Today NVIDIA GeForce3 chip boasts the highest performance of all the 3D gaming chips. In fact, its performance should be enough to ensure playable conditions for the already existing and some upcoming games.

However, human beings constantly strive for something better, and 3D games make no exception. Gamers always feel better when they can set higher resolution or enabled anti-aliasing, anisotropic filtering and so on. In case the user enables all these functions or does his best to obtain exorbitant fps rate at the highest resolution possible, even GeForce3 may turn out too slow.

A wise way out is to overclock the graphics card. Guided by our favorable experience in modifying and overclocking various graphics cards based on NVIDIA GeForce2 GTS, GeForce2 MX and PowerVR/STM Kyro II, we went further than ordinary overclocking and decided to modify graphics cards based on NVIDIA GeForce3. At our disposal we had ASUS V8200 and VisionTek GeForce3 that is why we put all effort into sweat them to the utmost.

Extreme Overclocking

Let's start with VisionTek GeForce3 card following NVIDIA's reference-design pretty precisely:

This graphics card is equipped with SC1175CSW voltage regulators from SEMTECH. Each of them represents a combination of two versatile 2-phase, synchronous, voltage mode PWM controllers intended for voltage regulation on graphics cards and peripheral devices. Two independent controllers, these chips are made of, can work in two distinct modes: as dual switchers or in a current sharing mode.

Current sharing mode serves to provide GeForce3 with stabilized core voltage. In current sharing configuration the SC1175CSW regulator can produce a single output voltage from two separate voltage sources (which can be different voltage levels) while maintaining current sharing between the channels. It implies that the Vout of the first controller is determined by the proportion between R1 and R12 resistor ohmages (see the typical application circuit scheme). The second controller takes the output voltage of the first controller as a reference. The outputs of both voltage regulator controllers work in parallel and that's where the term "current sharing" comes from, actually.

Typical application circuit scheme in this mode (click to enlarge):

On VisionTek GeForce3 the ohmage of R12 and R1 resistors (R832 and R810 on the card) makes 130Ohm and 770Ohm respectively. The Vout in this case is calculated with the following formula: Vout = 1.25x(1 + R12/R1) and makes 1.45V. Having shunted R1 with a 510Ohm resistor by soldering it right to the pins No 18 and 20 of SC1175CSW chip, we got the resulting ohmage of R1' equal to 303Ohm and the Vout became 1.78V:

Thanks to the second SC1175CSW microchip installed on the card to stabilize the graphics memory voltage, we got the opportunity to increase the graphics memory voltage as well.

In dual switcher configuration, two voltage feedback paths are provided for independent control of separate outputs. Here is a typical circuit scheme for the microchip with two independently working controllers (click to enlarge):

According to the specs, the DDR SDRAM chips used on GeForce3 are powered with two voltages: 3.3V for all the internal circuits and 2.5V for the input-output buffers. One of SC1175CSW controllers is exactly the one to power the internal circuits of the graphics memory chips, while the second one is responsible for powering the input-output buffers. The Vouts for both controllers are determined by the proportion between the resistor ohmages: R15, R14 for the controller providing 3.3V and R13, R11 for the one providing 2.5V.

The card resistors (R822, R824 - for 3.3V and R825, R829 - for 2.5V) with the ohmage of 170, 100, 110 and 100Ohm correspondingly provide 3.37V and 2.63V on the output. In other words, the memory voltages are somewhat higher than the nominal ones already. To make them another bit higher, we shunted R14 and R11 with 820Ohm resistors by soldering them to the pins No 3, 18 and 20, as it is shown on the photo:

    

The "new" ohmage of R14 and R11 resistors equals to 89Ohm for each of them and the Vouts of the voltage regulator controllers (calculated by the same formula) has turned to 3.64V and 2.8V respectively.

Higher Vcore and memory voltage produce a serious strain on the chips and lead to greater heat dissipation, so we had to take care of some more efficient cooling solution. We installed specially designed heatsinks from Thermaltake on the memory chips and equipped the core with a heatsink borrowed from a standard ND3 cooler. For better core cooling, we fastened a really helpful Thermaltake Volcano fan to the heatsink:

But even this advanced cooling system didn't satisfy us and we decided to resort to Peltier element. This highly efficient cooling system allowed GeForce3 to retain stability at 280-290MHz core frequency. However, when some condensate dripped from the element to the AGP slot, we held our breath with fear and only having made sure that that neither the board nor the graphics card were damaged we sighed with relief. From then on we made up our mind to avoid experiments like that and to give up using Peltier element. Without it the chip got overheated at 270MHz already and the system hung. The trouble is that GeForce3 chips surface is a concave and it's hard to dissipate the heat from this concave metal spot. We found a straight-forward solution and simply planished it to make flatter:

When the chip surface became flat, we planished the heatsink's foot too and enjoyed stable work of the card at high clock frequencies.

The modified VisionTek GeForce3 card worked well at 270MHz core and 600MHz (300MHz DDR) memory frequency. Let us remind you that all the NVIDIA GeForce3 graphics cards feature 3.8ns DDR SRDAM chips, therefore the achieved memory frequency is a really splendid result.

The PCB of ASUS V8200 graphics card differs a bit from the reference, however, it features the same core and memory voltage regulators. It is only the location of the elements that is different:

The resistors responsible for the Vcore (R134 and R138 on the card) are located on the rear side of the PCB. Their ohmage is 140 and 780Ohm respectively. NVIDIA GeForce3 chip on this graphics card receives 1.48V. We installed an additional 820Ohm resistor parallel to R1 (R138 on the card) and obtained "new" R1 ohmage - 400Ohm:

The Vout of the regulator netted 1.68V. Due to high quality cooling on ASUS V8200 the core didn't warm up as much as it did on VisionTek graphics card in case of increased Vcore, so we risked not to summon up any extra core cooing.

As for the memory voltage, we raised it in the same manner as on VisionTek GeForce3 (with the help of two 820Ohm resistors):

    

This modification enabled ASUS V8200 to work at 260MHz core and 580MHz (290MHz DDR) memory frequency.

Testbed

For the investigation we prepared the following testbed:

• AMD Athlon 1.2GHz CPU (133MHz FSB);
• ABIT KT7A (VIA KT133A based) mainboard;
• 256MB NCP PC133 SDRAM;
• Fujitsu MPE3084AE 8.4GB HDD.

Software:

• Windows 98 SE build 4.10.2222 A;
• DirectX8.1a;
• Quake3 Arena v1.27g.

The graphics cards were tested with Detonator 12.41 driver. To adjust the graphics core and memory clock frequencies, we used RivaTuner utility.

Testing Methods

First of all we would like to describe the conditions these graphics cards were tested in. Our testbed had no housing, so there were no problems with heat dissipation, which are so typical for a closed housing. To cool the graphics memory heatsinks, we put a conventional 80mm fan some 5-10 centimeters away from the graphics card. The room temperature remained within 20-23 degrees.

In our previous review devoted to a NVIDIA GeForce3 graphics card from VisionTek we have already handled its performance in different games in an overclocked state. Here we'll only check the performance in Quake3 Arena.

We tested VisionTek GeForce3 graphics card, since it appeared capable of working at higher frequencies and proved more stable than ASUS V8200. In order to obtain more exact figures, we ran each test 5 times and chose the maximum result out of the five obtained:

Performance in Quake3 Arena

For VisionTek GeForce3 tested in Quake3 Arena we used the following settings: for 16bit color modes we took 16bit textures and for 32bit modes the textures quality was set to 32bit. Texture and details quality was set to the maximum. Tri-linear filtering was enabled, while texture compression was disabled.

We tested the card in different resolutions at the following frequencies:

• nominal core and memory frequencies;
• memory overclocked to its max and nominal core frequency;
• core overclocked to its max and nominal memory frequency;
• core and memory overclocked to their maximum.

In order to better illustrate the performance gain resulting from extreme overclocking, we draw several graphs depicting the gain (in percents) relative to what we had at nominal frequencies:

The graphs show that at 800x600 and 1024x768 the performance growth is constrained by the CPU and the system as a whole reaching nothing more than 10% in case of extreme overclocking, though at higher resolutions it goes up to more than 30%.

In 16bit color mode core overclocking turns out more fruitful than graphics memory overclocking.

In 32bit mode they both yield similar performance gain, but at higher resolutions memory overclocking proves more efficient, whereas with an overclocked core the gain freezes at 10%.

It's an interesting fact that when the core is overclocked by 35% and the memory - by 30.43%, at the highest resolution the performance grows by 32% both in 16bit and 32bit color mode. It means that FeGorce3 overclocking is really good - the performance grows in direct ratio to the clock frequencies increase.

To assess the efficiency of core and memory overclocking, i.e., how well the architecture is actually balanced, we measured the performance at 1600x1200, when the main strain is produced upon the graphics card and the influence of all other components can be neglected. We measured the performance at nominal core and memory frequencies, with both core and memory overclocked to the maximum and at two intermediary frequencies taken with the same step. As a result, we got 4 different pairs of core and memory frequencies. We tried to make the graphs correct and easy to read, so the horizontal axis indicates the core and memory frequency increase (in percents) relative to the nominal frequencies, and the vertical axis shows the performance gain (in percents again) relative to the performance obtained at nominal frequencies:

As we have supposed, in 16bit color mode core overclocking brings about much more tangible results. The impact of the graphics memory clocking from 460MHz to 600MHz (while the core frequency remains nominal) is quite slight, but grants more notable performance gain when the core frequency gets higher.

In 32bit color mode it's better to overclock the memory. Core overclocking generates a good performance upturn too, especially if the memory frequency is fixed at an over-nominal level.

At last, due to the possibility to increase not only the core frequency but also the graphics memory frequency, the performance grows sharply in 32bit color mode as well. As you remember, we didn't observe this phenomenon with GeForce2 GTS and GeForce2 MX, where we managed to modify only Vcore and not the memory.

BTW, it's interesting to assess the performance drop when 16bit color mode is changed to 32bit during extreme overclocking:

With 200MHz core and 600MHz memory frequency the performance in 32bit mode is even higher than in 16bit color mode. Surely, no one will use a graphics card based on NVIDIA GeForce3 chip working at frequencies like that, but as far as we remember, it's for the first time that a graphics card runs faster in the harder 32bit mode.

At 1600x1200 the performance drop caused by the shift from 16bit to 32bit mode with different core and memory frequencies looks like this:

When the core frequency is fixed, higher memory clocking leads to a linear performance fading, and at 200MHz/600MHz the drop becomes negative. That is, at 200MHz/600MHz in 1600x1200x32 mode the graphics card is faster than in 1600x1200x16 mode.

This situation can be explained with the fact that at the frequencies that high the insufficient graphics memory bus bandwidth hardly tells on the performance in 16bit mode as well as in 32bit mode. The performance difference in this case is determined by how well the cache, memory controller and GeForce3 architecture in general are optimized for 32bit color and corresponding sizes of data units.

Conclusion

We can imagine the performance of GeForce3 Ultra chip if it happens to be manufactured one day. If it differs from GeForce3 only in core and memory clocking, then at 250/600MHz it should be approximately 25-30% faster than NVIDIA GeForce3.

And those of you who would like their GeForce3 to run faster and are not afraid of risks could try to repeat our experiment. Due to the opportunity to bring up both core and memory voltage, you can proportionally increase the working frequencies till purely fantastic heights - 270/600MHz - and enjoy your graphics card revealing excellent performance in the hardest color modes.

Reminder:

• This research is just a kind of experiment and shouldn't be taken as an appeal to taking up extreme overclocking and graphics cards resoldering.
• These modifications shorten your card's service life.
• Any sort of mechanical modifications deprives the users of the warranty.
• Should the graphics card or other components be wrecked, the users bear the complete responsibility for their actions.