iXBT Labs - Computer Hardware in Detail






Sapphire Atlantis RADEON 9700 Volt Modification

January 28, 2004


1. What is volt modification?

It's not easy to answer the question whether volt-modification is needed to survive when gaming requirements for hardware, in particular video adapters, are getting tougher, or it is a way to squeeze more potential from a video card to get more FPS in games. The video card tested today has a performance level sufficient for all existent games, that is why it matches the latter option. For low-level solutions such modification is a way to get more FPS needed for acceptable gameplay.

Volt-modification is an attempt to increase a video card's overclocking potential by changing its internal working voltages (memory, GPU, I/O circuits). But such attempt doesn't guarantee a higher performance. This procedure is identical to changing CPU and memory voltages from BIOS. One can resort to it if a given video card doesn't provide such software means.

2. What is it for?

For this video card volt-modification simply demonstrates its overclocking potential because its performance in gaming applications is not critical.

Reasons for volt-modification:

  1. It's always worth overclocking PC hardware;
  2. Desire to boost performance by inexpensive changes;
  3. For the love of it;
  4. Standard overclocker's principle: to get more for nothing.

I must say that we didn't aim at overclocking the video card extremely for revealing core and memory maximum potential (with various coolants) but wanted to get the fastest possible system that would match any home PC.

3. Materials and tools

The list below shows the tools we used, but it is not a reference set. Any component can be replaced at your option.

Materials and tools (commented):

  1. Skillful hands (without them, please, consider the following material as fiction);
  2. 25W soldering iron (or a little bit more powerful);
  3. Wires;
  4. A couple of precise variable resistors;
  5. A great number of additional resistors of different rated values;
  6. Tester (digital preferred);
  7. KPT-8 grease (used as thermal interface between heatsinks and chips).

You will also need documentation and circuit diagrams (see the details below).

Before starting volt-modification you have to carry out effective...

4. Preliminary testing

Well, it's time to have a look at the video card's specs. In this section we test the pre-overclocking performance. The scores will be further used for estimating the performance increase after overclocking.

Fig. 1. Sapphire RADEON 9700 vide card with reference PCB design

Video card features:
  • Video processor: R300
  • Video memory: 128 MB, Infenion 3.3 ns., 256bit interface, BGA package
  • GPU clock: 275 MHz
  • Video memory clock: 540 MHz


  • CPU: AMD Athlon XP Barton 2500+@3200+ (10.5x210MHz)
  • Mainboard: EPOX 8RDA3+ (driver v. 2.64, BIOS of 27.08.03)
  • Cooler Zalman 6000-CU
  • Memory: 2x256 MB M. Tec PC3200 (CAS Latency 2) Bus 210 MHz
  • HDD Seagate ST380011A 80 Gb, 7200 Rpm
  • Sound card: Creative SB Live 5.1 (driver KX v. 3536)
  • Display driver: Catalyst 3.10
  • OS: Windows XP + SP2 v. 1204

Test applications:

  1. 3DMark 2003 v.340, it directly depends on video card's speed and reflects all frequency changes.
  2. Quake 3 Arena ver. 1.32 (for revealing performance gain in old applications).
  3. Unreal Tournament 2003 (for revealing performance gain in latest applications).
  4. 3DMark 2001 SE v330 for testing speed of applications using DirectX7 and DirectX8.

We didn't make a special accent on testing the card's performance because the today's aim is to find out if volt-modification is justified. The test conditions are ordinary: running applications (Apache, Ftp server, VirtualCD, PowerStrip, LanChat, CdSlow) were not taken out of memory.


Test modes for Quake 3 Arena:

  1. 1600x1200x32 highest quality + 6xAA + 16xAF
  2. 1600x1200x32 highest quality

Test modes for Unreal Tournament 2003:

  1. 1152x864x32 highest quality

Test modes for 3DMark 2003 v.340 and 3DMark 2001 SE v.330: default settings

Scores for 275/540:

  1. 3DMark2003:

    Fig. 2. 3DMark 2003 at 275/540

  3. 2. Quake 3 Arena ver. 1.32 (demo0000 map dm17):


    Fig. 3. Quake3 Arena at 275/540

  5. 3. Unreal Tournament 2003:


    Fig. 4. Unreal Tournament 2003 at 275/540

  7. 4. 3DMark 2001 SE:


    Fig. 5. 3DMark 2001 SE

5. Cooling

The GPU has the following cooler:

Fig. 6. Standard cooler

This compact cooler is able to ensure a reliable operation of the GPU at the default clock speeds as air is taken out in all planes. But the cooler is not big enough for overclocking. Besides, the fan is pretty small and its fast rotation makes a lot of noise. Moreover, this particular fan turned out to be of low quality, and 2 months after its purchase it outcried all other fans and produced sounds unpleasant for human ears.

The video memory had no cooler at all.

Taking into account the layout of all the components let's design a new cooler.

5.1. Materials

The new heatsink is based on Thermaltake's Vulcano 6 Cu because it has the matching size, a big copper plate in the base to ensure better heat removal; besides, it's not that expensive.

Fig. 7. GPU heatsink

The heatsink (without a fan)) measures 80*40*62 mm.

The copper cooler marked AVC C6010B12H was used for the memory. Since we plan on overclocking the memory up to the upper limit, such solution will considerably increase heat removal. That is why we needed an effective passive (for quiet operation) cooler, and we have decided on that inexpensive copper heat sink.

Fig. 8. Heatsink for video memory

5.2. Designing, assembling and mounting

Now we turn to designing a cooler.

Before installation of the heatsink we made it twice lower - 20 mm. It's sufficient for effective cooler, and such design won't lock free PCI slots on mainboards which have the first PCI slot facing up (EPOX 8RDA3+). Such boards have a 35mm gap between the AGP and first PCI slots. We made holes in the sink base for projecting parts. Their size and location were experimentally calculated due to the difference in PCB designs, rated values (and sizes) of the parts and their location.

When we installed the sink for the first time it turned out that the metallic frame that protected the GPU seated unevenly. The sink couldn't tightly abut on the chip, and the latter quickly overheated. We didn't remove the frame as it could damage the GPU, but we made a cut (0.5 mm) in the sink base:

Fig. 9. Heatsink for GPU after milling

The mount was polished and the heatsink was tightly pressed against the GPU. After that the card worked steadily. The sink was outfitted with a fan from the AVC cooler (the fan was of very high quality):

Fig. 10. Cooler mounted on the video card

ATTENTION! The fan is shifted from the die center as far as the heatsink size makes it possible. It was done to prevent the dead area (area under the fan center without any air flows) above the die and make the latter blown with a strong air flow (which increases as the distance from the fan's center grows up).

The sink is screwed through the standard holes in the video card. Let's dwell on the procedure of pressing and clamping the sink to the card.

The screw-bolt has the following parts on underneath:

  • Vinyl washer (provides insulation of the sink mounting systems from the card's components and uniformly distributes the sink load);
  • Metallic washer (protects the vinyl washer from damages from the spring);
  • Spring (balances the load on the GPU die);
  • One more metallic washer (doesn't let the spring to get away and makes screw-tightening easier);
  • Nut (fixes the whole construction).


    Fig. 11. GPU heatsink mounting

After that we made 8 small heatsinks out of the AVC sink (sawed it into 8 ones), each being a bit larger than the memory chip. The remaining part of the sink used for cooling the GPU voltage regulator. 

The only problem in installation of the heatsinks was their mounting.

We didn't want to glue the chips because we could have problems when removing the heat sinks, and decided to fix them with clips. The principle of their operation is based on the metal memory effect: after putting it into a certain shape by bending and unbending it within certain limits the metallic part tends to return to the original state (before unbending). We used a copper wire 2 mm in diameter as a clip; its ends press down the heatsinks mounted onto the memory on both sides. Have a look at the result:

Fig. 12-14. Video memory cooler

Now it's necessary to provide effective cooling of the GPU voltage changer chips (because overclocking causes higher power supply, which makes the stabilizing chips warmer). For that purpose you must remove the heat distributing aluminum plate. ATTENTION! Remove it very carefully as it's attached to the chips with thermal glue. It's very simple to damage the chips - in our case one of them remained on the heatsink and the video card stopped working. Besides, it's almost impossible to find a substitute. We were lucky to find a similar chip from the damaged GF3Ti200 128 MB card (ABIT Siluro). After that the copper sink the mount of which was modified so that it didn't get in the radio elements' road was installed onto the voltage changer. The clamp fixed the system. Here it is:

Fig. 15. Cooler for the voltage changer

6. Volt-modification

6.1. Preparation

For volt-modification we need a soldering iron, a digital multimeter with the internal resistance at least 10 MOhm (it's needed to exclude the measurement error of the multimeter). In our case it was MASTECH MY-65. Wire (you'd better use a braid of three wires as it makes the connection to the control points more reliable and easily accessible), two precise variable resistors (they should be wire-wound resistors because of small Ohm steps), and their accuracy should be within 1-10%. Rated values: 30 to 100 KOhm.

ATTENTION! The theoretical maximum memory clock speed can be calculated by dividing 1000 by the memory access time in nanoseconds. In our case the access time was 3.3 ns, that is why the maximum possible working frequency was 303 MHz, or 606 MHz taking into account the DDR type. Before increasing the memory power supply the clock rate was 570 MHz, i.e. much lower than the limit. Memory voltage is usually lowered by manufacturers to separate the market segments. The fact that the memory clock speeds exceeded the theoretical ones after the volt modification procedure was to the credit of the given sample. That is why if it's possible to overclock memory on your card using the software means think well before starting volt modification (the experience show that the working clock speeds almost always exceed theoretical ones after such modification).

6.2. Memory volt-modification

In our case the memory volt-modification was carried out because the memory didn't properly overclock at the rated power supply which MUST be measured before the process is started. For that purpose we took the documentation from the manufacturer - in our case it's ISL6522 memory internal circuits power supply pulse-width modulation controller (fig. 16) from 

For the measurements you have to solder the wires to the outputs 8 (+ internal circuits (VDD) of the memory chips), 7 (GND common) and 5 (a padding resistor (which should be variable at the first state of resistance selection) is soldered between the 5 and 7 outputs). 

Voltage is measured between outputs 7 and 8, and it was 2.8V in our case. We DO NOT RECOMMEND that you take our values as reference ones because you will most probably have different ones as it depends on the card maker and its memory. Even the same cards from different lots can differ.

The output voltage of the internal memory power supply circuits is assigned by the resistance ratio in the feedback circuit. To increase the power supply voltage you should change this ratio by adding a padding resistor; it makes the voltage on the feedback circuit output lower, and the controller which compares it to the reference one increases this voltage up to the needed level, though the real voltage turns out to be increased.

Fig. 16. Reference diagram of pulse-width modulation controller ISL6522 (the padding resistor added is red)

The 47 kOhm wire-wound resistor of 10% tolerance was chosen as a variable one.

ATTENTION! While increasing the core and memory voltage we recommend that you solder a fixed resistor of 1.5 to 2 kOhm in parallel to the variable one. Without the additional resistor the variable resistance can drop much causing an increase in power supply on the respective blocks. These measures avoid damage of the card.

Now we proceed to the key operation - the voltage increase.

Measurements and parameters setting should be carried out very carefully and accurately as these circuits are very sensitive to any kinds of pickups, and even if you slightly touch the bared wire, the video card can hang making you restart the computer (the reset button won't help). The voltage must be controlled with a multimeter. You should lift it up in small steps reducing the variable resistance and controllong the variable resistor every time if you don't have an additional resistor in parallel to the variable one. The maximum voltage which can be reached in most cases is 3.2- 3.3V. However, you can try to get more at your own risk. At the higher voltage it can be insufficient to have only the passive cooler. For our video card the end-point output voltage of the internal memory power supply stages was 3.28V, which is a reasonable compromise of the danger of having the card damaged and its overclocking. After the voltage is lifted up you should carry out some tests with the memory clock increased, and if everything it okay, you must unsolder the resistor, measure the resultant resistance and replace it with an identical additional resistor (if you can't get it with one resistor, you can use a parallel of series connection of 2 or more resistors). The resistors should be covered with protective cambric. Here is how it looks like after all the operations:

Fig. 17. Modified memory power supply chip

6.3. GPU volt modification

To increase the GPU stability at the increased clock speeds and to prevent any artifacts in case of overclocking at the rated power supply (we could see a chessboard-like screen and a distorted shape of some 3D objects in some screen parts) you should raise the GPU power supply.

In contrast to the memory volt-modification, for the GPU you increase the voltage on a different power supply controller chip. The padding resistance and location of the control points change at that. 

The safety margin of the voltage changing is 10 - 30% relative the default value. The standard supply voltage makes 1.5V in many cases. This voltage is controlled by the double synchronous low-power direct-voltage controller with the combined current circuits SC1175 from 

The chip has two independent voltage control and stabilization channels, but in this case they are connected in parallel:

Fig. 18. Standard circuit diagram with parallel connection of the channels. (The padding resistor is shown on the diagram)


Fig. 19. Pin-out of SC1175 chip

You can calculate the resultant resistance by the following formula:

R1 = (R1xR1add.) / (R1+R1add.)

Vout=1.25 x (1+R12/R1)

The weak point of such approach is that you have to unsolder the resistors from the board because of the large capacitors at the converter output.

The precautionary measures and the principle of operation when increasing the memory voltage is similar to the memory volt modification.

If you see white ripples on the screen after turning it on after all the modifications, it means that you applied to much power to the GPU which can soon initiate irreversible processes and change the core structure (damage the internal circuits and some of the die's parts).

You must solder wires to the outputs 16 (+converter output stages), 18 (feedback output where you set voltage by changing the resistance, with a padding resistor added between this output and output 20) and 20 (GND common) to change and control the voltage.

The voltage is to be changed between outputs 16 and 20 of the power supply chip SC1175.

Like in case of ISL6522 chip the output voltage can be modified by changing the resistance ratio in the feedback circuit. A padding resistor is to be added between outputs 18 and 20.

NOTE! Similarly to the memory modification, you should solder an additional resistor of 1.5 to 2 kOhm in series to the variable resistor.

You must control the voltage increase and test the video card at every step.

The end-point voltage on the GPU and the padding resistance is to be calculated when the wanted result or the maximum working GPU clock is reached. In our case it was 1.65V, and the resistance was 9.1 kOhm.

The card reached very high scores at 1.65V, which is 10% greater than the default value. Here is how the video card looks like after the CPU volt-modification:

Fig. 20. Modified GPU power supply chip

6.4. Overclocking

After all the changes the card looked like that:

Fig. 21. Modified video card. Top view

Fig. 22. Modified video card. Back view

Here is the system with the video card installed:

Fig. 23. Assembled system

The card's operation is considered steady if it passes 3DMark 2003 three times without artifacts and Mother Nature four times without artifacts (this is the heaviest test, and problems, if any, will be quickly revealed).

The clock speeds our cards works steadily at are: 

  • GPU clock: 400 MHz.
  • Video memory clock: 620 MHz.

Note: it turned out that 3DMark 2003 is not enough: at first, the steady clock was found to be 410/635. The test packet passed it several times without any artifacts. However we noticed distortions in games 20-30 minutes after the game started. That is why we had to reduce the clock speeds.

Now we are going to test the card (the test scores of both the overclocked card and the original one are brought together in the diagrams):

  1. 3DMark2003:


    Fig. 24. 3DMark 2003 at 400/620


    Fig. 25. 3DMark 2003

    Fig. 26. 3DMark 2003 in GT1

    Fig. 27. 3DMark 2003 in different tests

    Fig. 28. 3DMark 2003 in Fill Rate benchmark

  3. Quake 3 Arena ver. 1.32 (demo0000 map dm17):


    Fig. 29. Quake 3 Arena

  5. Unreal Tournament 2003:


    Fig. 30. Unreal Tournament 2003

  7. 3DMark 2001 SE:


    Fig. 31. 3DMark 2001 SE at 400/620

    Fig. 32. 3DMark 2001 SE

The overall performance increase in 3DMark 2003 is 27%.

Performance gain in subtests:

  1. Wings Of Fury - 13%
  2. Battle Of Proxycon - 30%
  3. Troll's Lair - 17%
  4. Mother Nature - 39,5%
  5. Vertex Shader - 42.9%
  6. Pixel Shader 2.0 - 26.5%
  7. Ragtroll - 30.6%
  8. Fill Rate (Single-Texturing) - 52.8%
  9. Fill Rate (Multi-Texturing) - 63.7%

The overall performance increase in Quake 3 Arena:

  1. Highest quality + 6xAA + 16xAF - 33.8%
  2. Highest quality - 26.3%

The overall performance increase in Unreal Tournament 2003:

  1. Botmatch antalus: min - 10.5%; average - 1.4%; max - 4.4%;
  2. Flyby antalus: min - 10.4%; average - 12.8%; max - 55.5%;

The overall performance increase in 3DMark 2001 SE is 18.6%.

Performance gain wasn't estimated in the separate subtests of 3DMark 2001 SE.

The general performance increase is 25-30% in all applications. Judging by the 3DMark 2003 the picture looks logical: the heavier the test, the greater the speed drop because of the great load on the video card. The texturing speed is 1.5 times greater at least, and it will have a positive effect on games that deal with large textures. In Quake 3 Arena in the heavy mode the overclocking has a greater effect too as the load is greater. In Unreal Tournament 2003 the scores depend mostly on the CPU.

Now, when the card has a nice appearance, let's also modify its BIOS.

7. BIOS modification

There are several reasons why we decided to change the BIOS parameters:

  1. A higher overclocking potential. It can be obtained by reducing CAS from 2 to 1. It makes the latency lower and the card steadier at high clock speeds. But in our case the attempt to set CAS 1 resulted in artifacts at all clock speeds, even default ones. Such method can be used for video cards having the memory working at higher frequencies (presumably from 700 MHz).
  2. We didn't want to resort to software overclocking all the time. After all the tests were finished the following clock speeds were set: 400/620. They were flashed into BIOS.
  3. Problems occurring when one uses a software overclocking utility. At OS booting, at the moment of setting the new clock speeds for the video card we could notice 3-4 black "noisy" squares. The same artifacts appeared later, after the OS was booted, even without games. The PC reboot didn't help because even the POST display had distorted symbols. Only PC restarting could help (turning on/off). First at all we thought that we set too high video memory voltage, but the artifacts appeared when the video load was minimal. So, we decided to set new clock speeds in BIOS, and the problems disappeared and never appeared again for 2 weeks. Well, we suppose that this artifact takes place as the overclocker sets clock speeds incorrectly. But why did they appear while the system was working? Is it because the utility regularly restores the specified clock speeds? The questions are actually addressed to the developer of this utility.

7.1. Possible problems

There can be several problems:

  1. ATTENTION! Before flashing the BIOS remember the Flash ROM chip type of your video card so that you can be sure that the flash software supports a given Flash type. IMPORTANT NOTE! If you have SST M25P05 Flash ROM DO NOT use Atiflash for flashing in BIOS. In 99% of cases you will get the message Timeout: Poll rom idle after bulk erase. It means that the flash is erased but a new BIOS can't be recorded. The video card doesn't work anymore. Only a PCI card can help it... but the attempts to reflash the BIOS with Atiflash resulted in the same error, and other flash utilities do not see this adapter at all. The only program that could flash the BIOS was FlashROM 1.3 (ATTENTION! The versions over 1.3 do not help and display the message VGA Adapter not found).
  2. It's DESIRABLE that you use the video card's own BIOS saved on the disc in advance because BIOSes of video cards built on the same graphics core differ (the difference can be caused by any factor: manufacturer, video memory supply voltage, GPU supply voltage, the way the primary blocks get power, electrical scheme etc.). If you flash BIOS taken from a third card without checking characteristics of its nodes you can make its operation unsteady and even damage the card.

7.2. Results

After all those time-taking experiments with the BIOS we got only one important advantage: new clock speeds were flashed into the BIOS. At the same time we changed the TV standard from NTSC to PAL.

All the experiments gave us more food for thought.

Software used for video card overclocking and BIOS modification:

  • RivaTuner 2 RC14.1
  • Atiflash (mostly used for "corrupting" the BIOS)
  • FlashROM 1.3
  • ATi BIOS editor v. 1.1D

8. Conclusion

We spent quite a lot of time for all the designing and developing and got a card the performance of which increased by 25-30%. The Sapphire RADEON 9700 performs even better than the PRO version. This card can be named Sapphire RADEON 9700 XT. We described all the problems we faced when improving the adapter. And we hope that this article will help those who'd want to carry out volt modification of their cards. We believe that this review will help you decide whether it's worth modifying your card this way.

NOTE! Do not do it if: 

  1. you do not know why do you need it for;
  2. you have no idea about how it works;
  3. you are going to do it without thinking well before starting it;
  4. you are going to ask what supply voltage is;
  5. you are going to ask where and how to measure supply voltage;
  6. you are afraid of disassembling expensive devices; 
  7. you never used a soldering iron before;
  8. you do not know what a soldering iron is;
  9. you do not have any rudiments of knowledge and elemental skills in radio engineering and do not know how to calculate the resultant value of series and parallel connections of resistors, and if soldering is supernatural for you;


This volt modification process was recorded! The clip can be downloaded from here (95 MBytes!). We express our gratitude to Driver.Ru and Sergei Komarov.

Artem Sokolovsky (voltmod@tut.by)
Pavel Shilenok (voltmod@tut.by

Write a comment below. No registration needed!

Article navigation:

blog comments powered by Disqus

  Most Popular Reviews More    RSS  

AMD Phenom II X4 955, Phenom II X4 960T, Phenom II X6 1075T, and Intel Pentium G2120, Core i3-3220, Core i5-3330 Processors

Comparing old, cheap solutions from AMD with new, budget offerings from Intel.
February 1, 2013 · Processor Roundups

Inno3D GeForce GTX 670 iChill, Inno3D GeForce GTX 660 Ti Graphics Cards

A couple of mid-range adapters with original cooling systems.
January 30, 2013 · Video cards: NVIDIA GPUs

Creative Sound Blaster X-Fi Surround 5.1

An external X-Fi solution in tests.
September 9, 2008 · Sound Cards

AMD FX-8350 Processor

The first worthwhile Piledriver CPU.
September 11, 2012 · Processors: AMD

Consumed Power, Energy Consumption: Ivy Bridge vs. Sandy Bridge

Trying out the new method.
September 18, 2012 · Processors: Intel
  Latest Reviews More    RSS  

i3DSpeed, September 2013

Retested all graphics cards with the new drivers.
Oct 18, 2013 · 3Digests

i3DSpeed, August 2013

Added new benchmarks: BioShock Infinite and Metro: Last Light.
Sep 06, 2013 · 3Digests

i3DSpeed, July 2013

Added the test results of NVIDIA GeForce GTX 760 and AMD Radeon HD 7730.
Aug 05, 2013 · 3Digests

Gainward GeForce GTX 650 Ti BOOST 2GB Golden Sample Graphics Card

An excellent hybrid of GeForce GTX 650 Ti and GeForce GTX 660.
Jun 24, 2013 · Video cards: NVIDIA GPUs

i3DSpeed, May 2013

Added the test results of NVIDIA GeForce GTX 770/780.
Jun 03, 2013 · 3Digests
  Latest News More    RSS  

Platform  ·  Video  ·  Multimedia  ·  Mobile  ·  Other  ||  About us & Privacy policy  ·  Twitter  ·  Facebook

Copyright © Byrds Research & Publishing, Ltd., 1997–2011. All rights reserved.