The arrival of graphics cards based on ATI RADEON 9000 / RADEON 9000 PRO chips into the mass market means that there appeared new alternatives among the mainstream products.

These relatively inexpensive graphics cards have every chance to become popular among those who stick to the principle sounding something like "to get more but to pay less". Many people put this rule into practice by overclocking the newly bought graphics cards. Some quite risky guys go even further and plunge into PCB modification and extreme overclocking.

Today we're going to take the risks again and to check the overclockability of ATI RADEON 9000 Pro solution. Our guinea pig today will be Sapphire RADEON 9000 Atlantis PRO 64MB graphics card:

So, let's get started…

Cooling System Modification

To cool the graphics memory chips, I used a heatsinks from Thermaltake Memory Cooling Kit. I had to mount them a little to the side: there are capacitors between the memory chips that rise slightly above them and hinder proper installation of the heatsinks this way.

The heatsinks were installed onto sticky thermo-conductive pads from the Thermaltake Memory Cooling Kit.

For better chip cooling, I had to put off the rather weak standard cooler and install a specifically adjusted cooler for Slot A processors. The adjustment meant sawing off a few ribs so that the ribs didn't sit against the PCI slots, thus preventing proper graphics card installation. I first used this cooler during our extreme overclocking of NVIDIA GeForce2 GTS and GeForce2 MX based cards (see NVIDIA GeForce2 GTS and GeForce2 MX Voltage Tweaking and Extreme Overclocking) and it proved a helping hand to the poor overclocker :).

In fact, I had problems fastening the cooler to the graphics core, but eventually found a solution. The cooler was fastened with nylon braids passed through the heatsink ribs:

   

The locks of the braids don't fit in between the ribs and hold the cooler tightly. On the reverse side of the PCB the same locks were holding the braids.

   

As a result, the cooler was tightly pressed to the graphics chip surface, as if it was stuck to it. I also added some Russian made organic-silicon thermal paste aka "KPT - 8" to ensure the best contact between the cooler foot and the chip surface.

After all the preliminary work has been finished, the Sapphire RADEON 9000 Atlantis PRO graphics card got much more impressive looks:

Vcore Increase

The graphics chip in Sapphire RADEON 9000 Atlantis PRO receives 1.56V power voltage. It is supplied by the pulse regulator implemented in the IRU3037 chip from International Rectifier.

The output voltage of the regulator is determined by the R3 and R5 resistors (in the typical application circuit scheme) by the formula: Vout = 1.25x(1+R3/R5).

The typical application circuit scheme of IRU3037 looks as follows:

In order to raise the output voltage, I lowered R5 resistance (in the typical application circuit scheme) by soldering a 5.1kOhm resistor parallel to it. It appeared much easier to solder this one not to the resistor located beside the IRU3037 chip, but straight to the chip's pins.

   

As a result, we managed to increase the chip Vcore up to 1.78V, i.e. by the good 14.8% above the nominal value.

Vmem Increase

Sapphire RADEON 9000 Atlantis PRO is equipped with DDR SDRAM chips from Hynix with 3.6ns access time and HY5DV641622AT-36 marking. According to the manufacturer's documentation, these chips have to be power-supplied by a double power source: 2.5V - for inner circuitry and 3.3V - for I/O buffers.

On the graphics card from Sapphire, the graphics memory voltage regulators provide 2.6V and 3.01V respectively. Let's try to increase these rates.

The I/O buffers are supplied by an ordinary linear regulator based on discrete elements. Its output voltage is determined by the ratio of the PCB resistors. On the snapshot these are the one below the additional resistor and the one more on the left and lower.

To raise the voltage, I had to shunt one of the resistors with an additional 4.7kOhm resistor.

   

After the modification, the output voltage of the regulator made 2.89V, that is, 11.2% higher than before.

The inner circuitry of the graphics memory chips is powered by one more regulator implemented in the IRU3037 chip. Like the chip regulator, this is also a pulse one. Its output voltage is determined by the ratio of the on-board resistors and can be calculated with the following formula: Vout = 1.25x(1+R3/R5), where the resistors are marked according to the typical application circuit scheme.

In order to raise the output voltage of the regulator, I had to either shunt R5 resistor (according to the typical application circuit scheme) or solder up an extra resistor to the chip's pins, as we did before.

I chose the harder way: found the necessary resistor on the card PCB and soldered an additional 6.8kOhm resistor to it:

   

This modification yielded 3.65V on this regulator, i.e. the voltage got 21.3% higher than the nominal.

So, the new cooling system has been installed, the voltages have been increased to the utmost of their potential. Now let's see how the card is going to do in the benchmarks.

Testbed

Our testbed configuration looked like that:

• AMD AthlonXP 2000+ CPU;
• MSI K7T266 Pro2 v2.0 (VIA KT266A) mainboard;
• 2x256MB PC2700 CL2.5 DDR SDRAM by Crucial;
• Fujitsu MPF3153AH HDD.

We used the following software:

• 6.13.10.6118 driver for graphics cards based on ATI chips;
• Windows XP;
• 3DMark 2001 SE build 330;
• Quake3 Arena v 1.30;
• Serious Sam: The Second Encounter.

The test were run with the following settings:

3DMark 2001 SE:

32-bit frame buffer; 32-bit textures, 32(24)-bit Z-buffer, D3D Hardware T&L / Pure Hardware T&L.

Quake3 Arena:

32bit screen and textures color depth. Graphics quality settings at maximum. Tri-linear filtering and texture compression enabled.

Serious Sam: The Second Encounter:

The tests were run in Quality mode: 32bit screen color depth; "Quality" graphics quality settings.

Performance

The nominal chip and graphics memory working frequencies by Sapphire RADEON 9000 Atlantis PRO make 275Mhz and 550MHz (275MHz DDR) accordingly.

The maximum frequencies we managed to overclock the graphics card to and to have it run stably before any modification was made equaled 295Mhz for the chip and 620Mhz (310MHz DDR) for the graphics memory.

After the modification, the extremely overclocked card worked stably at 340MHz chip and 680MHz (340MHz DDR) memory frequencies.

So, we see that the ATI RADEON 9000 PRO chip showed better overclockability than ATI RADEON 8500, at least on this graphics card. The maximum chip frequency shown by ATI RADEON 8500 during extreme overclocking was 320MHz (see ATI RADEON 8500 Extreme Overclocking Experience). The nominal frequencies of ATI RADEON 8500 and 9000 PRO are equal: 275MHz.

It is evident that ATI RADEON 9000 PRO has fewer transistors than RADEON 8500 model: it generates less heat and must have been modified to be able to work at higher frequencies, so that to achieve higher manufacturing output. The result is here: higher frequencies we got by extreme overclocking.

The graphics memory of Sapphire RADEON 9000 Atlantis PRO was overclocked to 680MHz (340MHz DDR). It's certainly lower than those 704MHz (352MHz DDR) we had by ATI RADEON 8500. Both graphics cards, Sapphire RADEON 9000 Atlantis PRO and ATI RADEON 8500, feature the same 3.6ns memory chips from Hynix. The worse memory overclockability may be explained by a "cheaper" design of the card. It's targeted at the mainstream market while ATI RADEON 8500 was a high-end card when it came to the market. Well, maybe it's not the case at all and I just was unlucky with the graphics memory in this given card. I chose it at random, without any particular overclocking potential check.

So, let's check the performance gain we get as a result of extreme overclocking of ATI RADEON 9000 PRO.

Quake3 Arena is twice as sensitive to graphics memory overclocking as to chip overclocking. The game requires high fillrate and graphics memory bandwidth.

The simultaneous overclocking of the memory and graphics chip to 340MHz (23.6% higher than the nominal), the performance gain makesis 21.8%. The RADEON 9000 PRO based card is just a little short of its "fully-fledged" forefather, ATI RADEON 8500.

Here we have the same picture: the graphics memory overclocking proved more efficient.

A perceptible performance gain is only felt in the 1280x1024 mode (up to 17.7%). In lower resolutions, the processor speed is limiting the results, so that we cannot enjoy the overclocking magic in full.

Graphics memory overclocking is even more advantageous in 3Dmark 2001(Nature) than in the previous tests.

Note the graphics chip performance gain graph. It's remarkable that with the overclocked chip the performance gain doesn't grow up as the resolution increases, but on the contrary, gets gradually lower. Usually, the gain is higher with higher graphics card workload and lower "limiting" CPU impact that we saw in the previous tests. But here even in 800x600, the processor doesn't slow down in any way the graphics card performance. In higher resolutions, insufficient graphics memory bandwidth makes the un-overclocked memory a bottleneck in the system, thus diminishing the effect of the overclocked chip.

When we overclock both: the chip and graphics memory, the performance reaches 24.7%. It's even higher than the actual increase in the chip and memory clock-rates :). Of course, this curious result is nothing but a measuring error, which occurs as a result of approximations allowed by 3DMark 2001.

The extreme overclocking even helped ATI RADEON 9000 PRO to surpass ATI RADEON 8500 LE 128MB in the test that uses pixel and vertex shaders. Especially since RADEON 9000 PRO is rather poor at processing them.

Conclusion

So, ATI RADEON 9000 PRO chip has good overclocking potential and shows appropriate performance gain. The Sapphire RADEON 9000 Atlantis PRO graphics card based on this chip is an excellent proof of the fact.

Of course, we might wish something more than 23.6% frequency increase and about 20% performance growth. Those are not the figures we are used to during extreme overclocking and voltage increase, as we usually look forward to some more striking results. However, let me remind you that firstly, I raised the voltages not too "extremely" and didn't want to reach the very maximum frequencies at any rate and secondly, the card I had at my disposal was an ordinary product from an ordinary manufacturer and it claimed no "high overclocking potential".

Nevertheless, our overclocked card fell just a little behind ATI RADEON 8500 LE 128MB in Quake3 Arena, left it far behind in Serious Sam: The Second Encounter and outperformed it in 3DMark2001 SE (Nature).

By the way, one more thing worth mentioning: ATI RADEON 9000 PRO chip used in the tested card from Sapphire was stable working at the nominal frequency even when we reduced the chip voltage down to 1.4V. I also dismantled the active cooling from the chip and installed a small passive heatsink they usually use in NVIDIA GeForce2 MX200/MX400 based cards. There were no problems or complaints about the card's performance in these conditions.

So, all of you who are irritated by the noise produced by the system block, who have already made all other devices shut up and want to do the same with the graphics card, may take an ATI RADEON 9000 PRO based card and eradicate the noise in such an original way.

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.