by Tim Tscheblockov
10/07/2003 | 11:01 PM
It is no secret to anybody who keeps up with the GPU market situation that R360 and NV38 chips from ATI and NVIDIA are nothing more but “overclocked” versions of their predecessors, NV35 (NVIDIA GeForce FX 5900 Ultra) and R350 (ATI RADEON 9800 Pro). ATI has been steadily pushing up the frequency of the good old R300, while NVIDIA has “officially” overclocked a relatively new chip.
In our today’s article we will try to get ready to the arrival of NV38 and R360 by estimating the overclocking potential of NVIDIA GeForce FX 5900 Ultra and ATI RADEON 9800 Pro. We will squeeze all their fps out of them by the well-known method – increase of the GPU and graphics memory voltages.
For our today’s tests we took an MSI graphics card based on the NVIDIA GeForce FX 5900 Ultra GPU from the NBOX kit and a PowerColor R38-C3 card featuring ATI RADEON 9800 Pro.
The gorgeous kit from MSI called NBOX will represent NVIDIA GeForce FX 5900 Ultra based graphics cards in our today’s review. NBOX is designed stylishly and laconically: you will not see any of those beauties or monsters or terminators or any other hi-tech embellishments on the box:
The paper skin conceals an even more taciturn package – a black box with a single “NBOX” logotype on the lid:
The graphics card lies in its bed of microcellular rubber. It sits quite tight, but you will have no problems removing it, because the bed has special cuts for your fingers:
The rest of the kit lies in the bottom section of the package. In order to access it, you take up the section with the card. You can see a strap made exactly for this purpose in the snapshot (on the left). When you lift the card’s case up, you will hear a sound of the outgoing sticker that has been keeping it closed.
So, what do we have here? Let’s see:
The first thing my searching eye caught was the optical USB mouse with MSI’s logo and a Command & Conquer: Generals mouse pad.
Let me say this from the very beginning: MSI’s mouse with its aluminum buttons and blue highlighting of the wheel looks cool, but it is not that handy in games. I guess hardcore gamers will prefer to go on using their own carefully selected mice with them. As for the pad, it is just perfect: thin, handy, with a special anti-slide covering.
Next, we have CDs with drivers and utilities from MSI as well as three full games: Unreal Tournament 2, Battlefield 1942 and Command & Conquer: Generals. Each game comes with a booklet explaining the way to play.
Lastly, the package contains cables and adapters: an S-Video cable, an adapter from the composite connector to 2 RCA and 2 S-Video connectors, a DVI-to-D-Sub adapter and an auxiliary power cable.
Now, let’s say a few words about the card itself. The core of the NBOX kit is a standard FX5900U-VTD256 model based on NVIDIA GeForce FX 5900 Ultra:
The cooler located at the front side of the PCB covers both: the GPU and the memory chips. The air pumped in by the fan is then directed along the ribs of the heatsink that form a radial pattern. A layer of thermal paste helps to improve the thermal contact between the heatsink and the chips, which it is cooling.
It is often hard to make sure that a monolithic cooler like that sits tightly on all the chips. However, in our particular card, the cooler did press against each chip on the front panel quite nicely.
The backside cooler consists of two parts. One is responsible for taking the heat off the graphics memory chips, while the other does the same with the PCB spot right beneath the GPU. The contact to the PCB is achieved by means of a thick and stiff thermal pad:
I doubt this cooler makes a great contribution to the cooling of the graphics processor, though. Anyway, it is good because it produces an air stream to blow at the ribs of the heatsink installed on the memory chips:
It would be more useful if this heatsink had a little bit of thermal paste for better contact with the chips. As it is, it simply presses against them. Well, this is a minor drawback. The card works stably at its regular frequencies, and if you are into overclocking, you can find some paste. :)
The last thing about the cooling system is the metal it is made of. All the heatsinks on the card look as if they were made of copper, but they are actually produced of some aluminum alloy. I have made a small cut with an ordinary paper knife and you can see a metal of another color:
The panel of the card carries D-Sub, DVI-I and composite connectors.
At the opposite side of the PCB, there is a connector for additional power supply. We see it on any GeForce FX 5900 Ultra based card.
Once again, the card features the NVIDIA GeForce FX 5900 Ultra GPU and has 256MB of graphics memory onboard. The memory chips are made by Hynix and have 2.2ns cycle time:
The regular frequencies of the card in the 3D mode are standard: 450MHz for the graphics chip and 850MHz (425MHz DDR) for memory.
The special unit integrated into the GPU shapes up the television signal for the video-out, while the popular SAA7108 chip from Philips is responsible for decoding it. The Sil164CT64 chip from Silicon Image supports the digital output (DVI).
So, MSI’s FX5900U-VTD256 model included with the NBOX kit is a well-made graphics card following the reference design. This product is distinguishable from other graphics cards for its original looks, quality cooling solution and very low noise level. Its voice practically dies away against the roar of other system components.
As for the package and accessories, I think that top-end graphics cards must come packed like the NBOX from MSI. When you count out a few hundred dollars, you wish to have something more than a standard set of the card, a couple of cables and a CD with drivers – all in a tasteless carton, don’t you?
Winding up this section, I’d like to mention the overclockability of the card. The FX5900U-VTD256 worked stable at 570MHz GPU and 950MHz (457MHz DDR) memory frequencies. This is a good, but not very impressive performance. But this is only the beginning! :)
The GPU voltage regulator is based on the ISL6569 chip from Intersil. This chip is an “intellectual”, pulse-width controller for a two-phase impulse voltage regulator. It features protection against voltage and current surges in demand, “soft activation”, balancing of currents in the channels, temperature stability of the output voltage <1% and so on and so forth. According to the specifications, it is recommended for the use with AMD Hammer processors.
Well, this is actually the first time I ever meet a two-phase voltage regulator in a gaming graphics card. They are usually seen in mainboards as regulators for powerful CPUs. Well, if the NVIDIA GeForce FX 5900 Ultra chip does consume a lot of power and requires highly stable power supply, the two-phase regulator is a must. But this is not the only discovery…
The output voltage of the regulator is set by sending a binary code to the DAC inputs of the chip:
Setting the output voltage by means of a binary code is nothing new, but the peculiarity of the ISL6569 chip is its ability to adjust the voltage “on the fly”. When the input data are changed, the controller smoothly increases or decreases the output voltage of the regulator to the new value.
It’s clear that such a “clever” regulator with a variable output voltage is not just for show. Watching the GPU voltage, I found that it really changes depending on the work mode:
Thus, NVIDIA engineers have endowed GeForce FX 5900 Ultra based graphics cards and the Detonator driver with the ability of controlling the GPU power supply voltage.
Why does the voltage increase in 3D? As the GPU and memory frequencies go up in 3D applications, we may suppose that the higher voltage helps the graphics card to be stable at higher clock-rates. But this doesn’t seem to be the case. It’s rather otherwise: the voltage goes down in 2D compared to 3D (as well as the frequencies do). And this is done for the card to save energy and generate less heat when it is not required to show miraculous speed.
As it turns out, it is only necessary to change a part of the driver’s code to increase the voltage in the 3D mode, or program the regulator itself. I guess we will soon see utilities that allow adjusting the voltage of GeForce FX 5900/5900 Ultra based cards. For now, I have to look for a hardware way: I’m not too much into writing drivers to make such a utility myself. :)
The card’s voltage regulator suits perfectly for an external intervention: the ISL6569 chip has a special OFS input, connected to the current source and a divisor. By connecting a resistor between the OFS input and the common wire, it is possible to increase the output voltage of the regulator above what is set by the binary code at the DAC input:
The voltage offset value when a resistor with resistance of R is connected equals (100µA*R)/10.
The OFS output of the ISL6569 chip is connected in the MSI card with the common wire through a chain of resistors with zero resistance. Thus, it is easy to increase the voltage by simply substituting one of the links with a resistor with the desired resistance. I soldered wires instead of the resistor for better convenience:
And then soldered a variable resistor (with 33kOhm nominal) to the ends of the wires:
During extreme overclocking experiments, I found acceptable a resistance of 30kOhm. Thus, the GPU voltage in the 3D mode was 1.69V, that is, it grew 0.29V higher.
You should keep in mind that the voltage offset doesn’t depend on the input data sent to the DAC of ISL6569. That is, the GPU voltage is increased in all operational modes. Accordingly, the heat dissipation increases calling for better cooling of the graphics processor.
We will deal with cooling and overclocking a bit later. Right now, let me complain to you about my memory overclocking experience.
After examining the memory voltage regulators, I raised the voltage of the internal circuitry and I/O buffers (VDD and VDDQ) by 0.2V. But overclocking was a complete disappointment. The maximum memory frequency the card could give out remained at the same: 950MHz (475MHz DDR).
Increasing the voltage by 0.2V more, I had a second try. It was the same: 950MHz.
I did everything I could think of – increased and decreased the voltage around the nominal, tried to vary VDD and VDDQ independently, then took the experiments over into the heat chamber and below-zero temperatures. Everything was in vain; the maximum frequency remained 950MHz.
At last, I gave up the idea of increasing the graphics memory voltage. It looks like 950MHz is the limit for the NVIDIA GeForce FX 5900 Ultra reference design and Hynix memory chips.
So, the card is ready for the tests:
We only have to replace the graphics chip and memory cooling solution with a more efficient one.
The NVIDIA GeForce FX 5900 Ultra graphics chip has a copper casing. It is mounted on the die and protects it from chipping as well as distributes the heat along the surface of the heatsink. However, to get the best from overclocking, we’d better put off everything that’s in between the chip and the heatsink. So, I took the GPU casing off:
This operation was performed with a sharp paper knife. Those, who venture upon such an experiment despite the warranty, should better be VERY careful. The multi-layer wafer of the chip has “fragile” current-conducting paths that are too easily damaged with an unsteady hand.
So, here is the naked NVIDIA GeForce FX 5900 Ultra in person:
The white smudge in the screenshot is the thermal interface (KTP-8 paste) between the die and the water-unit. To take the heat off the GPU, I decided to use Thermaltake Aquarius II water-cooling solution. Because of the water-unit, I had to abandon the standard cooler, which had covered the memory chips, too. For the memory chips, I used heatsinks from an NVIDIA GeForce FX 5900 Ultra based card designed according to the reference.
The assembled system with the MSI card installed looked like this:
In order to push up the overclocking potential of the card, I took the system to a heat chamber that maintained a constant temperature of -15oC throughout the tests.
So, MSI’s card featuring the NVIDIA GeForce FX 5900 Ultra GPU is ready for the tests. Now, we will try to make up a good competitor to it by modding a RADEON 9800 Pro based graphics card.
Unlike MSI’s NBOX, the RADEON 9800 Pro based card from PowerColor comes in a traditionally colored paper box:
The box includes the card, CDs with drivers and utilities, demos of computer games and a full version of Summoner. There are also S-Video and RCA cables, a DVI-I-to-D-Sub adapter and a cable for auxiliary power supply:
The PowerColor R98-C3 card follows the reference design and thus resembles any other RADEON 9800 PRO based product:
A standard aluminum cooler is mounted on the graphics chip. Its only peculiarities are the protrusion in the sole for better contact with the die and the unusual shape of the fan blades. The blades are bended so that one part of the blade works as a vane in a centrifugal pump, while the other as an ordinary blade of an ordinary cooler:
This cooler is quite noiseless, at least no worse than the two coolers used on the MSI card.
The card carries the ATI RADEON 9800 Pro graphics chip and 128MB of DDR SDRAM graphics memory from Samsung with 2.86 cycle time:
The regular frequencies of the card are 380MHz graphics chip and 680MHz (340MHz DDR) memory.
The card is equipped with DVI-I, D-Sub and TV-Out connectors. The integrated units of the R350 processor are responsible for outputting the digital and television signals:
As all RADEON 9800 Pro based cards, the one from PowerColor has a connector for additional power supply:
The maximum frequencies the card reached during regular overclocking were 460MHz for the chip and 740MHz (370MHz DDR) for the memory. Not much. But this is with the standard cooling system and without any voltage regulator modifications.
Well, no one prevents us from doing this modding, the warranty be damned. :)
The graphics chip voltage regulator is based on the dual-channel pulse-width SC1175CSW controller from Semtech. This time it works in the current sharing mode, when both channels serve one demand:
The nominal graphics core voltage on the PowerColor card is 1.7V.
We can increase the voltage at the output of the regulator by introducing an additional resistor. It is marked with blue in the scheme. I used a variable resistor with a resistance of 33kOhm, soldering it to pins 18 and 20 of the SC1175CSW chip. In order to prevent any regulator damage or graphics core damage by setting too low or zero resistance, I also switched on a “rescue” resistor with resistance of 1.5kOhm in parallel. You can see it in the snapshot:
After a few experiments with the output voltage of the GPU voltage regulator, I decided to stick to 1.94V value.
After experimenting with the regulators on the PowerColor card, I found that the memory chips could only be sped up when the inner circuitry (VDD) voltage was increased, while tweaking the input-output circuitry (VDDQ) didn’t practically affect the overclocking results.
The VDD regulator in PowerColor’s card is an ISL6522 chip from Intersil. The typical connection circuit for this chip is shown below. I also marked the resistor, which resistance should be reduced to increase the output voltage:
For details, please refer to the description of the chip on the manufacturer’s website.
Again, I used an additional resistor with a resistance of 2.7kOhm, connecting it with wires to pins 5 and 7 of the ISL6522 chip:
Thus, I increased the output voltage of the regulator from 2.95V to 3.35V.
I had to make heatsinks for memory chips cooling myself. Well, in fact I just took a chipset heatsink, sawed it into several pieces and polished all burrs off:
Sticky thermal pads from the Thermaltake Memory Cooling Kit were used to fasten the heatsinks onto the memory chips. Well, I know there are ways to make a better cooling, but still it is better than nothing. :)
The same Thermaltake Aquarius II water cooling solution was used for cooling the RADEON GPU. This is how the card looked just before the tests:
And here’s the testbed with PowerColor RADEON 9800 Pro based card installed:
The card was tested under the same conditions as the MSI NBOX: in the heat chamber at -15oC:
So, we have our cards ready, let’s get to the tests…
Our testbed configuration looked as follows:
We used the following software:
The anisotropic filtering settings and texture level of detail were set in the drivers as “quality”. All the tests were performed with my own demo record made in Unreal Tournament 2003, the DM-Inferno level.
The graphics cards were only tested in 1600x1200 resolution. At such a high resolution, the CPU and the system don’t practically limit the performance and we will have the maximum relative performance gain achieved from overclocking the graphics card.
I ran Unreal Tournament 2003 in 32-bit color and with maximum graphics quality settings: Texture Detail – Highest, World Detail – Highest, Character Detail – Highest, Physics Detail – High, Character Shadows – ON, Dynamic Lighting – ON, Detail Textures – ON, Projectors – ON, Trilinear Filtering – ON, Decals – ON, Coronas – ON, Decals Stay – High, Foliage – ON, Use Blob Shadows – OFF.
The maximum frequency the NVIDIA GeForce FX 5900 Ultra chip on MSI’s card could handle during extreme overclocking was 675MHz. This seemed an outstanding achievement and a new record!
However, I was not joyous for long. Having tested MSI card at different graphics chip frequencies, I got very curious results. Below you can see the graph showing the relation between the graphics core frequency and the card’s performance at a fixed memory frequency:
As you see, when the GPU clock-rate grew above 625MHz, the card ran more slowly than at its regular frequencies. Moreover, the results don’t change at higher graphics chip frequencies.
It is known that NVIDIA GeForce FX 5900 / 5900 Ultra chips have integrated thermal diodes. It seems like the temperature of the GPU is too high at frequencies above 625MHz (in spite of our water-cooling solution) and, following the instinct of self-preservation, the driver drops the frequencies of the card even below the standard 3D mode ones. This is also evident from the GPU voltage value. It should be 1.4V plus my addition, while at frequencies above 625MHz it is actually 1.2V + my addition.
In some cases, the graphics card was going through the test at the overclocked core frequency, but then would hang up for a moment, lower the GPU voltage and drop down to lower working frequencies. It continued through the rest of the test then, but certainly with unsatisfactory results.
Thus, although the card proved to be operational at 675MHz GPU frequency, it did best at 625MHz graphics chip working frequency. This frequency will be considered the maximum GPU overclocking result.
Now, let’s estimate the gains we achieve by overclocking NVIDIA GeForce FX 5900 Ultra in various work modes. I drew a graph showing the performance increase from overclocking as compared to normal frequencies. I also put down the gain of the GPU frequency.
As we have expected, the graphics core overclocking is most advantageous in the modes that use anisotropic filtering. GeForce FX 5900 Ultra spends many clock cycles of its pixel pipelines to perform AF and a higher GPU frequency gives it a boost.
When full-screen anti-aliasing is enabled, we have a smaller gain. FSAA dramatically increases the memory bus workload, but we have kept the memory frequency intact here.
Overall, the performance gain was twice lower than the GPU frequency gain even in the most “responsive” mode, about 20% at maximum. So, overclocking the GPU only makes no sense: let’s speed up the memory.
There were no problems with overclocking the memory o MSI card. The maximum frequency I hit was 950MHz. After that, the test scene transformed into a chaotic commotion of polygons of every color and size imaginable. The graph below shows the performance of NVIDIA GeForce FX 5900 Ultra with an overclocked memory and a fixed GPU frequency:
Well, the performance gains are rather small. Let’s view them in percents and compare to the graphics memory frequency gain:
Again, there are no surprises. We have a maximum feedback from memory overclocking in FSAA modes, and minimum – in AF modes.
Overall, the graphics memory proved less willing to speed up than the GPU. We’ve got only 5% performance gain at maximum.
When overclocking both the graphics core and memory to their maximum, 625MHz / 950MHz (457MHz DDR), we have of course higher results than if speeding them up separately:
The diagram below shows the relative performance gain in different work modes. I also put down the graphs of the relative GPU and memory frequency growth.
So, extreme overclocking made the NVIDIA GeForce FX 5900 Ultra card run about 20-30% faster. Overclocking brings similar advantages in every work mode, which indicates a good “balance” of the product. The maximum performance gain was in 1600x1200 with FSAA and AF enabled – as this was the hardest operational mode for the card.
Unlike NVIDIA GeForce FX 5900 Ultra GPU, the graphics processor from ATI and cards that have it don’t offer hardware monitoring capabilities or a tricky power supply circuit. Thus, they behave quite predictably during overclocking. From a certain frequency up, the screen is all littered with “lost” pixel blocks. If you don’t mind this, but keep on driving the core frequency up, the system will hang up. The ATI RADEON 9800 Pro chip on the PowerColor card worked stable up to a frequency of 540MHz. The diagram below offers you the test results for GPU overclocking while the memory frequency remained fixed:
As you see in the diagram, the card performs poorer at 560MHz GPU frequency than at 540MHz. It is just because the GPU in the PowerColor card doesn’t yet hang up at 560MHz, but already makes a jumble of the screen image. Such a work mode quite naturally doesn’t tell well on performance.
Here is the graph of the relative performance gain against the relative GPU frequency gain:
The ATI RADEON 9800 Pro graphics chip is faster during anisotropic filtering than NVIDIA GeForce FX 5900 Ultra. However, overclocking was most effective in modes with enabled anisotropic filtering, too. A higher GPU frequency isn’t much of a boost for the PowerColor card in FSAA modes.
Overall, the effect from the graphics core overclocking is comparable to the “responsiveness” of NVIDIA GeForce FX 5900 Ultra: 20% at maximum.
PowerColor graphics card turned to have a speedier memory than MSI NBOX. The maximum memory frequency we managed to obtain and when the card worked fine was 840MHz. Below are the benchmarking results with the overclocked memory and a fixed GPU frequency:
And here is the relative performance gain:
My overclocking the graphics memory of the RADEON 9800 PRO based card was less fruitful than it was in case with the NVIDIA GeForce FX 5900 Ultra card. Most probably, ATI GPU has more than enough of memory bandwidth, even though it has eight pixel pipelines compared to four by NVIDIA chip. This is all thanks to its HyperZ III technology, which includes an efficient texture-caching algorithm as well as algorithms for work with the pixel units and Z-buffer. But again, because of this efficient HyperZ III, memory overclocking results into a smaller advantage than expected.
The highest performance gain is when full-screen anti-aliasing is enabled. The relative performance gain is then three times smaller than the relative memory frequency gain. In other cases, the performance gain lags behind the frequency growth by four times.
When I overclocked both the graphics processor and memory, I got a very perceptible speed boost:
This is how it looks in percents:
ATI RADEON 9800 PRO was overall better at overclocking than NVIDIA GeForce FX 5900 Ultra, yielding from 24% to 31.7% speed increase.
The results suggest that the overclocked ATI RADEON 9800 PRO card shows its best with enabled anisotropic filtering. In this case, the performance of the graphics card is mostly determined by the core speed, and we have enough of this – extreme overclocking gave us a 42.1% GPU frequency gain.
As for full-screen anti-aliasing, the overclocked RADEON 9800 Pro card didn’t have much to offer. The memory frequency gain was small compared to the GPU frequency increase and, moreover, this card is overall rather indifferent to memory overclocking.
Summing it up, I would say that ATI RADEON 9800 Pro is less “balanced” compared to NVIDIA GeForce FX 5900 Ultra, but this doesn’t prevent it from showing a higher performance boost when both, the graphics core and memory, are overclocked.
Now, let’s have a final reckoning. First comes the NVIDIA GeForce FX 5900 Ultra graphics card in MSI’s implementation:
The results of the ATI RADEON 9800 Pro based graphics card from PowerColor:
Overall, the ATI RADEON 9800 PRO based card from PowerColor showed a higher performance gain as a result of overclocking than the NVIDIA GeForce FX 5900 Ultra based card from MSI. However, if we compare the results of the two cards, this advantage will seem rather small:
The MSI card outperforms the PowerColor one in the “Quality” mode and with FSAA enabled. The same is true for extreme overclocking, in spite of a higher performance gain shown by ATI RADEON 9800 Pro.
The RADEON 9800 PRO performs anisotropic filtering faster, although with lower quality. That’s why the PowerColor card is ahead of the MSI one in the modes with enabled anisotropic filtering. Extreme overclocking only makes the gap wider.
So, the two competitors – MSI NBOX and PowerColor R98-C3 – proved to be worth each other in our extreme overclocking race. The performance gain both cards provided was about 20% - 30%. It is a solid surplus above the ordinary frequencies of those cards, and the upcoming “officially overclocked” chips – R360 and NV38 – will hardy offer anything like that.
As our testing showed, both: the 0.13micron GeForce FX 5900 Ultra chip and the ATI RADEON 9800 Pro made with an already out-dated 0.15micron technology, have a good reserve of frequency left. By the way, NVIDIA’s transition to the 0.13micron process doesn’t necessarily mean that 0.15micron is completely obsolete. The excellent results the RADEON 9800 PRO chip showed today are a striking proof of the opposite.
But back to our cards. The PowerColor product offers higher overclockability of the GPU and memory and exceeds MSI card in the performance gain value. Moreover, ATI RADEON 9800 PRO behaves much more predictably during overclocking.
MSI NBOX card loses somewhat in terms of overclocking potential, but makes up for it with its gorgeous package, elegant looks and rich accessories set.
And if we take into consideration only the benchmarking results shown in the last diagram, we would come to a conclusion that both cards are in fact peers. Of course, there is no reason to choose any of them basing on the results of one gaming benchmark only. Well, this review was not about comparison and choosing between NVIDIA GeForce FX 5900 Ultra and ATI RADEON 9800 PRO, it was about extreme overclocking! That’s why the following note seems necessary:
Reminder:
This link takes you to the Unreal Tournament 2003 demo record I used in the tests: tim_demos.rar (616KB).