by Tim Tscheblockov
02/27/2004 | 04:04 AM
Every new generation of graphics cards offers more performance to the user. As progress continues, today’s mainstream cards easily leave behind the solutions that used to be the fastest two-three years ago. The graphics processor itself is becoming more complex, with more transistors per die. The “transistor density” of the today’s graphics processors can now be considered as high as that of the most sophisticated and high-tech component of a personal computer, the central processor.
As graphics processor and memory technologies evolve, power consumption and heat dissipation of graphics cards grow ever higher, so it’s hard to think of a fast graphics card that wouldn’t require a cooling system. Such specimens died out at the end of the last century when NVIDIA’s TNT chip was the pinnacle of 3D graphics. Today you can only meet a graphics card with a passive heatsink or without a heatsink at all at the very bottom of the low-end sector, while products on the other end of the price scale dissipate almost as much heat as some CPUs do. So the cooling system of a top-end card today is not just a “heatsink + fan” combo, but a contraption of plastic, aluminum, copper and heat pipes, a sample of engineering art, with both functional and esthetic purposes. Such contraptions as well as “heatsink + fan” coolers usually do their job properly, but often at the expense of your aural comfort.
Some manufacturers equip their products with low-noise or passive cooling systems and emphasize it in the model name or in the specifications. Cards like that are scanty, though, and they usually come at a higher price compared to their ordinary counterparts.
So what can you do to make this annoying device quieter?
First, you can install a passive cooling system yourself. They are available in the market, for example Zalman, a company specializing in low-noise cooling solutions, offers a series of systems of this kind for cooling the graphics cards. Running a little ahead, I should confess that they don’t always provide the necessary result (we will talk about two cards from Sapphire with such systems today).
Second, you can reduce the voltage of the cooler and, accordingly, reduce the noise, or, third, you can design a low-noise or passive cooling system yourself.
When you switch to passive cooling or reduce the cooler rotation speed, you run the risk of making your card unstable because of overheating of its components. On the other hand, even a non-modded card with its standard cooling system may get to circumstances where overheating is inevitable, for example, if your system case cannot provide proper airflow around the graphics card.
There is also the last way to solve this overheating problem altogether: we must reduce the very heat generation of the card. Is it possible? Yes, it is!
Our problem is simple: we need to find a way to reduce the heat dissipation of the graphics card, at the same time keeping it stable. Overclockers know that the relation between power consumption and heat dissipation of a central processor and its frequency is nearly linear. There are physical reasons for that: processors consume the largest portion of their total power in current surges when switching logical elements made from CMOS transistors. The higher the clock-rate, the more switches per second there are, and accordingly, the higher is the power consumption/heat dissipation. Thus, one of the ways to reduce the generated heat is to reduce the clock rate.
This is not our way, though. People who are into extreme overclocking certainly know that the CPU or GPU generates more heat when its voltage has been increased than when you overclock it without increasing the core voltage. This is another fact, explained by Ohm’s law: the relation between the power consumption/heat dissipation of a processor and its voltage is not linear, but rather quadratic.
Thus, it is better to oppose the high heat dissipation of an electronic chip by reducing its voltage rather than by reducing its operational frequency. We can leave the clock rate alone and reword our problem a little bit: our task is to reduce the amount of heat generated by the graphics card by reducing the voltages of the GPU and graphics memory while keeping the card stable at its nominal frequencies.
The testbed configuration looks as follows:
We used the following software:
I tested the graphics cards in an open testbed rather than in a system case and without any additional air cooling. I chose this test method as a compromise since each particular model of a system case may have affected the results in a certain way. Available system cases differ greatly in the thermal environment they provide for the devices inside, but graphics cards usually have lower temperatures in well-designed cases with proper airflows than in an open testbed. Thus, it is even harder for the graphics card to work in an open testbed without extra cooling than in a good system case. Please, keep this fact in mind.
On the contrary, bad system cases may fry the graphics card much more than the open testbed does, but the reduction of heat dissipation we are about to accomplish is going to help the devices even under most unfavorable conditions.
The room temperature was 21°C during the tests.
I used an infrared “gun” thermometer for measuring temperatures of the graphics processor and memory. The exact locations are shown in the snapshots.
The back side of Sapphire RADEON 9600 XT Ultimate Edition and RADEON 9800 PRO Ultimate Edition cards is covered with a heatsink, so I took the temperature data for the second card from the heatsink surface on the face side of the PCB, above the graphics core.
Such measurements are prone to be inaccurate, of course, but only RADEON 9600 XT card of all we are going to review today has a thermal diode integrated into the graphics processor from which we took very precise data. With other graphics cards, you should be aware that the real temperature of the graphics die was higher than the number we put in the diagrams (the temperature of the PCB against the VPU). However, the difference between the measured and real temperatures remains stable and we can judge the relation of the temperatures with an acceptable degree of correctness. The rotational speeds of the fans (for cards that featured active cooling) were measured using an optical tachometer:
We tested our graphics cards in two modes: “Idle” (no running applications and Windows Desktop on the screen) and “Burn” (highest workload – 3DMark03 runs endlessly with its default settings). I kept the system in each test mode for 30 minutes for the temperatures to become stable. After that I measured the temperatures of the VPU and memory.
Our criterion of stability at reduced voltages is the card’s ability to run a set of 3DMark03 tests with default settings for 1 hour and then work for 1 hour in Far Cry Demo with the “heaviest” settings.
I will only deal with graphics cards on ATI RADEON chips. That’s not because I prefer ATI to NVIDIA, but rather because my own home computer had a RADEON 9500 “modded” to RADEON 9700. I wanted to make it cooler and less noisy and then experimented with cards on other RADEONs to see how profitable my “anti-extreme” overclocking would be with products of different classes.
So, we start out with cards equipped with standard active cooling solutions.
This card was made from the RADEON 9500 following the well-known manipulations, so everything we say about the RADEON 9700 is also true for RADEON 9500-based graphics cards following RADEON 9700 design.
PowerColor RADEON 9700 128MB remains decently cool during work and its cooler is not very loud. Many other RADEON 9500/9700-based cards are similarly unpretentious, not requiring any modifications. In my personal case (in a system assembled from the Shuttle SS51G barebone) little noise from the PowerColor card turned to be quite noticeable, while the power consumption and heat dissipation of the card appeared higher than I wanted them to be.
Let’s get to business.
The graphics processor of the PowerColor RADEON 9700 128MB card is supplied with power through an impulse voltage regulator based on SC1175CSW controller from SemTech. This chip has two independent channels, set up for current separation in our case. The output voltage of the regulator is determined by the ratio of the resistances of resistors in the feedback circuit. All we need to do is to change this ratio in the desired way. It’s simpler to reduce the resistance of one of the resistors by attaching an additional resistor (if we were into extreme overclocking, we would shunt the other resistor).
The voltage regulator for the internal circuitry of the graphics memory chips (VDD) is based on ISL6522 chip from Intersil. The output voltage of the regulator, like in the previous case, is determined by the ratio of the resistors in the divider included into the feedback circuit. So our actions are absolutely the same in this case.
After reducing the voltage of the internal circuitry we should do the same with the I/O buffers (VDDQ). According to the requirements of the memory manufacturer (in our case, Infineon with HYB25D128323C-3.6 modules), VDDQ shouldn’t exceed VDD. The regulator of the I/O buffers of the graphics memory is designed on the same chip as the regulator of the internal circuitry.
The last onboard regulator is responsible for termination circuits (VTT) and, according to the specifications of the graphics memory chips, VTT must be half of VDDQ. This regulator is based on the same chip as VDD/VDDQ regulators, so you only have to find the necessary resistor on the PCB.
The snapshots below show voltage regulators of the VPU and memory (VDD/VDDQ/VTT) and exact locations where I soldered up additional resistors (I used trimming resistors with a resistance of 22kOhm).
I didn’t measure the resistances of the additional resistors with which I achieved the necessary result because the dividers may have different initial resistances on cards from different manufacturers or even from two lots by the same manufacturer. The ratio of resistances is more important here than their absolute values.
The following table shows the initial and final voltages of the VPU and memory. Again, these are the minimal voltages at which the graphics card remained stable at its regular frequencies.
Nominal Voltage, V
Reduced Voltage, V
As a result, we have the VPU voltage reduced by 15% and the voltages of the internal circuitry and the I/O buffers of the graphics memory reduced by 26% and 15%, respectively.
This is what the graphics card looked like during the tests:
The graphics processor and memory immediately reacted to my modding by becoming cooler:
We’ve got excellent results: the temperatures of the graphics processor and memory went down by 4-5°C in the Idle mode and by 8-10°C in the Burn mode.
By limiting the power ration of the fan we nearly halved its speed, from 4020rpm to 2410rpm. The card got silent, producing a kind of soft whisper. The tradeoff was quite acceptable: the card was 3-4°C warmer in the Idle mode and 4-5°C warmer in the Burn mode.
So we’ve reached our goal: by reducing the voltage we made our RADEON 9700-based graphics card both: quieter and cooler.
PowerColor RADEON 9800 PRO is a modern graphics card with a faster graphics chip. R350 VPU (RADEON 9800/9800 PRO) is in fact a slightly modded R300 (RADEON 9700/9700 PRO), particularly it works at a higher frequency and uses higher-voltage power. No wonder then that R350 consumes more power and generates more heat than R300. So the VPU voltage regulator in RADEON 9800 PRO reference design deserves such honor as a separate heatsink on some field transistors. The cooling system changed, too, as the heatsink area became larger and the fan acquired an elegant twist in the blades (to produce much more noise, to my opinion):
Nothing happened to the graphics memory. The chips still have no heatsinks. Under workload, they become much warmer than the memory on RADEON 9700 based card because of their higher frequency and voltage. Let’s see what we can do under such unfavorable conditions.
The design of PowerColor RADEON 9800 PRO differs from the design of the previous card, although all voltage regulators are based on the same chips. So I won’t repeat myself but will just show you a few snapshots with those regulators and places where I put additional resistors:
Note that the nominal voltage of the graphics processor in PowerColor RADEON 9800 PRO is about 0.2V higher than of RADEON 9700! The nominal voltage of the memory chips was also higher by this card. Besides that, VDD of the memory chips is balancing at the very limit of the acceptable voltage range as recommended by the manufacturer (Samsung and K4D26323RA-GC2A chips): 2.96V against the acceptable maximum of 2.94!
This high voltage of the internal circuitry is probably meant to provide stability at high nominal frequencies and under unfavorable conditions, or maybe this is a step towards overclockers?
Nominal Voltage, V
Reduced Voltage, V
The relative reduction of the VPU voltage was bigger than by the previous card: 17.7%. The memory chips, on the contrary, refused to be stable when VDD and VDDQ were dropped down by more than 14.6% and 19.4%, respectively. Obviously, the high clock rate makes it impossible to reduce the memory voltage any further.
So, now the graphics card is ready for our tests:
By reducing the voltage of the cooler from 12V to 7V we dropped its rotation speed from 3810rpm to 2650rpm. That’s a smaller reduction than in the pervious case, but the noise nearly vanished altogether. With less efficient cooling, the temperature of the graphics processor became 2-4°C higher, like by the previous card.
Voltage reduction proved to be of more effect by RADEON 9800 PRO than by RADEON 9700. The VPU was 6°C cooler in the Idle mode and 13-15°C cooler in the Burn mode. The temperature of the memory chips was even more sensitive to the voltage reduction becoming 5°C lower in the Idle and 30°C lower in the Burn mode!
That’s just an excellent result! The modded RADEON 9800 PRO card became both: cooler and quieter than the non-modded RADEON 9700-based card. That’s the more satisfying because PowerColor RADEON 9800 PRO is a more advanced and faster graphics card.
Now we will discuss a few devices with passive cooling systems.
The Ultimate Edition of RADEON 9600 XT from Sapphire features an original design and carries a passive cooling system consisting of two aluminum heatsinks connected by a heat pipe:
This construction becomes very warm during operation. The cooling might be more efficient if the big metal sticker with the Sapphire logo on the face side of the PCB didn’t cover the heatsink ribs, and the heatsinks themselves were better suited for working without air cooling (that is, if they had needles instead of ribs).
The VPU voltage regulator on this device is based on RT9202 controller from RichTek. As usual, we need to change the ratio of resistances in the feedback circuit to reduce the power supply voltage.
Internal circuits and I/O buffers (VDD and VDDQ) of the graphics memory chips are supplied with power by impulse regulators on Intersil ISL6522 and RichTek RT9202 chips, respectively. Again, the method remains the same: find the resistors in the feedback circuits and solder up additional resistors to them to change the resistances ratio in the desired way.
The regulator for the termination circuits is based on RT9173 chip from RichTek. It takes VDDQ as the basic voltage, so you don’t have to do anything – VTT changes with VDDQ.
The following snapshots show you where the VPU and memory voltage regulators are and where I put additional resistors:
RADEON 9600 XT chip gives out relatively little heat, although has the highest frequency of all RADEON 9xxx series chips: it has fewer transistors than R300/R350, it is manufactured with finer 0.13-micron technology and it works on only 1.3V voltage.
Nominal Voltage, V
Reduced Voltage, V
The graphics chip voltage was reduced by 11.5% and it was stable at its original frequency, 500MHz. The graphics memory on the RADEON 9600 XT also works at a high frequency, so I only reduced the voltages of the internal circuits and I/O buffers by 12.7% and 11.1%, respectively.
The card is ready for the tests:
RADEON 9600 XT features an integrated thermal diode, so the temperature measured for the VPU in the following diagram is very accurate:
RADEON 9600 XT dissipates little heat, and the cooling like the one we see in the Sapphire card handles this heat dissipation well enough: 66°C under workload is an acceptable result.
By reducing the voltages we make the graphics processor and memory cooler by 3-4°C in the Idle and by 4-6°C in the Burn mode. Not much, actually. If the cooling system were more efficient and the graphics chip warmer, the effect of the modification would be more conspicuous.
The last graphics card in our today’s experimental investigation is Sapphire Ultimate Edition of RADEON 9800 PRO. That’s a real graphics monster with its 256MB of DDR2 memory and a huge cooling system from Zalman consisting of two heatsinks connected with a heat pipe. Every memory chip is covered with its own heatsink:
When the graphics card was loaded with some heavy work, the temperatures of the graphics and memory chips became really worrying. My apprehensions came true: after half an hour of work in the Burn mode, there appeared artifacts on the screen – the card was unstable even at its regular frequencies because of unacceptable overheating.
Let’s get to our anti-extreme business. The regulators of the VPU and memory are based on the same chips as by RADEON 9700 and RADEON 9800. So I just show you a few snapshots where the regulators and the connection points for the additional resistors are indicated:
The card doesn’t have a regulator for the termination circuits: they are implemented in the DDR2 chips from Samsung.
Nominal Voltage, V
Reduced Voltage, V
I couldn’t achieve any better results. One of the reasons for that is probably the complex wiring of the card with long signal lines, while the graphics memory works at high frequency (700MHz).
As a result, the graphics chip voltage became 8% lower and the graphics memory voltage – 7.5% lower (only for VDD). The card remained stable in our tests with these voltages active.
Here is the card just before we started our tests:
Let’s see if there are any advantages of the small voltages reduction at all:
80-85°C under workload seems to be very warm. Let me remind you that the diagram contains temperature values taken not from the graphics processor and memory chips, but from the heatsink surface.
Having reduced the voltages, we managed to drop the temperature of the VPU and memory by 8-9°C in the Idle mode and by 8-14°C in the Burn mode. The result is pretty good, and the card passed all our tests after the modification, but the temperatures are still dangerously high.
It is indisputable that you should install this graphics card into a good system case or provide some extra air cooling. In other words, it’s quite difficult to turn RADEON 9800 PRO 256MB from Sapphire into a “Cool’n’Quiet” graphics card.
So our anti-extreme modification proved to be the most efficient with graphics cards on RADEON 9700 and 9500 graphics processors with 128MB of graphics memory and standard cooling systems. After you switch the cooler to 7V, the temperatures of the VPU and memory grow up by a negligible value, so the overall result of the modification is the card’s becoming both quieter and colder.
RADEON 9600 XT-based card from Sapphire with its passive cooling system works OK without any modifications, since this graphics processor has relatively low heat dissipation. You may want to reduce the voltages of the VPU and memory in this card only if you’ve got a bad system case with improper airflows.
Sapphire RADEON 9800 PRO 256MB is the most powerful and “hottest” device we reviewed today. The DDR2 memory chips from Samsung generate too much heat for their passive heatsinks to handle it efficiently. Their temperature is always higher than that of the memory chips on the other cards, even if they don’t have any heatsinks on. The passive cooling solution from Zalman works at its limit with RADEON 9800 PRO. It’s clear you should set a fan to blow cool air at this card or think about proper airflows in your system case to have no problems with it.
Reduction of VPU and memory voltages make their temperatures notably lower and solve the problem of overheating when there is no additional air cooling. The graphics card stops overheating and works for several hours without any artifacts, although there is no guarantee you won’t see any problems in the future.
So, it is quite possible to apply your own kind of Cool’n’Quiet technology to your graphics card. You can modify it by reducing the VPU and memory voltages or switch the standard cooler to 7V power. You can simply reduce the rotation speed of the standard cooler fan. Or you can buy a passive cooling system like the ones from Zalman and take care of proper airflows generated inside the system case.
The choice of the way to comfort is your own! :)