by FastSite
07/20/2002 | 12:00 AM
In this article we reviewed the most advanced and expensive creation of the VisionTek company: the VisionTek Xtasy GeForce4 Ti4600 graphics card based on the most powerful gaming chip available today, NVIDIA GeForce4 Ti4600. <%BANNER[article]%>
Keeping in mind the successful experience we had with extreme overclocking of VisionTek Xtasy GeForce4 Ti4400 (see our review), we will try to answer the question: is the trouble worth it? How high will the performance growth of the today's fastest graphics card be and does it make sense to take chances and subject cards like that to extreme overclocking torture?
The retail VisionTek Xtasy GeForce4 Ti4400 graphics card is shipped in the same paper box as VisionTek GeForce4 Ti4400:
Besides the card itself, the package contains CDs with drivers and utilities (Cyberlink PowerDVD and PowerDirector), user's manual made as a folding booklet, and a VIVO joint plug-and-socket splitter for two S-Video Outs:
The sticker on the fan tells that the graphics card is manufactured by VisionTek, in all other respects the card looks as usual. VisionTek Xtasy GeForce4 Ti4600 follows the NVIDIA reference design and is equipped with VGA Out, DVI-I and Video-In/Video-Out.
The Video-In/Video-Out support is implemented via the widely spread Philips SAA7108 chip, which is installed on the front side of the PCB.
The "heart" of the graphics card is the today's most powerful NVIDIA's creation: the GeForce4 Ti4600 chip working at 300MHz frequency:
The card has 128MB of graphics memory. These are DDR SDRAM chips from SAMSUNG with 2.8nm access time. The graphics memory works at 650MHz (325MHz DDR):
So, here we are with the real "monster" of 3D gaming graphics. Let's see what it has to show us in case of extreme overclocking. Before we start, let's introduce some changes to the graphics card cooling.
To take the heat off during extreme overclocking experiments we need something more powerful than a standard graphics core cooling solution.
The Thermaltake Mini Copper Orb cooler suited us best. We had already disassembled it and used for our extreme overclocking investigation of ATI RADEON 8500 and VisionTek Xtasy GeForce4 Ti4400:
Before installing the cooler, we had to work on the surface of the graphics core. We polished it with fine sandpaper so that it became absolutely flat, and not concave:
The concavity of the copper plate covering the core showed itself quite curiously: its "deepest" spot was right above the die. So after sandpapering the spot, the metal layer covering it remained unerased. It can be clearly seen in the snapshot.
The cooler was stuck to a layer of the KPT-8 grease made in Russia. Then the ribs were tightly pressed to the card by putting nylon bands through the holes in the card PCB.
I didn't install heatsinks onto the graphics memory chips. To be exact, I did try to install them, but as practice showed, installing heatsinks onto the memory chips didn't have any positive effect on overclocking performance. However, if the card is placed in a case with poor airflow, the heatsinks will be essential, though. Every memory chip of the kind may produce up to 2W of heat and it's not that little, we should say.
After all the sufferings the VisionTek Xtasy GeForce4 Ti4600 card turned looking like this:
Maximum frequencies we managed to make our VisionTek GeForce4 Ti4600 graphics card work at stably during ordinary overclocking (i.e. without raising the voltage), were 315MHz for the chip and 750MHz (375MHz DDR) for the memory.
VisionTek Xtasy GeForce4 Ti4600 as well as all the NVIDIA GeForce4 Ti4600 based graphics cards following the reference-design are provided with SC1102CS voltage regulator from Semtech responsible for the power supply of the graphics core. The chip is a controller for pulsed power supply units and DC/DC commutators. Output voltage of the regulator is set by the resistance of the resistors in R8/R7 devisor (in the typical application circuit scheme) and calculated according to the following formula: Vout=Vref*(1+R8/R7). Here Vref is reference voltage, which equals 1.256V for SC1102CS. The resistor numeration in the formula is taken from the typical application circuit scheme:
To raise the graphics core voltage it's easiest to reduce R7 resistance (in the typical application circuit scheme) by shunting it with a parallel resistor. That's what we did: soldered a 1kOhm resistor to pins 11 and 14 of the SC1102CS chip:
So, the R8/R7 ratio (in the typical application circuit scheme) changed and the regulator's output voltage grew from 1.66V to 1.81V.
During the graphics card torture, we raised the graphics core voltage even higher, up to 2V, but it didn't bring any additional clock-rate growth during overclocking. 340MHz is the maximum frequency the graphics core worked at during our extreme experiments. By the way, the maximum frequency we managed to achieve when overclocking NVIDIA GeForce4 Ti4400 chip was absolutely the same. 340-350MHz is evidently the top frequency possible, determined by the inner architecture and manufacturing technology of NVIDIA GeForce4 Ti4600/Ti4400 chips.
In the end, I decided to stop at a rather small voltage increase: 9% only. Well, what's the purpose of making the graphics core work at full stretch if the terminal rate can be achieved by a very small increase in voltage?
Graphics memory power supply is implemented by means of the SC1175CSW regulator. Like SC1102CS, it's a pulse regulation controller, but with two independent channels. By NVIDIA GeForce4 Ti4600 based cards, one channel of the regulator provides power for the internal circuitry and graphics memory I/O buffers (VDD/VDDQ); and the other channel regulates the impedance load voltage (Vtt) for sync circuits. The sync circuitry layout can be found in the documentation for DDR SDRAM graphics memory chips from Samsung, which are used in NVIDIA GeForce4 Ti4600 cards:
Interesting, that of all NVIDIA GeForce4 Ti/MX graphics cards, only NVIDIA GeForce4 Ti4600 (i.e. "fastest") video cards have sync circuitry like that.
The situation with the Vtt value appeared very curious: in theory, the increase in the graphics memory chips voltage should lead to the proportional increase of the Vtt value (Vtt should be equal to 0.5xVDDQ). However, I found no positive or negative effect of proportional increase of the Vtt value on the functional stability and reaching the maximum frequency. So, I decided not to complicate the issue and leave Vtt alone. As for Vref, there are no problems about it at all. You can derive it immediately from the VDDQ by simple resistor devisors for every graphics memory chip.
As the voltage of internal circuitry and I/O buffers of the chips used in NVIDIA GeForce4 Ti4600 coincide, VDD and VDDQ are connected with one another and the chips power supply is provided via the single channel of the SC1175CSW regulator. The output voltage of the regulator is set by the resistors ratio in the devisors of the following formulas: Vout1=Vref*(1+R15/R14) and Vout2=Vref*(1+R13/R11), where Vref for SC1175CSW equals 1.25V and resistor numeration is set according to the typical application circuit scheme:
The default voltage of the graphics memory chips is 2.83V.
Vout1 was increased with the help of R14 720Ohm resistor, which was soldered to pins 18 and 20 of the controller chip.
The graphics memory voltage increased up to 3.09V. When I tried to raise graphics memory voltage just a little higher, the card lost stability, there were ripples and falling-out pixels - even during the start-up process. So I decided to stop at 3.09V: the graphics memory voltage was only raised by 9.2%.
The maximum graphics memory frequency when the card remained stable was equal to 800MHz (400MHz DDR).
So, after the modification the top frequencies of the graphics core and memory by VisionTek Xtasy GeForce4 Ti4600 were 340MHz and 800MHz (400MHz DDR) respectively. To tell the truth, this is quite good if we bear in mind that the voltages of the graphics core and memory were only raised by 9%. The chip frequency growth was only 13.3% while the graphics memory frequency was increased by 23.1%.
Our test system was configured as follows:
We used the following software:
The tests were run in the following modes:
3DMark 2001:
32bit frame buffer, 32bit textures, 32(24)bit Z-buffer, d3D Pure Hardware T&L.
Codecult Codecreatures:
We tested with default settings.
In this review we decided to limit ourselves to testing in the "hardest" benchmarks only - Codecult Codecreatures and 3DMark 2001 Nature. For a more detailed review of the functioning of overclocked NVIDIA GeForce4 chips and the dependence of their performance upon core and graphics memory frequencies, refer to our article "VisionTek Xtasy GeForce4 Ti4400 Extreme Overclocking Experience".
The performance growth after extreme overclocking is about 20%.
In Codecreatures the performance growth was lower, which is actually not at all surprising: during extreme overclocking of VisionTek GeForce4 Ti4400 we have already found out that this test is more sensitive to the growth of the graphics chip frequency than to that of the memory frequency. The chip frequency was extremely overclocked by only 13.3%.
The 3DMark 2001 - Nature test is more dependent on the graphics memory bandwidth, so the extreme overclocking results were higher as the graphics memory frequency got 23.1% higher:
Well, the results of extreme overclocking of VisionTek Xtasy GeForce4 Ti4600 can hardly be called outstanding.
The reason is that the nominal frequencies of the graphics chip and memory are close to the top frequencies, which are determined by the chips architecture and manufacturing technology. It's no use even to hope for a good overclocking potential in such circumstances.
The graphics core is especially hard to overclock: 13% frequency growth is too little compared with what we had at the time of NVIDIA GeForce2/GeForce3. Evidently, when creating GeForce4 Titanium series NVIDIA engineers were trying hard to squeeze the maximum out of what they had: the modified NVIDIA GeForce3 core and 0.15micron chip manufacturing technology. So, the poor overclocker has been left little to himself.
The situation is very similar to what happens in the processor market: among processors on the same core, low-end models have higher overclockability than top-end ones. The development of the family towards higher clock rates (so to say, "official" overclocking performed by the manufacturer) leads to the reduction of the frequency potential used by overclocking fans. And only when new processor cores are rolled out, the overclocking epidemic rekindles with a new vigor and doesn't go out until the reserve is used up again…
It seems to be the same with graphics cards: "free cheese" lovers are biding their time till new chips from NVIDIA, ATI, Matrox and others come out…
We'd like to point out the high quality and reliability of the VisionTek Xtasy GeForce4 Ti4600 card, which has successfully made it through most subtle tortures, graphics cards had ever suffered in our test lab :-).
And in the end, our usual Reminder: