by Alexey Stepin
01/21/2004 | 11:02 PM
From the X-Bit Labs glossary: “Extreme overclocking – the process of increasing the operational frequencies of the graphics processor and the local memory of a graphics card, usually far above the frequencies recommended by the manufacturer of the card. Extreme overclocking always involves increasing the voltages of the GPU and memory by manual modification of the power circuitry to ensure higher stability of the card at high frequencies. Extreme overclocking may crash the card completely, and by modding the card you automatically lose your warranty”.
Our today’s article is dedicated solely to our extreme overclocking experience with newest and most powerful graphics cards.
As you know, nearly every processor and memory chip can work at a frequency higher (and sometimes much higher) than the nominal frequency, usually after a slight increase of the voltage. In spite of the consequences, enthusiasts go for extreme overclocking to get extra frames-per-second in their favorite 3D games. They modify the power circuitry of the card (volt-modding) to get more voltage for the GPU and memory. Volt-modding adds stability to the card when it is working at high frequencies, but also brings in the risk of sending your card into the other world, since it works in a much harsher thermal environment. An inveterate extreme-overclocker would show you a collection of “dead hardware”, but this only makes the game more exciting. But what if you are lucky and the modded graphics card works alright? How much of the extra performance do you usually get?
This time we decided to experiment with the top solutions available in the today’s graphics market, the latest-generation graphics cards on GeForce FX 5950 Ultra and RADEON 9800 XT. The first guinea pig should be well known to you as we have already described it in detail in one of our previous reviews (see our article called NVIDIA GeForce FX 5950 Ultra against ATI RADEON 9800 XT: Shader Wars).
As for the RADEON 9800 XT, we preferred a SAPPHIRE RADEON 9800 XT ULTIMATE Edition card. What’s so exceptional about this card? It features an impressive cooling system, which simply begged to be tested in and out in our lab. We will start working with it shortly. Right now, take a look at the card as it arrived into our test lab:
A simple package of a moderate size with the picture of an alien is an example of good taste – no gaudy eye-wresting colors, only silver, black and deep-blue. There is a transparent window at the backside of the package through which you can cast a glance inside to see the following items:
The Redline utility from Sapphire deserves a mention in this review. Unlike system software from other companies, Redline asked for a serial number printed on the CD envelope during its installation. After installation, we got access to the following pages of the main window of this program:
This is a curious fact, but Redline from Sapphire works with practically any other graphics card. For example, I ran it on a GeForce FX 5950 Ultra and on an S3 DeltaChrome S8 without any problems, although some settings were naturally missing. What are the benefits of Redline? It depends on your particular needs. If you are a gamer or an overclocker, this utility may please you with its overclocking options and the ability to create setting profiles for each particular application. Otherwise, Redline will hardly be of any use to you. Moreover, it launches at startup to sit in the Taskbar and every icon in the Taskbar means loss of several megabytes of the system memory.
The accessories to the Sapphire Atlantis RADEON 9800 XT are a welcome gift; we’ve got a modern DirectX 9 game, to the bargain. You also receive a Zalman ZM-OP1 fan, which is actually part of the card’s cooling system. Let me explain.
As you see in the snapshots, the graphics card features a massive cooling solution, and this is actually why we decided it would be interesting to try this card in our extreme overclocking competition. A comprehensive cooling solution will guarantee that the card survives our cruel tests safely.
The ZM80C-HP system from Zalman was originally developed to make the card noiseless. The heat from the graphics core is distributed over the big heatsink on the front side of the card and is also transferred by a heat pipe to the back side of the PCB where a similar-looking heatsink is mounted. Thus, the total size of the heat-dissipating surface is 1200sq.cm. That’s enough for dissipating heat produced by a powerful processor like the R360. The heat pipe itself is made of copper, with gold coating. The junctions between the pipe and other elements of the system are glued with thermal paste.
It seems like heat is also taken off the back side of the GPU, but that’s not quite right. The platform there only carries the backside heatsink and has a groove for the thermal pipe. The memory chips are cooled down by small aluminum heatsinks that have nothing to do with the main cooling system. It seems they are too small to perform their function, but that fan from Zalman helps here.
A special thin fan from Zalman, the ZM-OP1 model, is installed on the butt-end of the upper heatsink so that the airflow it generates goes to the two heatsinks cooling down the GPU and to the heatsinks on the memory chips.
The fan is connected via a standard 3-pin connector, so you can plug it right into the mainboard. However, you will hardly like the noise it produces. Zalman, an experienced maker of noiseless cooling solutions, offers a solution to the problem. They include a special adapter with the fan that allows connecting it to 12V or 5V power supply. In the first case, it works at its full speed (about 3,000rpm), in the second – at half of it (1,500rpm). When working at half of its speed, the fan makes no noise, but remains quite efficient at cooling.
The biggest disadvantage of the described cooling system is its enormous bulky size: the PCI slot next to the AGP will certainly be occupied. There is an advantage to it, though. The height of the fan is twice the height of the cooling system, so it can cool the graphics card and the PCI card below. The assembled system looks massive and it is such: the card feels heavy in the hand as it weighs about 700grams (only the heatsinks of the cooling system weight 325grams). That’s quite natural for a Zalman product, as some of the company’s coolers are as heavy as 750grams and more.
I won’t discuss the PCB design, since it doesn’t differ from the reference design (you can refer to our previous reviews for details, namely to the article called NVIDIA GeForce FX 5950 Ultra against ATI RADEON 9800 XT: Shader Wars). Otherwise, this card differs slightly from other products featuring the RADEON 9800 XT VPU. The frequencies for the core and the memory are 412MHz and 365MHz (730MHz DDR), respectively.
The Overdrive technology works smoothly, unlike in some other cards I worked with, but you shouldn’t rely on it much. When it is enabled, the VPU works at 418MHz most of the time. The bonus of 6MHz is ridiculous.
“Manual” overclocking produced more tangible results: the core worked at 445MHz and the memory at 395MHz (790MHz DDR). This RADEON 9800 XT provided good 2D image, just like other cards on this VPU, in resolutions up to 1600x1200@85Hz.
The inherent disadvantage of this card is its clumsy cooling system. If you have many PCI cards in your system, you may encounter certain difficulties during installation of the Sapphire RADEON 9800 XT Ultimate Edition as the fan of the cooling solution will press against large expansion cards. Moreover, as I have said above, you should keep the first PCI slot free. There are exceptions like the ABIT NF7-S 2.0 mainboard, which we used in our testbed. The PCI slots on this mainboard are moved one position away from the AGP, as you remember.
Having taken a close look at the card we found that the power circuitry of the memory on RADEON 9800 XT is the same as we saw on RADEON 9800 PRO, save for ratings of some elements. The memory power supply circuit is still based on the Intersil ISL6522 chip and looks like that:
In order to change the voltage of the current that powers the memory, we need to reduce the resistance of the R2 resistor marked with a circle in the scheme. It has a resistance of 562Ohm rather than 1kOhm as in RADEON 9800 PRO. So we soldered up a trimming resistor with the resistance of 10,000 Ohms to pins 5 and 7 of the chip. Having reduced the resistance of the resistor to 2.5kOhm, we made the memory voltage grow from the nominal 2.69V to 3.12V.
We thought it would be not quite safe to continue increasing Vmem, so we stopped at that. We took the voltage measurements from the C746 capacitor (the point where you should take the measurements is marked as VMem in the scheme). The memory worked stable at 400MHz (800MHz DDR) and we couldn’t go up any further. The bonus of extra 10MHz compared to ordinary overclocking is of little value, but we probably deal with a simplified design of the PCB that hasn’t been designed for work with high-frequency memory chips.
As for the VPU voltage, it was not that simple. The circuit we saw in RADEON 9800 XT had nothing in common with the one of RADEON 9800 PRO, so it took some time for us to perform the modification. As it turned out, the VPU power circuitry on the RADEON 9800 XT card is based around the FAN5240 chip from Fairchild Semiconductor.
A curious fact: according to the manufacturer, this chip is specifically designed for the voltage regulator of mobile Athlon XP processors. Its output voltage can vary from 0.925V to 2.0V by sending different 5-bit combinations to its inputs (VID4-VIA0 pins).
Our measurements showed 1.78V VPU voltage, but the combination at the input of the chip corresponded to 1.4V. It meant that there was another way for changing the VPU voltage, besides changing the bit combination on the input. I didn’t like the idea of cutting and soldering lines on the PCB and had to read the specifications of the FAN5240 to find another way. According to the specs, pins 17 and 18 of the chip (marked as VCore Output Sense) serve for voltage control as well as for protection and monitoring. The typical connection circuit from the manufacturer differed from what we saw on our graphics card: there was an R1597 resistor with a resistance of 1620Ohm between pins 17 and 18. In fact, pin 17 was connected to the common output. So, I decided to connect a trimming 22kOhm resistor in parallel to the R1597.
I soldered one wire from the trimming resistor to pin 18 of the FAN5240 chip, and the other wire to a small bond area next to and attached to the R1597 resistor.
We took our measurements from the point marked as VCore in the scheme.
Our idea was correct, as the Vcore was growing when we were reducing the resistance of the trimming resistor. We stopped on 1.9V, when the trimming resistor stayed at 4.8kOhm. As a result, we made the VPU work at 500MHz. The card was actually stable at 510-520MHz core frequency, but produced some visual artifacts in 3D, so we made it work at 500MHz in our performance tests.
An additional 120mm 9W fan was installed to blow at the butt-end of the graphics card. It pushed air along the entire PCB and cooled down the graphics chip and memory chips similarly. This solved the problem of heat dissipation, although I can’t say the card was cool. An external thermal sensor reading data from the aluminum sole of the graphics chip cooler showed 55-60°C under workload and 46-47°C in the idle mode. For example, the same temperatures before the volt-modding were 46°C and 40-41°C, respectively. The difference is quite perceptible, considering that our sensor was external and the real graphics core temperature was even higher. Here is my advice: don’t save on cooling when you are overclocking your graphics card. I think a water cooling solution might help. A solution like that is noiseless, while our graphics card was not: you wouldn’t work in such a noise, if you are not completely deaf.
Overall, SAPPHIRE RADEON 9800 XT ULTIMATE Edition turned to be not the best choice for extreme overclocking. We’ve got low frequency gains, although spent a lot of time and effort. You will see the performance gains from our volt-modding later, right now we will deal with a solution based on GeForce FX 5950 Ultra.
This graphics card, as well as GeForce FX 5900 Ultra, uses Intersil ISL6569ACR chip in its voltage regulator.
This chip, just like the FAN5240 from Fairchild, allows adjusting the output voltage “on-the-fly” by changing the type of the input signal at the digital VID0..VID4 inputs. We discussed this chip in greater detail in our article about extreme overclocking of the NVIDIA GeForce FX 5900 Ultra and the ATI RADEON 9800 PRO. The graphics card uses the flexibility of the chip to send a voltage of 1.1V to the GPU at startup, 1.2V in 2D and 1.6V in 3D applications. Just like with the GeForce FX 5900 Ultra, we used the OFS input of the ISL6569ACR controller to increase the Vcore here.
According to the technical documentation, when there is a resistor with a resistance R between the OFS input and the “Ground”, the output voltage of the regulator goes up by V=(R*100mkA)/10.
The PCB of the GeForce FX 5950 Ultra carries an R554 resistor with zero resistance, through which the OFS input of the regulator is grounded. To increase the GPU voltage, I soldered up a variable resistor with a resistance of 22kOhm (through wires, for convenience) instead of the zero one.
I increased its resistance thus raising the GPU voltage. I preferred to stop at 10kOhm, which resulted in an increase of 0.1V for the Vcore in all operational modes. In other words, it was 1.2V at startup, 1.3V in 2D and 1.7V in 3D.
Vmem modding was somewhat more complicated. An impulse regulator on the HIP6012CB chip from Intersil supplies power to the internal circuitry (VDD) of the graphics memory chips. The chip is connected in the following manner:
The output voltage of the regulator is defined by the resistances of R2 and R3 resistors according to the formula: V=1.27*(1+R3/R2). To boost the regulator’s output voltage, we can reduce the resistance of R2 resistor by soldering up another resistor in parallel to it.
I used a 5.6kOhm resistor, soldering it through thin wires for convenience:
This modification helped me to increase the voltage of the memory chips internal circuitry from 3.2V to 3.44V.
Input/Output circuitry (VDDQ) is powered by another regulator, based on ISL6225CA chip, which was specifically designed for memory voltage regulators. The output voltage for each channel of this dual-channel regulator is determined by the ratio of resistances in the feedback circuit. This is a simplified scheme of the chip:
The output voltage is determined by the formula: V=0.9*(1+R1/R2). We can increase it by reducing R2 resistance (marked with a red circle in the scheme above). Having found the necessary resistor on the PCB, I attached an additional 4.7kOhm resistor in parallel to it:
As a result, the voltage of the I/O circuitry increased from 2.53V to 2.88V. Now that the microsurgery has been done, I can start overclocking the thing.
We had more luck with this graphics card: 600MHz graphics core and 1010MHz memory frequencies are good, since without volt-modding the GPU was only stable at 515MHz. Memory frequency gain was only 10MHz, but this fact has a simple explanation – the memory chips reached their frequency ceiling.
The testbed was configured as follows:
We tested the graphics cards in the following operational modes:
We used the following benchmarking games and applications:
First-person 3D shooter:
Third-person 3D shooter:
Real-time strategy games:
So let’s now check what advantages extreme overclocking brings to the high-end graphics cards in modern games and applications.
Overclocking brings us small profits in RTCW: Enemy Territory. Volt-modding adds 2fps for the SAPPHIRE RADEON 9800 XT in 1024x768. GeForce FX 5950 Ultra wins the highest resolution, adding 10fps due to volt-modding. ATI’s Overdrive technology gives no perceptible performance advantage. So, we can conclude that this particular game is more dependent on the processor speed, rather than the graphics card performance.
We get different results after turning full-screen anti-aliasing (FSAA) and anisotropic filtering on. RADEON 9800 XT produces the same fps when simply overclocked and when volt-modded. The same happens to GeForce FX 5950 Ultra. Overdrive looks useless again.
Overclocking is more or less efficient in 1600x1200 (5.5fps for RADEON and 5.2fps for GeForce FX). There is no effect from Overdrive technology at all.
Extreme overclocking would hardly make sense if you play OpenGL games. We will discuss Direct3D games in a second.
As you might have guessed, volt-modding is most effective, but there is little difference between overclocked and modded SAPPHIRE RADEON 9800 XT, while the modded GeForce FX 5950 Ultra did much better than the non-modded but overclocked self. The performance of GeForce FX greatly depends on the frequency and the 85MHz difference between simple and extreme overclocking is valuable enough.
RADEON 9800 XT gets more from extreme overclocking when FSAA is enabled. This mode requires more memory bandwidth, and the RADEON card has slower memory than the GeForce FX.
There is certain advantage from overclocking, but it is too little for SAPPHIRE RADEON 9800 XT. Overdrive is practically useless again.
It’s the same in the highest quality mode: overclocking of the GeForce FX 5950 Ultra is more rewarding than that of the SAPPHIRE RADEON 9800 XT.
Halo: Combat Evolved was practically indifferent to overclocking; only modification and extreme-overclocking of the GeForce FX 5950 Ultra ensured a more or less perceptible outcome. 53fps in 1024x768 resolution is nice for this most fastidious and hungry game, but I doubt you should risk your card for the sake of extra 10fps.
Splinter Cell was even more indifferent to overclocking than Halo.
The modification of the GeForce FX 5950 Ultra provides a certain performance boost in this game, while the extreme-overclocking of the RADEON 9800 XT results in nothing.
The same is true for the modes with FSAA and anisotropic filtering. If you own a RADEON 9800 XT and want to speed it up, use traditional overclocking, without any volt-modding. GeForce FX 5950 Ultra also gains little from the modification, it is not worth the trouble.
Volt-modification of the graphics cards again provides a negligible performance gain: RADEON 9800 XT adds up 9fps at best, and GeForce FX 5950 Ultra – 11fps.
The effect from overclocking is even less visible in the modes with FSAA and AF. Overdrive is practically useless.
This flight simulator once again proves that there is not much benefit from extreme overclocking: you get 1-2fps at best. On the other hand, this game is hardly an adequate benchmark.
We see the same picture in the “hard” modes. If you are a hardcore pilot, don’t go for overclocking otherwise you will be really disappointed.
There is no reason to overclock your top-end graphics card, if you play this Formula 1 simulator: there is no performance gain at all in the “pure speed” mode.
There are some extra fps you win from overclocking, but that is not exactly what you have hoped for, yeah?
Overclocking is useless for strategy brains and for C&C admirers in particular. You only get as much as 1fps in 1600x1200 on the modded GeForce FX 5950 Ultra.
The hard operational modes were more favorable to overclocking, but the performance gains are too small for me to recommend you going for extreme overclocking and running the risk of damaging your graphics card.
AquaMark3 is not the kind of application where you could get a boost from overclocking. There is some advantage, of course, but not a big one.
It’s no better with FSAA and AF: the slight performance growth lies in the measurement error range. Thus, AquaMark3 is another item in the list of programs that are indifferent to overclocking.
Modded cards were somewhat faster than their non-modded selves. Overall, overclocking brings no surprises.
We see the same situation with FSAA and anisotropic filtering. In spite of extreme frequencies, the fps values are very low.
It is hard to evaluate the efficiency of extreme overclocking in this test. GeForce FX 5950 Ultra sped up more, but never reached the level of RADEON 9800 XT. RADEON was reluctant to get faster even after our volt-modding.
Some overclockers get so carried away by the competition with their fellows trying to get a higher score in this benchmark. If you are one of them, compare your results to ours:
The modded cards are faster than the non-modded cards with overclocked working frequencies. The RADEON and the GeForce got a similar performance gain, and the first remained the leader.
Now, let’s view the results for each test:
Nothing new compared to gaming benchmarks: extreme overclocking is just a little more profitable than simple overclocking.
Nothing changes in the hard mode, although the modded GeForce FX 5950 Ultra outperforms the modded RADEON 9800 XT.
Extreme overclocking adds only 2-4fps over ordinary overclocking.
After FSAA and anisotropic filtering have been enabled, the gain from overclocking is often less than 1fps.
That’s the same situation as we have just seen in the second gaming 3DMark test.
With FSAA and anisotropic filtering enabled, the gain from overclocking is often less than 1fps, sometimes 1 or 2 fps.
The Mother Nature test is a big problem for any graphics card that is poor at processing pixel shaders version 2.0. Sky-high frequencies can’t help GeForce FX 5950 Ultra here. Extreme overclocking brings in little advantage again.
Overclocking adds some speed to the cards when FSAA and anisotropic filtering are on, but they still perform slowly. Through extreme overclocking RADEON 9800 XT managed to hit the 30fps mark, though.
So our today’s tests confirmed a well-known truth: overclocking, however extreme, doesn’t cure low performance where the performance is basically low.
There are two cases with overclocking top-end graphics cards. In old games you achieve a bonus of tens of frames per second, but the total fps value is measured in hundreds. So it doesn’t make much sense to overclock the card – you can’t tell 120fps from 140fps. There is only one exception – heaviest scenes when the speed can go down to 20-30fps. Overclocking may help here. For the modern games, there is another pitfall: overclocking usually adds 2-5fps to the total of 30-50fps. That’s too little and is certainly not worth the trouble of overclocking the card.
The heavily-advertised Overdrive technology from ATI proved to be absolutely useless in practice. That’s because of the very principle behind this technology: only the VPU is overclocked, while the memory chips remain at their nominal frequency. Moreover, this technology depends on the VPU temperature, and the threshold values are set up in such a way that the graphics core nearly always works at 418MHz. That’s a negligible frequency gain that provides a performance gain of the same small proportion.
As for extreme overclocking, I think the effort you spend and the risks you run are not worth the results you get (at least, for these two cards). However, extreme overclocking of the GeForce FX 5950 Ultra appears more efficient than that of RADEON 9800 XT due to the higher overclocking potential of the 0.13-micron graphics core. RADEON 9800 XT gains little from the modification, but this VPU is quite powerful as it is. Of course, our article won’t stop a hardcore overclocker, but if you are just considering taking a soldering iron and voltmeter into your hands, you’d better think twice as you will probably get a few extra fps, but also more noise and heat.
Analyzing our today’s overclocking experience, I would point at a few performance peculiarities of the current graphics processors that cost about $500.
The ultimate edition of RADEON 9800 XT from Sapphire uses the tested and reliable reference design from ATI Technologies and carries a unique and efficient cooling system. If you need performance and silence, consider this card. The cooling system may make it interesting for overclockers. The appearance and accessories of the product are also up to the mark, and I recommend it to anyone who’s shopping for a modern and fast graphics card.