How to overvolt 680m memory

[Khenglish]

I've had a couple people ask me how I overvolted my 680m's memory, so here's a guide for 4GB Clevo cards:

The memory has its own VRM separate from the larger core VRM. The memory has a hardwired voltage based on a voltage divider. Default voltage is 1.35V for all 680m models. 680mx and 780m have 1.5V on their memory.

The middle of the voltage divider is always .7V. What you do is lower the resistance of the resistor between the .7V point and GND causing more current flow, making the voltage drop between Vout and .7V increase, thus raising Vout.

Annoyingly the resistor to modify is on the backside of the 680m, so you need to take the card out.

General area:

http://i.imgur.com/0zHaOVH.jpg

(the glue is to try to keep the graphite on the resistor over time. more on that later)

Exact resistor:

http://i.imgur.com/E3ZSXBr.jpg

Now doesn't everything sound nice and easy? That's what I thought at the time. It turns out that when the 680m is off that lots of stuff looks shorted together even though it really isn't. Even better is that measuring resistance one way or the other will produce different results. The only explanation I have for that is that the memory FETs are not symmetric and have enough subthreshold conduction to change what the multimeter sees.

And what looks shorted with the system off? Vout to GND! This means that you will measure the same value across Vout and .7V as .7V to GND, so you can't tell what the output voltage is. I did some math though to determine what the resistance values likely are. I measured a resistance of 11.73k one way, and 11.50k the other. Assuming that these are the only 2 resistors in parallel, that means that the .7V to GND resistor (the one we want to mod) is really 25.4k, and the other resistor is 21.79k. (the Vout to .7V resistor is immediately to the left of the cap above the .7V GND resistor). You could desolder them and measure them, but I never bothered to since I figured it'd be a great way for things to get messed up. You can measure the memory voltage with the card running from the pair of 330u caps on the front side of the card, but that requires running the card with the memory cooler off. It should be safe with the card idle. I never did it though.

I ended up pencil modding the resistor down to where I measured 10.63k. If I did my math right this means that I raised the voltage to 1.48V. I found measuring the resistance between pin 4 on the VRM to GND, which can be found on one end of most caps, to be an easier way to measure the resistance that trying to touch the ends of the tiny resistor.

Back to why I had glue on the resistor. I had issues with the pencil graphite settling and changing the resistance value over time. Twice it went up. Once it went down. The alarming time was when it went down. I closed up the laptop measuring 10.59k. The next day my GPU started crashing constantly. Measured the resistance again and it had dropped to 9.37k, which is well over 1.5V. To do this mod a proper and non-lazy way unlike me, you should probably buy small SMD resistors and piggy back one on the existing resistor. That way you know that the voltage will remain constant over time. You could also just touch a known resistor to the existing one to help figure out what the resistance and thus output voltage really is. The glue actually took the graphite off the resistor, so don't try it.

I did this mod to a 4GB clevo card. I do not know if the dell and MSI cards have the same memory voltage circuit. Maybe they do though and cards like the 675mx do too. I don't know.

Before the mod I was running memory at 4700. After this I run it around 5100. I also replaced the cooling pads with copper shims. Before that my memory only ran in the low 4600s. A problem is though the shims are thicker than the pads, and I overtightened the screws and warped the card... don't do that. It looks like I might be getting a 780m because I did that.

I can't guarantee that you won't blow up your card doing this. Don't blame me or make Dell or whoever pay for it if you do.

[Clyde]

I recommend that you read this: http://forum.techinferno.com/general-notebook-discussions/1847-nvidia-kepler-vbios-mods-overclocking-editions-modified-clocks-voltage-tweaks-104.html#post35831

[Khenglish]

I recommend that you read this: http://forum.techinferno.com/general-notebook-discussions/1847-nvidia-kepler-vbios-mods-overclocking-editions-modified-clocks-voltage-tweaks-104.html#post35831

Ha nice memory swap.

And I see on your vmem mod that you raised the .7V-Vout resistance instead of lowering the .7V-GND resistance. Why didn't you do the easier mod to lower the .7V-GND?

Also there is an extra VRM on the other end of the PCB close to the GPU die. Do you know what it controls? Vcore is at the top of the card so it's not that. I'm guessing either memory controller voltage or PCI-E voltage. Memory controller voltage would explain your low overclock despite replacing the memory and raising the voltage. It would also explain why the 780m's memory clocks so much higher that 680m memory. I would certainly play with this if my card was not such a PITA to get working right due to the warp I put in it.

Also I had memory heat issues over 1.5V. You might clock higher at a more conservative voltage like 1.45V.

[Clyde]

I used this method because it was easier for me to add a little resistor soldered on top of existing than replaced it. Moreover, it seems to me that other solutions did not provide a stable voltage. Never mind. My conclusion was that the 3GHz memory operation mode require a different memory chip pin out. At 1.5V the temperature rise became significant (+5 C) but for me did not cause throttling and exceeding Tjmax.

The second voltage regulator (centrally located) is of course the core voltage controller.

[Khenglish]

I used this method because it was easier for me to add a little resistor soldered on top of existing than replaced it. Moreover, it seems to me that other solutions did not provide a stable voltage. Never mind. My conclusion was that the 3GHz memory operation mode require a different memory chip pin out. At 1.5V the temperature rise became significant (+5 C) but for me did not cause throttling and exceeding Tjmax.

The second voltage regulator (centrally located) is of course the core voltage controller.

I am referring to the VRM in green: http://i.imgur.com/lDZqo1L.jpg

[Clyde]

Unfortunately I have no idea what it is.

Oddly enough the GTX780M use the same memory chips like GTX680M (FC04 not FC03), but changed the inductor (1R0 for R56) and capacitors (330uF for 470uF).

Maybe 3GHz memory operation mode it's not a matter of memory chip pin out but the appropriate filtering?

[Khenglish]

That's an enormous clock gain for just a change in filtering. But memory voltage regulation is just a single phase, so maybe a 42% capacitance gain does help a lot.

Notice that all core caps went up to 470u as well. The core inductors did not change.

Do you know how power phase based voltage regulation works for DC? I've only ever done it with AC. I'm guessing that since you end up making DC that the load capacitance does not need to have a certain ratio with inductance? IDK

[Clyde]

Changing the inductor (1R0 for R56) and capacitors (330uF for 470uF) work. Though with some artifacts. But it is rather the fault of poor quality components used and soldering. Above 2900MHz are too many artifacts and I fear for the card.

http://www.3dmark.com/3dm11/6675176

For DC the inductor behaves as a resistor. But in this case we are doing with pulsating current.

Here is a voltage regulator datasheet

[Khenglish]

Nice! Unmodded 680m has horrendous voltage ripple then with such good results. Have you thought of going even higher on the capacitance?

And did you replace the core caps as well?

[Clyde]

I'll try to going higher.

I did not have the parts for changes all capacitors. But I already thought about dismantling 7970M for parts.

[Khenglish]

I'll try to going higher.

I did not have the parts for changes all capacitors. But I already thought about dismantling 7970M for parts.

Well I had 30 470u 30V caps arrive today. I think I'm going to go higher capacitance than you

30V is overkill, but I got 30V so I could also put caps between Vin.

[Khenglish]

Well I added a 470u cap to the 2 existing 330u. Made a slight improvement, but nothing like what you got clyde. It looks like it's the inductor swap that made the big difference.

I wonder why a 1uH inductor is used in the 680m, but only a .56u in the 780m. I feel like the lower inductance inductor would be cheaper, so why was the higher inductance ever used on the 680m?

[Clyde]

IMHO, You can not add freely capacitors because you limit voltage in the circuit (as did Clevo in HD7970M). Inductor R56 (with lower inductance than 1R0) pass higher current, which we have slightly suppress-filter with higher capacitors, but not with twice the value.

Edited May 8, 2016 by Clyde
Fixed evident mistake: it was current, should be voltage

[Clyde]

NVIDIA GeForce GTX 680M video card benchmark result - Intel Core i7-2920XM Processor,CLEVO P170HMx score: P8985 3DMarks

Result GTX680M after 1.5V+inductor+capasitors+memory chip mod (1.05V vbios). No artifacts up to 2700MHz.

[Khenglish]

NVIDIA GeForce GTX 680M video card benchmark result - Intel Core i7-2920XM Processor,CLEVO P170HMx score: P8985 3DMarks

Result GTX680M after 1.5V+inductor+capasitors+memory chip mod (1.05V vbios). No artifacts up to 2700MHz.

Nice improvement. Still pretty short of the 780m though. Only variable left that I can think of is memory controller voltage, but I feel that nvidia might just directly use the memory voltage for that.

By adding a 470u cap I went from 4860 to 4937 on 1.455V (measured while running) fully stable. Will add another cap or 2 and possibly inductor swap soon, as well as raise the voltage a bit more. I had to back the voltage down since it kept drifting too high on me with pencil mods. Plan now is to piggyback a 100k resistor between .7V and GND. I could get 5100 stable at higher voltage before adding the cap, so I expect 5200 or so eventually. What clocks did your memory do before the filtering change?

[Clyde]

http://forum.techinferno.com/general-notebook-discussions/1847-nvidia-kepler-vbios-mods-overclocking-editions-modified-clocks-voltage-tweaks-104.html#post35892

The inductor-capacitors mod works well. 2600MHz vs 2700MHz without any artifacts for GTX680M and K5KM.

I need a closer look at the GTX780M to go higher.

[Khenglish]

Just added another 660u to the memory. Should have around as low ripple as you now. 5012 made artifacts after a few minutes. 4976 has been running fine for quite a while now, but I am expecting it to be unstable, with the next step down being fully stable.

Still "only" running the memory at 1.45V

Will probably start on the core caps tomorrow. 1032 is almost stable so this should do the trick and maybe even get me 1045.

[tmash]

hey thanks to your quick guide, it gave me an idea where to start with memory OC/OV on my ancient 770m lol

I was able to go from ~2000 mhz memory to ~3200 with samsung memory @ 1.51v (stock 1.35v), used 4B pencil as i noticed its less grainy and thus dust and airflow doesnt wipe it out easily i guess, though still havent tested it >1mo.

 

I also had to raise vbios tdp limit since the core started to throttle after memory overvoltage, and now its perfectly stable.

 

http://i.imgur.com/3iRJwed.jpg

http://i.imgur.com/eA3MHcc.jpg

http://i.imgur.com/V7T8MoL.jpg

 

 

offtopic, did you actually remove the backside memory from your 980m? " MSI 980m 4GB (no pointless backside memory wasting power) " thats awesome for OCing...

Edited April 30, 2016 by tmash
imgur problem

[Khenglish]

@tmash

 

Nice clocks. Kepler has a good IMC, but for reasons unknown memory never clocked well on the 600m series cards. Good to see the 770m was better.

 

You can't remove memory from a card. On the 16 chip cards 16 bits of memory interconnect are wired to each chip. If you pull a chip you lose 16 of 256 bits. 8 chip cards meanwhile have 32 bits wired to each chip.