the poor 2GHz 970? I found its memory ID resistor too:

6q6iGuQ.jpg

 

Memory improved from 2.1 GHz to 6.264 GHz. Surprisingly the memory was slower than it was on the 980m. I expected the 970's vBIOS to have looser timings built in to run the memory faster. As for why the memory was over 100MHz slower than the 980m, 980m actually has better memory cooling than the 970. With the core at 61C I read the 970's backside memory at 86C with an IR thermometer. The Meanwhile the 980m has active cooling on all memory chips, so they will be cooler than the core. In addition, the 980m's memory traces are slightly shorter, which may also help.

 

The 980m at 6.93 GHz is still slower than the 8 GHz that the 970 was capable of with the same memory. I'm not sure why this is. Maybe memory timings are still an issue. Maybe since MSI never released a Hynix version of the 970 meant leftover timings for an older card like a 680 were run, instead of looser timings that should have been used (I know in system BIOS tons of old, unused code get pushed on generation after generation). I don't know, just guessing. Talking to someone who knows how this stuff works would be great. I still want 8 GHz.

 

Some more pics. Here's one with the 970 about to get its 3rd and 4th Hynix chips:
Xuu0dxa.jpg

 

Here's my 980m with all memory switched to Samsung. Sorry for the blurriness:
7B8bHQm.jpg

 

So in summary:

 

1. It is possible to mix Samsung and Hynix memory, or switch entirely from one manufacturer to another, with some limitations.

 

2. There is a resistor on the pcb that is responsible for telling the GPU what memory manufacturer is connected to it. This affects memory timings, and maybe termination. It has a large impact on memory speed, especially for Hynix memory. This resistor value can be changed to another manufacturer. It is not guaranteed that the vBIOS will contain the other manufacturer's timings. If it does they may not be 100% correct for your replacement memory.

 

3. If you take a card meant for Hynix memory, you can mix Samsung memory of the same size if it is a faster memory. If the memory is the same speed, the penalty for running Samsung with Hynix timings may hurt memory clocks.

 

4. If you take a card meant for Samsung memory, you cannot mix any Hynix memory without MAJOR clock speed reductions without also changing the memory manufacturer resistor. It is not guaranteed that the vBIOS will contain the other manufacturer's timings, or if it does 100% proper timings for your specific memory.

 

5. For Kepler cards the Samsung resistor value is 45k, and for Hynix 35k. For Maxwell cards the Samsung resistor value is 20k, and Hynix 15k.

 

Next up is changing the hardware ID to be a 980 notebook. Clyde  also found HWID to have an impact on the number of CUDA core blocks enabled. In about a month I can get a hold of a 970m that someone is willing to let me measure the resistor values on. It has the same pcb as the 980m. Does Nvidia still laser cut the GPU core package? We will find out.

 

  • Thumbs Up 8

Share this post


Link to post
Share on other sites

@Prema@Khenglish Hey guys, I have a dead 980M that I am trying to resuscitate by replacing the MOSFETs along the top edge of the card. I have completely destroyed one and damaged another one, but it seems that no amount of heat will get them off. I am using a hot air station and went up to 500°C and it's like it is glued on. All of the black plastic is broken off of one of them and I even tried using a soldering iron directly on the metal substrate that is exposed on that one. It seems like the solder refuses to melt.

 

Have you run into this before? Any suggestions how to get them off?

Share this post


Link to post
Share on other sites

@Khenglish Been waiting for this type of guide for a while now. Looking forward to see how things go with the 970m. Question is if the timings can be manipulated via vbios for the new memory, maybe @Prema would be someone to consult with about this? 

 

P.S. Nice job, promoting it. :)

Share this post


Link to post
Share on other sites

My first read article for the year,great motivation to get the year started,people with your motivation and talents sure make our computing experience a much happier one.Was always interested in doing memory swaps on ddr chips,now seeing done on a GPU is amazing to say the least.

Share this post


Link to post
Share on other sites

Hi @Khenglish

From what I remember in the technical documentation of the Hynix and Samsung chips, there is differences in pin out.
As you showed, not significantly enough to work side by side. However, most probably it's the reason of decrease its speed in combination.
Swap for a faster chips from the same manufacturer I already did twice in my damaged K5000M (2 pieces Hynix T2C to ROC) and in GTX680M (Samsung FC04 to FC03 complete set). All work well but even after hardware OV no fireworks.
Recently I realized that the swap of all the chips from another manufacturer and probably a higher speed is possible through GDDR5 manufacturer ID resistor.
For increase it clock speed in GTX980M I suggests to use GDDR5 manufacturer resistor ID values from GTX970.
Regards

Edited by Clyde

Share this post


Link to post
Share on other sites
5 hours ago, Clyde said:

Hi @Khenglish

From what I remember in the technical documentation of the Hynix and Samsung chips, there is differences in pin out.
As you showed, not significantly enough to work side by side. However, most probably it's the reason of decrease its speed in combination.
Swap for a faster chips from the same manufacturer I already did twice in my damaged K5000M (2 pieces Hynix T2C to ROC) and in GTX680M (Samsung FC04 to FC03 complete set). All work well but even after hardware OV no fireworks.
Recently I realized that the swap of all the chips from another manufacturer and probably a higher speed is possible through GDDR5 manufacturer ID resistor.
For increase it clock speed in GTX980M I suggests to use GDDR5 manufacturer resistor ID values from GTX970.
Regards

 

Where did you find pinouts? I only found a Micron datasheet with pin descriptions. I never found any Samsung or Hynix.

 

I did use the 970's resistor value. It matched the 970m's value at 20k.

Share this post


Link to post
Share on other sites

I was wrong, The differences were between pin out of Hynix and Elpida chips. Rather I did not have of Samsung pin out documentation, or at least I can not find anything.

links to Hynix technical documentation no longer work. :( I only have documentation of Elpida EDW2032BBBGxx chip.

 

  • Thumbs Up 1

Share this post


Link to post
Share on other sites
Just now, johnksss said:

@Khenglish

Is there a place to buy just the chips themselves?

 

No, or else I would have done that instead of crippling a 970. Getting preballed chips also would have saved 75% of the total time on this.

Share this post


Link to post
Share on other sites
2 minutes ago, Khenglish said:

 

No, or else I would have done that instead of crippling a 970. Getting preballed chips also would have saved 75% of the total time on this.

And you only replaced half of them as well?

Share this post


Link to post
Share on other sites
Just now, johnksss said:

And you only replaced half of them as well?

 

No, all of them. The MSI 980m does not have backside memory.

Share this post


Link to post
Share on other sites
Just now, Khenglish said:

 

No, all of them. The MSI 980m does not have backside memory.

Ah, got it. Thanks.

 

Going to test that 20k ohm resistor and see if mine is rated the same as well.

Share this post


Link to post
Share on other sites

This is exactly what I've been looking for..I have a Evga 770 GTX SC 2GB version, I'm decently versed in messing with the bios on Graphics cards, and laptops etc, and recovering from fails using my lil bus pirate

and decent with soldering

 

I was pondering, could such a thing be attempted with upgrading a 2GB card to a 4GB card using higher density ram chips of the same brand, I know that the 4gb version of my card doesn't use extra chips, it uses higher density chips on the same pads

 

I also know the bios of the older ATI/AMD cards had the memory timing tables in the bios that you can select or alter, but from the looks of your post it appears to be selected in hardware via a resistor of a certain value for nvidia based cards, is that likely to be the case for the desktop editions of the cards aswell?

 

if tldnr:

is it a possiblity to upgrade the ram on a 2gb nvidia card when a 4gb is avalible for purchase?

and is it likely the desktop cards have a similar resistor

 

edit: why would I want to try this? cause tinkering moar power! and excuse to finally get a hot air station and some bga templates

 

Edited by Raxis

Share this post


Link to post
Share on other sites

I expect the higher density chips to work, but I don't know if you can get the higher capacity. I learned that memory actually does contain data on its capabilities, but most BIOS never bother reading it, so if it doesn't ask who made it, I doubt it asks its size.

 

Flashing the 4GB card's vBIOS on to yours should do the trick in enabling the capacity by my understanding, if you can get a hold of it.

Share this post


Link to post
Share on other sites

Hi Khenglish,

   I was wondering if you could tell me the value of one of the resistors on your 980m card?   In the photo of the bottom of the 980m, it the one above the 'D503' in the top right.  

 

Basically someone gave me a dead 980m mxm card identical to the one in your photo.  Comparing the underside of the card with the photo you posted I've notice some components are missing.  From the solder it looks like they were broken off rather than never being populated.   I found a schematic of the Clevo p65x on the web and worked out the value of most of the missing components. Unfortunately the schematic is not for an MXM card so some bits are different, explicitly the part to enable power on the card which includes that resistor.

 

It's impressive that you managed to change the ram successfully!  Thanks for writing about it.    

Share this post


Link to post
Share on other sites
7 hours ago, Timp74 said:

Hi Khenglish,

   I was wondering if you could tell me the value of one of the resistors on your 980m card?   In the photo of the bottom of the 980m, it the one above the 'D503' in the top right.  

 

Basically someone gave me a dead 980m mxm card identical to the one in your photo.  Comparing the underside of the card with the photo you posted I've notice some components are missing.  From the solder it looks like they were broken off rather than never being populated.   I found a schematic of the Clevo p65x on the web and worked out the value of most of the missing components. Unfortunately the schematic is not for an MXM card so some bits are different, explicitly the part to enable power on the card which includes that resistor.

 

It's impressive that you managed to change the ram successfully!  Thanks for writing about it.    

 

No worries.  I Stuck a 4.7k on and amazingly the card is working!!  

Share this post


Link to post
Share on other sites

Hello!

You only one man who was try do this and post in internet, i have questions.
Say me please:
1. What will be happen, if replace memory chip 'K4G20325FD-HC03' (256mb from 2GB) to 'K4G80325FB-HC03' (1GB from 8GB)?
All chips manufactured 'Samsung'.
2. Is this will be worked?
3. How many 'capacity' of all memory i will get if i replace 2 from 8 chips?
Thanks!
Regards!

Share this post


Link to post
Share on other sites

Create an account or sign in to comment

You need to be a member in order to leave a comment

Create an account

Sign up for a new account in our community. It's easy!

Register a new account

Sign in

Already have an account? Sign in here.

Sign In Now

×

Important Information

By using this site, you agree to our Terms of Use. We have placed cookies on your device to help make this website better. You can adjust your cookie settings, otherwise we'll assume you're okay to continue.