Tweaking the fastest AGP Based System - Part 1

Overclocking/ by massman @ 2008-03-25

While the rest of the mainstream world moves on to PCI-Express and DDR3, we take a step back to the previous generation hardware. Our aim is to build a scalable AGP benchmark platform with sole purpose: breaking overclocking records. Fueled by HWBot competition we squeeze the best AGP performance from a budget-friendly Asrock motherboard. Which memory goes the fastest? DDR1 or DDR2? Let us find out.

  • prev
  • next

Voltage modifications


In the spirit of overclocking we are attempting to improve on the original, taking the standard and developing it into something greater. As such the hardware will always have limiting factors in this quest for perfection, none more so more as voltage limitations.

We can consider two ways to improve an overclock, run the silicon colder, increase the voltage the silicon runs at, or both (usually necessary!). As the silicon chips are in essence vast arrays of transistors there are associated losses with these devices switching, this is where the heat comes from and heat is the enemy of overclocking. Increased heat causes a number of problems in microchips including increase track resistance and thermal noise. Understandably increasing the voltage will add to the losses, as will increasing the switching frequency (or overclock); therefore it is always wise to upgrade your motherboard’s cooling when increasing the voltages of the components.

The AsRock 4CoreDual-VSTA is not the most forthcoming motherboard in terms of voltage options. In fact there are only two real adjustments, memory voltage and AGP voltage which have all the adjustment of “low”, “normal” and “high” – hardly precise…

In our pursuit of reaching the overclocking limits of this board a few extra voltage options would be needed:
  • CPU Core Voltage
  • CPU VTT Voltage
  • Northbridge Voltage
  • Improved AGP Voltage Control
  • Improved Memory Voltage Control
  • Readpoints For All the Listed Voltages Above
So let’s take a look at the board.

CPU Voltage modification

1) CPU Voltage read point

Easiest voltage read point to get is the CPU core voltage, simply look for an empty capacitor pad near the socket area and measure from ground to the positive leg, marked on the board in white.

Madshrimps (c)

2) CPU Voltage modification

The CPU voltage is controlled by a ST Microelectronics L6714 4 phase PWM controller (.pdf). Normally with a vmod the idea is to alter the feedback bias resistor network by running a suitably sized variable resistor in parallel with the lower (grounded) resistor. The ST controller however as a rather clever little addition to it, an OFFSET input at pin 23 that can be used to introduce an offset to the feedback voltage resulting in an elevated voltage output. A 20Kohm variable resistor in place of the shunt resistor provides a suitable scaling for the vcore voltage.

Madshrimps (c)

3) CPU Voltage Vdroop

Vdroop is a recently much debated value. Previously considered the scourge of CPU overclocking it is now slowly becoming accepted as a necessary evil for longevity in a CPU. The AsRock has as an excessive droop, especially at elevated voltages. While this board is certified for quad core use I’m reluctant to say it’ll overclock a quad core with any great success. My personal experience with an Intel X6800 running at 1.6volts was a voltage droop under load of nearly 60mV. This can also be adjusted by adjustment of the external droop network, an example shown here from a user called gustep12:

Madshrimps (c)

Memory voltage modifications

1) Memory voltage read point

Similarly for the Memory voltage we find an empty capacitor pad at the top of the motherboard that will provide a suitable read point.

Madshrimps (c)

2) Memory voltage modification

The memory is controlled via a Richtek RT9202 synchronous buck PWM controller (.pdf) where we alter the feedback network directly via pin 6 using a 50Kohm variable resistor to ground.

Madshrimps (c)

CPU VTT and AGP/MCH modification

1) Cpu vtt read point

The CPU VTT read point is not so clear, but can be found close to the northbridge where there are two MOSFETs located side by side. The CPU VTT voltage can be read off the drain on the MOSFET as shown.

Madshrimps (c)

2) Agp/mch read point

When hunting for the AGP and Northbridge voltages it became apparent that the two were In fact sourced from the same point, a MOSFET located between the AGP and PCI-E x16 slots. An empty capacitor pad here allows for measurement of the AGP/NB voltage.

Madshrimps (c)

3) Cpu vtt and agp/mch modifications

Voltage control for the CPU VTT and VAGP/VNB can be found near the top of the board beside the DDR/DDR2 sockets. Again the modification involves altering the feedback network, this time by directly altering the feedback network of the opamp. A 10Kohm variable resistor is used of the CPU VTT while a 20Kohm is more suitable of the VAGP/VNB feedback network.

Madshrimps (c)

As the AGP and NB voltages are interconnected we are forced to increase the AGP voltage in order to raise the northbridge voltage, in order to increase our FSB/memory speeds. Normally the AGP voltage is around 1.5volts with a usual limit at about 1.7volts. Those of you who remember overclocking on AGP will be no doubt aware that while increasing the AGP voltage did not bring much in the way of improved overclockability, increasing it beyond 1.8volts did run the risk of killing certain cards. Therefore we have a limitation on the northbridge voltage when using an AGP based graphics card.

Taking voltmods even further

Above mentioned issues highlights one of several limitations of the 4coredual-VSTA motherboards design. Due to cost cutting measures we find that circuits have been cut down to their “bare minimum”. Notice that a lot of the capacitor areas are empty, the AGP and northbridge are supplied from the same simple linear MOSFET power supply circuit, the memory power supply for DDR and DDR2 is sourced for the same circuit – as the JEDEC standard for DDR is 2.5 volts I am curious to know how the AsRock engineers are able to boot the motherboard in such a way that it can tell the difference between the memory types and set the voltage accordingly, or is there a situation where DDR2 memory is supplied momentarily with DDR voltages before being reset? Further testing of this power circuit also showed that it wasn’t particularly suitable for high current requirements either, tending to droop under load when using TCCD and Micron “fatbody” based RAM. Testing with Winbond BH-5 based memory seemed to be more suitable, even at elevated voltages of 3.4-3.5volts under load.

To improve upon these design limitations we are forced to take some seemly drastic measures, and start to replace pre-existing components.

For the CPU voltage the best line of action is to upgrade the capacitors with lower ESR (ebullient series resistance) ones to allow the current to flow with less obstruction. My personal choice is Samxon GC and GD series ultra low ESR capacitors that come with a good reputation behind then. You can also at this stage increase the capacitance of the circuit within reason to aid in current capacity. I choose to replace all the capacitors on my board in am attempt to improve the overclockability of various components.

Madshrimps (c)

The AGP/NB supply circuit is the Achilles’ heel of the 4CoreDual-VSTA. The MOSFET itself is rather under rated for its duties and has a very low current limit of around 25A. If you were to increase the output voltage to around 1.75volts and attempt to run an AGP card as well you would soon find that the MOSFET literally burns up taking the board with. If you do want to run at higher voltages on the AGP/NB it is imperative that you use a heatsink on this MOSFET.

Madshrimps (c)

While I had removed the surrounding capacitors I decided to unsolder this MOSFET as well and replace it with an upgraded model capable of a far larger current. For added protection I installed a small heatsink on top.

Madshrimps (c)

Finally the memory power circuit must be able to provide an adequate voltage scale between 1.8 to nearly 4volts in order to encompass DDR2 and DDR in all its guises. To achieve this I have removed the secondary inductor, effectively disconnecting the existing power circuit and attached my own linear design directly to the board. The simpler design has a high current rating and features a variable output from 1.4volts to 4.3volts. It should be noted however that there is no means to automatically switch the voltage scale between DDR and DDR2 so caution must be exercised.

Madshrimps (c)

Further improvements could be brought around by replacing the AGP/NB power circuit with a better off board design, separating the AGP and NB power lines, or improving the CPU power circuit further with upgraded inductors.
  • prev
  • next