This is a guide for people who aren’t shy with a soldering iron to really unleash the Kepler-based GTX 680. It is not necessarily for the faint of heart!
Everything here is reproduced with permission from its original author, legendary overclocker/modder (and EVGA engineer) TiN. The original post can be seen here. Save for formatting and editing (i.e. spelling/grammar), everything here is his.
Welcome! During a few days of overclocking EVGA GTX 680 cards with Vince, I did some basic modifications which I can share with fellow overclockers who still remember how to hold a soldering iron in their hand.
Most of the modifications are common between different GPU’s, so those who bench with extreme cooling should be very familiar with measuring resistance, calculating proper VR (variable resistance) value, checking for any shorts, finding ground and using DMM (digital multi-meter). So I will not go into crazy details about every small thing possible, but will give some notes and ideas how to deal with the GTX 680 to reach its maximum overclock.
Those who are interested in specs, card features, and comparison analysis can easily find that info in reviews and articles about new-gen 28nm NVIDIA GPU (Editor’s Note: We have an EVGA card in-house and in testing to review for you!), so here we focus straight on real business with soldering.
First let’s take a brief look at the VRM section of the card.
We see a staggered MiniFit JR (which is uncommon in PC’s), with two 6-pin power inputs for +12 V, and empty space for a second regular 6-pin connector. The GTX 680 has a lower TDP than monsters like GTX 480 or 580, so 6-pin is usually enough for most overclocking conditions, even on water or subzero.
The memory VRM is similar to one used since GTX 480, two phase with discrete FETs.
The main VRM is made to support 5 phases, but final reference boards were de-featured to 4 phases, which is enough for air cooling.
The GTX 680 also uses input power sensing circuitry to calculate input power and adjust it’s speed and voltage to meet the desired performance and power limits.
On the back we see a new idea of having the main GPU PWM as a soldered-on module with related passives. This was probably used for debugging and testing the VRM with different PWM controllers. You can also see one more sensor for input power.
On the bottom part, we can see six empty footprints for tantalum capacitors for better decoupling of the GPU power plane. Luckily for external power users (specially EVGA EPOWER owners), this GTX 680 design is really friendly modification-wise. The GPU power copper plane is located on the outside surface of the PCB, so you can easily shhrrr-shhrrr protective solder mask and solder crazy AWG10 wiring directly, without risk of breaking capacitor pads or tiny islands on the PCB. Also GTX 680 has two PLL LDO’s; each of those are similar to ones on GTX 580 and 480.
From my brief experience, Kepler does not need crazy voltage and current to run fast clock speeds. We were able to hit 1350-1400 MHz on air cooling with just VID mod to 1.3 V on GPU. So I will describe modification to the onboard power first, and then we will talk about external power module using an EVGA EPOWER 8-phase or 10-phase retail board.
GPU Voltage Modification
The main PWM for this GPU is a Richtek RT8802A. Luckily for us, the full datasheet on this PWM controller is available publicly on the official vendor web-site. It’s quite simple 5-phase controller with VID control. It has one drawback though – OVP is not adjustable and fixed internally to offset from VID setting, so a good way to prevent triggering protection during heavy loading/overvolting is to use VID mod.
I took multi-position DIP switch, and connected one side of its poles to ground, and another side to VID pins. To be able to toggle VID, I used 4.7 kΩ pull-up resistors to +5 V, so when the switch position is open, I get logical “1” on the VID signal. If the switch position is shorted, I get a logical “0”. The task is then simplified to just finding needed voltage in the VRD11 table in datasheet above and setting the proper levels for the switches.
NVIDIA limited the voltage range to max 1.21 V by hardware, having MSB VID6 and VID7 tied to 1 and 0 by hardware (direct connect to power and ground by trace). So if you want VID more than 1.21 V, then need to cut VID6 connection to VCC and manually connect to GND. Because VID7 is already tied to GND, I just shorted those two pins together with piece of thin wire.
So total modification looks like this:
This worked OK, but when I tried to push 1.5 V on one of the cards, a FET in one of phases failed with a bad smell, so be careful with going high volts. As usual – all modifications are provided AS IS and use at your own risk.
Also, it doesn’t hurt to populate the back of the card with solid caps which was supported by design on the reference cards initially.
Memory Voltage Modification
Memory is simple to modify. It has a Richtek 2-phase PWM without any advanced controls, so you just need the usual VR’s for feedback adjustment (marked red in the photo above). If you want to try to push the memory high, you can try playing with PWM Frequency and memory OCP tuning. There are also three empty footprints for tantalum caps, so I usually solder a pair of 330 uF 4 V caps there.
PLL voltage modification
This is a simple LDO with feedback adjustment. Each one powers separate circuitry in the GPU, but at default they are set to provide very close voltage at about 1.05 V. So no special magic here.
As was with the majority of NVIDIA cards since G80, there is always some thermal shutdown circuitry present to disable power when the GPU temperature reaches an overheating threshold level. It’s good thing for safety to prevent GPU burnout if you forgot to put your fan, but not so good for subzero overclocking. To prevent power shutdown when temperature freezes below -70 °C, we need to disable over-temperature protection.
This can be done by removing the tiny resistor shown on photo above. If you want to revert it back – you can just put a short there, that resistor is 0 Ω anyway on reference cards.
Power Limiter Modification
To fool the power limitation set by NVIDIA for Kepler, we can reduce the resistance of current shunts. The reference card uses three 5 mΩ resistors to measure input power. If we short them or solder on an extra current shunt in parallel, we can decrease the power draw values reported to GPU. This is a must if you want to avoid being capped by power and TDP.
So that’s it; not very complicated even for beginners. (Editor’s Note: HAH!) It is always best practice to check GPU, memory and PLL resistance before powering up the card, just to make sure nothing has been shorted by accident. Good resistance for GPU is in the range 1.5-10 Ω, for memory 100-140 Ω, for PLL a hundred or so Ω.
For those who are tired of removing shims on other GPU solutions, there is good news for GTX 680 – there is no need to remove anything. The protective metal shim surface is thinner than the GPU die, so any KPC TEK-9 LN2 containers will fit perfectly on GPUs using the standard mounting.
Crazy LN2 Overclockers Part
Toggling VID each time and risking to get OVP/OCP with the stock VRM is not always the best option, so if someone really wants to push these cards to their limits and beyond, then using external power is a good choice. The GTX 680 draws so little power compared to an overclocked GTX 580 that using wires to hook up an extra power board makes practical sense. I measured ~270-300 W power draw when the card was running 3Dmark11 P at 1400MHz with 1.4 V, which is a really good result for its performance.
The power plane for GPU voltage is placed on outer surface of the PCB, which makes attachment of EPOWER really easy, taking ~30 minutes. I think pics of the actual process will say it better than thousands of words.
To be able to run the card with EPOWER, only one modification to the stock VRM is needed: cut the single trace marked below red.
Make sure that you connect EPOWER by the shortest wires possible, and use as many wires as possible.
I was a little bit busy with a lot of other things, so I did not go crazy with wiring, just a bunch of 10 AWG grounds and direct PCB-PCB connection for power.
k|ngp|n and I tested about six GTX 680 cards, four of which we tried with EPOWER and got ~1600-1700 MHz without problems with only 1.5-1.6 V set by EVBOT on EPOWER, with only -50 to -60 °C container temperature. We pushed to the max tonight and got 1900 MHz with some little voltage push and -150 to -160 °C. To be honest, that was one of best OC sessions ever with well deserved single card WR in 3Dmark11 P. We would not stop at this and will push for more in the coming days. Stay tuned!
Editor’s Note: You can check out TiN & k|ngp|n’s single GPU world record 3DMark 11 Performance run here at HWBot. We’ll be back with a review of EVGA’s GTX680 when we’re finished with it!