kingpincooling.com

Go Back   kingpincooling.com > Overclocking & Modding > Hardware Modding

Hardware Modding Soldering, deciphering datasheets, tricks and tips.

Reply
 
Thread Tools Search this Thread Display Modes
  #1  
Old 11-28-2015, 12:23 AM
Khenglish Khenglish is offline
Ambient
 
Join Date: Mar 2012
Posts: 28
Default Switched 7/8 memory chips, then broke last chip...

After raising the memory voltage from 1.35V to 1.48V (I tried for 1.5V), and still having lackluster memory clocks on my 980m of just under 6400MHz stable, I decided to switch out the memory for the chips from a 970. The 970 gets the 980m memory, and 980m gets the 970 memory. I swapped 7 of the chips over to the 980m, but then ripped 3 pads off of the final 970 chip, leaving me with a single 6ghz Hynix chip left, holding back the memory clocks:



I spent at least 15 hours on this project to swap the memory, only to screw up at the very end making the whole time completely wasted. If anyone has any means of sourcing a Samsung K4G41325FC-HC28, please please please contact me. Most cards 960 and stronger have them, so if you have a dead card from some mod gone wrong, please let me know so I can have one of its memory chips.


Results thus far for the curious:

The 980m never had any negative affects from running with mixed memory chips. Worked fine, and overclocked the same as with the original memory. There was a time where I missed heatsink contact with the Samsung chips, and even then they still did not limit my overclock. I tested it with 1,2, 4, 6, and 7 chips switched over to the Samsung (please help me find a chip to test all 8! I want 8GHz!).

The 970 is WEIRD. I expected max stable memory clocks to drop from 8ghz to between 6 and 6.5ghz. Instead they dropped to 2ghz (yes, 500MHz * 4). Interestingly as I switched more chips to the Hynix, memory clocks improved from 2, to 2.1, to 2.2 ghz. When it goes unstable it shows characters on the screen like it's a crashed NES. I ran 2.1ghz with half the chips switched over and it was fully stable for hours of testing.

I don't get why the 970 has problems while the 980m does not, considering that they both have GM204 cores. Maybe memory timings set in the vBIOS are an issue? If so I have no idea how to edit them. Maybe having full Hynix would correct the 970's behavior. Either way, I want that damn 8th memory chip! Anyone with a means of getting me a chip, or an idea on how to get one please let me know.
Reply With Quote
  #2  
Old 11-30-2015, 04:03 PM
Bassplayer Bassplayer is offline
Sub-Zero
 
Join Date: May 2011
Posts: 112
Default

Regarding your last question... I'd definitely say it's down to memory timings. The superior Samsung chips likely ran tighter timings on the 970. Perhaps the Hynix chips are not capable of running the timings and are going into a fail-state of ~2GHz.

Now... onto the surgery. Did you take photos of the process? Did you have the IC's reballed? Did you use a hot air station? Such an amazing job... please share more!
Reply With Quote
  #3  
Old 12-02-2015, 12:49 AM
Khenglish Khenglish is offline
Ambient
 
Join Date: Mar 2012
Posts: 28
Default

Quote:
Originally Posted by Bassplayer View Post
Regarding your last question... I'd definitely say it's down to memory timings. The superior Samsung chips likely ran tighter timings on the 970. Perhaps the Hynix chips are not capable of running the timings and are going into a fail-state of ~2GHz.

Now... onto the surgery. Did you take photos of the process? Did you have the IC's reballed? Did you use a hot air station? Such an amazing job... please share more!
So I thought the chip was dead with 3 pads missing, and the first schematic I saw looked like 2 of the pads were address lines. It was the wrong schematic though (google image search cannot find any GDDR5 ballouts). Eventually I found the right schematic buried within a manufacturer's PDF (Micron I think). 2 pads were power (nice to have, but not needed), and the 3rd was some active low reset signal. I checked the 970 if this reset signal was used or not. It was shorted to GND, which is often done for unused signals. As for the power pads ripped off, one still had a sliver of its trace showing, which I managed to get a solder ball on. I figured there was a high enough chance that the chip would work, so I tried it, and it did.

As for the clock results... prior to the mod my max stable memory clocks were 6396 MHz. After they increased to 6425 MHz. Pitiful for the work that went into the mod. I still have the pcb resistor in place indicating Hynix memory. I am getting someone else's card to fix because he broke it when trying to raise the memory voltage. His has Samsung memory, so I can use that to find the resistor which sets the memory manufacturer. I am hoping that the BIOS reads this resistor and will load a more Samsung friendly memory timings table for higher clocks. Changing the resistor will likely just change what GPU-z says though. Too bad Nbitor is no longer updated to change memory timings.

I only ever took 2 pictures of the card while operating on it, and both of them were with 8 chips in place. I was always a little fearful seeing the $700 GPU with memory chips missing, so I always added them back onto the card as soon as possible. For every chip I removed, I always had its replacement reballed and clean beforehand to minimize the duration of this stressful period. I did all the work myself. I used a variable temperature and air flow rate heat gun, an IR thermometer, variable HEAT soldering iron (it says it is variable temperature, but it lies), .45mm 63/37 solder balls, GDDR5 stencil, Kimbo flux, MEK (vs IPA cleans better. safer for electronics health, but not your health), and a desoldering braid that I don't remember the brand of because I quickly cut a foot off the spool of it from a lab and walked out before anyone who might care saw me.

This is the only other pic I took. It's of the first chip I replaced: http://imgur.com/JXVyaWy

Last edited by Khenglish; 12-02-2015 at 02:04 AM.
Reply With Quote
  #4  
Old 12-02-2015, 03:53 PM
Bassplayer Bassplayer is offline
Sub-Zero
 
Join Date: May 2011
Posts: 112
Default

Wow... That's some serious work. Do any of these 970's come with Samsungs by default? You could flash the BIOS to get some more Samsung-friendly timings if so.
Reply With Quote
  #5  
Old 12-10-2015, 03:27 AM
Khenglish Khenglish is offline
Ambient
 
Join Date: Mar 2012
Posts: 28
Default

Quote:
Originally Posted by Bassplayer View Post
Wow... That's some serious work. Do any of these 970's come with Samsungs by default? You could flash the BIOS to get some more Samsung-friendly timings if so.
Most 980m cards do have Samsung chips. Unfortunately they are all 8GB cards, which probably won't play nice with my 4GB card. The flash chip looks like it won't fit in my programmer, so I'm hesitant to try in case it does brick it.

I am still trying to find the 35k resistor which sets the memory manufacturer on the board. It is very elusive and not in the logical places.
Reply With Quote
  #6  
Old 12-11-2015, 04:21 AM
Khenglish Khenglish is offline
Ambient
 
Join Date: Mar 2012
Posts: 28
Default

So timings are 100% the problem. I did find the manufacturer resistor and changed it to Samsung. I can now bench up to 7.5GHz, but my 100% stable clocks dropped to maybe 6.5ghz at best. This is a huge instability range which can only be caused by subtimings being too tight.

http://www.3dmark.com/3dm11/10651219

Oh and in the process I nicked a PCI-E lane and now the card is x4... hopefully I can fix the lane.

Update:

So the PCI-E issue was not due to the nick, but having dog fur in the mxm slot. x16 restored.

And the poor stability was due to using old liquid metal and not getting die contact on the top 1/3 of the die. 6925MHz is now stable. At 1.53V though with this memory I should be getting 8GHz, so still room to go. Timings still seem to be a problem.

Next up is finding the 970's memory manufacturer resistor and save it from 2.1GHz.

Last edited by Khenglish; 12-11-2015 at 07:06 PM.
Reply With Quote
  #7  
Old 12-17-2015, 01:54 AM
Khenglish Khenglish is offline
Ambient
 
Join Date: Mar 2012
Posts: 28
Default

So I found the 970's memory ID resistor. Changing it to Hynix did nothing and max clocks are still ~2100 MHz effective. It looks like the vBIOS lacks a Hynix timings table.

It's an MSI lightning, so it's a custom PCB and I can't flash another vendor's vBIOS for Hynix support.

So yeah... this 970 is screwed unless there's a BIOS modder who knows how to alter Maxwell timings, or there's another vendor's vBIOS that is compatible with the MSI lightning.

The 980m meanwhile is still going strong with a ~500 MHz improvement from switching the memory.

UPDATE:

Nevermind, it does have Hynix timings. There's just a memory range between 2200MHz and 5Ghz that just doesn't work. Card can now do ~6500.

Last edited by Khenglish; 12-17-2015 at 03:41 AM.
Reply With Quote
  #8  
Old 12-19-2015, 09:43 PM
Lays Lays is offline
Ambient
 
Join Date: Jul 2015
Posts: 40
Default

Quote:
Originally Posted by Khenglish View Post
So I found the 970's memory ID resistor. Changing it to Hynix did nothing and max clocks are still ~2100 MHz effective. It looks like the vBIOS lacks a Hynix timings table.

It's an MSI lightning, so it's a custom PCB and I can't flash another vendor's vBIOS for Hynix support.

So yeah... this 970 is screwed unless there's a BIOS modder who knows how to alter Maxwell timings, or there's another vendor's vBIOS that is compatible with the MSI lightning.

The 980m meanwhile is still going strong with a ~500 MHz improvement from switching the memory.

UPDATE:

Nevermind, it does have Hynix timings. There's just a memory range between 2200MHz and 5Ghz that just doesn't work. Card can now do ~6500.


You have an MSI Lightning 970?
Reply With Quote
  #9  
Old 12-22-2015, 01:08 AM
Khenglish Khenglish is offline
Ambient
 
Join Date: Mar 2012
Posts: 28
Default

Quote:
Originally Posted by Lays View Post
You have an MSI Lightning 970?
Had one. The deal was I buy it from a friend for $250. If after swapping memory the card works with the memory at at least 6ghz, he'd buy it back for $200. After fixing the memory manufacturer resistor it can now do 6.2ghz, so I sold it back.

I was surprised the memory only did 6.2ghz. I would have expected the 970 to have looser timings than the 980m for the memory to clock higher. Instead it clocked 200MHz lower at the same voltage.
Reply With Quote
  #10  
Old 12-27-2015, 01:56 AM
Lays Lays is offline
Ambient
 
Join Date: Jul 2015
Posts: 40
Default

Quote:
Originally Posted by Khenglish View Post
Had one. The deal was I buy it from a friend for $250. If after swapping memory the card works with the memory at at least 6ghz, he'd buy it back for $200. After fixing the memory manufacturer resistor it can now do 6.2ghz, so I sold it back.

I was surprised the memory only did 6.2ghz. I would have expected the 970 to have looser timings than the 980m for the memory to clock higher. Instead it clocked 200MHz lower at the same voltage.
Since when did the 970 Lightning exist? lol
Reply With Quote
Reply


Currently Active Users Viewing This Thread: 1 (0 members and 1 guests)
 
Thread Tools Search this Thread
Search this Thread:

Advanced Search
Display Modes

Posting Rules
You may not post new threads
You may not post replies
You may not post attachments
You may not edit your posts

BB code is On
Smilies are On
[IMG] code is On
HTML code is Off

Forum Jump


All times are GMT -5. The time now is 01:06 AM.

Contact Us - www.kingpincooling.com - Archive - Top
Powered by vBulletin® Version 3.8.3 Copyright ©2000 - 2017, Jelsoft Enterprises Ltd.