Originally Posted by geoffrey

Introduction
During my research for higher performance I've come across many different electronic schematics. Most of you know the existence of usual voltage mods and current protection tricks, let me tell you that for the NVIDIA 8800GTS we are having quite a hard time finding out what is holding back our overvolted cards. From what it seems, a over current protection kicks in whenever the GPU is heavy loaded (i.e. when high current is flowing through the GPU). The screen freezes and the PWM voltage module reads 0V. In order to know how we can fix this we must first understand how the PWM module exactly works and how the current sense network can be modified. For GeForce 8800 cards NVIDIA has chosen for the Primarion PX3540 PWM controller, here is a small schematic:



For most people this won't look very clear, but let me tell you that this controller isn't that much different compared to many others I've seen, it only has more options and more PWM phases... Let's us first have a talk about how this type of voltage converter actually works.

+12V

Modern PC's have 3 common voltage levels available which feed most of its components, 3.3 Volt, 5 Volt and 12 Volt. Back in the old days the 3,3 and 5V line were used to feed the processors and DRAM modules, though over time the power usage of some modules become so high that is was in fact more attractive to use the 12V line as feed for CPU's and GPU's, it is much easier to supply high power at those voltage levels. Though, 12V is far beyond the maximum supply voltage for the high switching transistors found in modern microprocessors, therefore we needed a circuit which converted the 12V DC voltage coming from the PSU into the 1,5 Volt which is needed to feed the CPU. Such circuits are called DC-DC converters because they convert a certain DC voltage into a lower level DC voltage. Have a look at the following picture:



The above picture shows a 12V DC power supply, a switch, a coil, a diode and a certain load which could be a lamp for example. The lamp will only light up whenever the switch is closed, that makes perfect sense to anyone. Now, if we rapidly set the switch to open and closes states, then the lamp will light on/off too, and if the on time is on par with the off time then the average voltage would be 12/2= 6 volt. If we would measure the voltage over the lamp we would see a square wave, something like this:

_|-|_|-|_|-|_

Now, in order to no longer switch on/off the lamp you could add a powerful coil in line with the lamp. Properties of coils tells us that they will work against the current flowing through them, therefore you will get a sawtooth kind of current flowing through your lamp, it will no longer have to switch on/off, we have created a kind of DC voltage though there is still lot of AC characteristic inside it. If you use your scope you will see something like this over you lamp:

/\/\/\/\/\/\

What the above picture doesn't show is the capacitors which are needed to filter the ripple out. After the coil one should add few capacitors, depending on their size the voltage measured over the lamp will become more or less a stable 6V voltage line. The caps aren't placed in line with the lamp but rather placed over the lamp, in parallel if that is easier to understand. You now will now have decent supply voltage which also could be used to feed CPU's, DRAM modules, ...
In real life that switch isn't set on/off twice a second, but rather few hundred thousand times per second. We won't dive further into the whys here, you must only know that for this rapid switch we can't use human power to make the switch flip levels, nobody is that fast. We could use mechanical switches like a relais, though mechanical switches tend to have a very limited life cycle, after a month a relais could be burned for example. Mosfets on the other hand can be turned on/off almost unendless and are perfect for such operations, in computers we find N-type mosfets to switch the 12V line on and off.


GPU nominal supply voltage

In the above example we successfully converted 12V DC into 6V DC, but how can we build a DC-DC converter which is capable of delivering 1,5V on it's output? It's all in the name in fact... PWM, Pulse Width Modulation. With this technique we are going to change the on and off times of the switch/mosfet.
In our above example we've switched the mosfet on for 0,5s, and switched it back of for 0,5s. The total wave length would be 1s, and the on state would be 50% of that total wave length, we speak of a duty cycle of 50%. Now, if we change the duty cycle to only 25%, then the mosfet would be switched off for 75%, and the average voltage would be only 25% of 12V -> 3V. If you understand well, the duty cycle defines the output voltage, the lower the duty cycle, the lower the average voltage, and the lesser light your lamp will create.

Originally Posted by geoffrey

PWM modules explained

To give you an idea of current PWM modules I've added a simplified connectivity drawing of the NX2415 PWM controller:



You can see the 12V line coming from you PSU at the top right of the picture above. This voltage feeds dual mosfets who will do their own switching algorithm. If mosfet 1 is "on" then mosfet 2 will be "off" and so on. The mosfet have a special connection pin called the 'gate'. Applying a voltage at that gate will make the mosfet more or less resistance free, it will let the 12V voltage through. The pins on the PX2415 controller are therefore called gate drivers in most cases, because they will actually drive the mosfets on and off. Again you can see the coil which creates a sawtooth like voltage line which is being further rectified to a usable output voltage via big capacitors.
In order to check whether or not the output voltage is good, each controller has a feedback loop which compares the circuit its output voltage against an internal set voltage. We won't dive any further into this, my main goal today was to focus on the over current protection circuit. New current sensing techniques learned us to add a RC-circuit (R from resistor, C from capacitor) parallel over the power coil, in this construction the voltage over the capacitor will increase in a linear way with the current flowing through the inductor. If we read this voltage back into our controller this means that we are actually reading in the amount of current flowing through our power circuit. Now look at the lower left side of the controller schematics... You'll notice a resistor divider which task is to set a certain voltage level at the OCP pin of the NX2415 controller. A resistor divider works as following: when you have a 5V DC supply, and you connect two 1k Ohm resistors in series between the + and - connectors of that supply, then you will get around 2,5V in between the resistors and ground. Depending on the size of those resistor, you will measure more or less voltage at that same point. If the first one has a resistance of 2k Ohm instead of 1k Ohm, then you will read less voltage in between the resistors and ground because more voltage has 'dropped' over the first resistor.



That's how it works, this way we can change the reference voltage created inside controller to anything useful worth comparing with.

Over current protection

Like I said, the voltage over the Capacitor of the RC network is being read into the controller, inside the controller this signal is compared with the voltage level which we made at the OCP pin. The higher the current flowing through the inductor, the higher the voltage over the sensing capacitor (Cs), and the closer Cs will come to the voltage level applied at OCP, the over current trip point. After OCP has been tripped the controller will shutdown itself in order to protect itself against any damage and your pc will most likely freeze or reboot. We can work around this be changing the resistor values at the OCP pin, though we don't have any specific details about pin connections and functionality with the Primarion PX3540 PWM controller, we are forced to find out how the RC-network works so that we can manipulate the signal that is being read by the controller.

Originally Posted by jmke

clearly no CPU bottleneck, pretty good numbers
would drop AA and go for high res

Originally Posted by jmke

I can recommend the latest version of Rivatuner which also support unlinked clocking;