by Tim Tscheblockov
09/07/2004 | 10:42 PM
As you know, NVIDIA Corp. has recently rolled out a new series of mainstream graphics processors to replace the obsolete chips of the GeForce FX 5700 line. For our detailed report on the capabilities and features of the new GeForce 6600/6600 GT GPUs follow this link. Our second report on the subject was about the performance of the new products in games and benchmarks. The review you’re now reading is going to provide yet another point of view on the new GPU series – that of the overclocker. That is, I’m now interested in how “power-hungry” and “hot” the GeForce 6600 GT is, how much power it consumes. I will also examine the overclockability of the sample we have in our test lab and will try to increase it with a volt modification, i.e. by increasing the voltage of the graphics core.
The GeForce 6600 GT inherits the architecture of NVIDIA’s GeForce 6800 series chips but differentiates from them in having half the vertex and pixel pipelines and being manufactured with a thinner 0.11-micron tech process.
The memory bus of the GeForce 6600/6600 GT has also been dissected – 128 bits are left instead of the 256-bit bus of the GeForce 6800 series. The top GeForce 6600 GT model comes with high-speed GDDR3 chips, while the low-end GeForce 6600 is equipped with ordinary TSOP-packaged DDR SDRAM.
We received a reference sample of the GeForce 6600 GT for our tests.
You should already know from our reviews that new GeForce 6600 GT graphics cards are superior in speed as well as in functionality to products on GPUs of the previous generation, the GeForce FX 5700 series. It is a remarkable fact that the new cards appear simpler in design and smaller in size than the older ones: GDDR3 chips have termination circuitry inside themselves, while the use of advanced components in the voltage regulators has lead to simplification of the voltage circuitry.
The back side of the PCB is covered with a shield. You can see details of the GPU and memory voltage regulators and a scattering of small surface-mountable components, mostly ceramic capacitors, there. The biggest chip on the back side of the PCB is the Philips SAA7115HL that’s responsible for digitizing the signal from the card’s video input.
The SLI connector is located at the upper edge of the card; it allows linking together two graphics cards. This feature is of small importance now as there are no mainboards with two PCI Express x16 slots, but the potential opportunity of building a SLI configuration around a mainstream graphics card is nice as it is.
But let’s get back to the card. It is based on the NVIDIA GeForce 6600 GT graphics processing unit that operates at a frequency of 500MHz.
The snapshot shows there are no AGP-to-PCI Express bridges on board or on the GPU’s wafer. PCI Express support is an inborn possession of the GeForce 6600 GT.
The reference card carries 128MB of 2.0ns GDDR3 memory from Samsung:
The nominal frequency of the graphics memory is 500 (1000DDR) MHz. The memory chips have no cooling system proper – they are supposed to be cooled by the air from the GPU cooler:
The contact spot of the heatsink looks like the rest of its surface: there are no copper insertions or polishing:
The cooler as if says, “This graphics processor is a cool guy”. In any case, the cooling system doesn’t look as mighty as the reference one on NVIDIA GeForce 5900 Ultra/5950 Ultra cards. Mark that the GeForce 6600 GT is obviously better than these two monsters in performance. Well, you should already know everything about the performance characteristics of the GeForce 6600 GT (in synthetic as well as real-life tests), and it’s time you knew everything about its power consumption and heat dissipation.
I’ll try to assist you in finding this information.
Power consumption issues were seen to by the developers of the PCI Express bus. It surpasses the Accelerated Graphics Port (AGP) in this respect. As a response to the higher power consumption of modern system components, the PCI Express specification lists advanced energy-saving functions, and a graphics card in the PCI Express x16 slot can receive much more power than one installed into the AGP. The developers of the new standard accounted for the unavoidable growth of the power consumption of graphics cards and realized that the maximum of 42 watts as provided by the AGP wouldn’t be enough.
So, the maximum amount of power a graphics card may get from the PCI Express x16 slot is about 75W, although originally a smaller value (like 60W) was proposed. The developers must have miscalculated the rate of the power consummation growth of modern graphics cards and had to adjust the specification, assigning more contacts for transferring electricity and raising the consumption bar to 75W.
Graphics cards in the “desktop” PCI Express variant are powered by two lines, 3.3v and 12v. The specification of the PCI Express x16 slot sets the maximum currents on these lines at 3amp and 5.5amp, respectively. The total power consumption on these lines, i.e. the maximum consumption of a graphics card without additional power connectors, is easily calculated to be 75.9W.
NVIDIA’s GeForce 6600 GT reference card has no additional power plugs – there’s place left for it on the PCB, but it is not soldered in:
So, the power consumption of the GeForce 6600 GT cannot exceed 75.9W. But how can we measure it with more precision? It’s done simply, in the same way as I tested the AGP cards in my previous reports. I isolate the power contacts of the PCI Express x16 slot and send this power directly to the card, bypassing the slot.
The PCI Express x16 slot is divided into two parts: the bigger part contains signal lines, and the smaller part consists of power lines. On the front side, the smaller part of the slot contains three 12v contacts and one 3.3v contact.
The back side of the slot has two 12v contacts and two 3.3v contacts.
In total, we have five contacts of the 12v line which has a higher maximum current value (5.5amp) and only three contacts of the 3.3v line (a maximum current of 3amp).
I measured the power consumption of the GeForce 6600 GT in two modes: Idle and Burn. There were no running applications in the Idle mode; the Windows Desktop was on the screen in 1280x1024x32bit@75Hz display mode. For the Burn mode I started up Doom 3 in 1600x1200 resolution with the maximum graphics quality settings and forced 4x full-screen antialiasing and 8x anisotropic filtering.
I tested the card both at its regular frequencies and at 615/1100MHz clock rates (these were the maximum stable frequencies of the card at ordinary overclocking, i.e. without modding the cooling system and interfering with the core and memory power circuitry).
The results follow:
So, the NVIDIA GeForce 6600 GT consumes rather small amount of power and, accordingly, its heat dissipation is low, too: 18.5W in the Idle mode and 48W in the Burn mode. This is no wonder since the GeForce 6800 GT is the first graphics processor from NVIDIA to be made with the thinner 0.11-micron tech process, and it is made up of fewer transistors in comparison to the GeForce 6800/GT/Ultra.
Overclocking affects the power consumption of the card but slightly: it grows to about 19W in the Idle mode and to 51W in the Burn mode.
Note that the relatively low power consumption in the Idle mode is achieved without such tricks as frequency reduction or GPU voltage reduction: the GeForce 6600 GT supports independent 2D/3D clock rates, but has identical clock rates in these modes (500/1000MHz) like the top-end GeForce 6800/GT/Ultra chips. The voltage on the graphics processor in 2D and 3D modes is practically the same – 1.431 and 1.477v, according to my measurements.
I will examine the GPU voltage circuit shortly. Right now, here’s a diagram for you that shows the power consumption characteristics of other graphics cards:
So, the GeForce 6600 GT consumes less power than the NVIDIA GeForce FX 5900 Ultra/5950 Ultra, GeForce 6800 Ultra/6800 GT, ATI RADEON X800 XT PE, RADEON 9800 XT, but more power than the NVIDIA GeForce FX 5700 Ultra or the GeForce 6800. You already know the standing of the GeForce 6600 GT among these products, so you can make up an opinion on your own.
Let’s take an in-death view of the power consumption of the GeForce 6600 GT:
The card takes only 0.15W through the 3.3v line of the PCI Express slot and this value doesn’t depend on the clock rates or the load. Thus, almost all the power the card consumes comes from the 12v line. The consumed current on this line is 4.07amp and rises to 4.35amp at overclocking.
So, the GeForce 6600 GT consumes just a little less power from the 12v line than the GeForce 6800 Ultra, and the recommendations about the PSU for the 6800 Ultra – with a good current reserve on the 12v line – seem to be true for the 6600 Ultra, too.
Replacing the standard cooling system with something more efficient, for example with a water-cooling solution, and pulling up the voltages are the traditional means of increasing the overclockability of a graphics card.
The graphics card under question has GDDR3 memory chips from Samsung, known for their low overclockability which doesn’t grow much even after a voltage increase. So I didn’t tamper with the memory voltage on the reference card at all – it wasn’t worth the trouble.
The overclockability of the graphics processor is another matter, as we deal with NVIDIA’s first GPU made with the 0.11-micron tech process. The thinner tech process and the smaller transistor count in comparison to the GeForce 6800 series should render this chip capable of working at higher clock rates.
Let’s check it out. The core voltage regulator is based on the ISL6534 chip from Intersil. It is a dual-channel pulse-width controller capable of driving a line regulator. One of the channels supplies power to the GPU. Unlike the regulators on GeForce FX 5950 Ultra or FX 5900 Ultra cards, which have digital inputs for setting the output voltage level and capable of adjusting the output voltage “on the fly”, this chip uses a resistor devisor for setting the output voltage.
Curiously enough, the NVIDIA engineers made this regulator change the output voltage “on the fly”, too. Receiving the control signals from the GPU’s registers, two transistors attach resistors with preset resistances to the devisor, thus adjusting the voltage value. Voltage regulators on NVIDIA GeForce 5900 XT cards employ the same idea, by the way.
In order to lift the voltage of the graphics processor, you only must reduce the resistance of one of the divisor’s resistors. That’s exactly what I did:
You can see the controller chip of the voltage regulator in the top left corner of the snapshot, while the output voltage control scheme and the two resistors of the devisor are in the center. The arrows point to the spots I soldered an additional variable resistor to.
By changing the resistance of the variable resistor and evaluating the overclockability of the card at different GPU voltages, I found the maximum GPU frequency at which the card was stable – 640MHz. Thus, the frequencies of the reference graphics card were 640/1100MHz at extreme overclocking.
The voltage of the graphics processor was 1.233v at Windows XP startup, 1.562v after the load of the graphics card driver, in 2D mode, and 1.629v in 3D applications. Before the modification, these voltages were 1.233v, 1.431v and 1.477v, respectively.
So, 10% GPU voltage boost in 3D mode allowed me to increase the GPU clock rate by 28% above the nominal. That’s good.
Of course, I couldn’t achieve this result without replacing the standard cooling system. As usual, I used the water-cooling solution Aquarius II from Thermaltake for my extreme overclocking experiments.
I tested the GeForce 6600 GT graphics card on the following testbed:
I tested the graphics card in 1600x1200 resolution and at the maximum graphics quality settings. The driver’s settings of the anisotropic filtering optimization were left intact, i.e. enabled. I also performed the tests in 1600x1200 with forced 4x full-screen antialiasing and 8x anisotropic filtering.
So, here are the numbers for Doom 3 (the standard demo2 record):
You see that the overclocking of the GPU in the simple mode (without full-screen antialiasing and anisotropic filtering) is more rewarding than memory overclocking. This is natural since I could only raise the memory frequency y 10%, while the GPU clock rate grew by 28%.
Forcing 4x FSAA and 8x AF, I increase the load on the memory subsystem, so memory overclocking brings in a higher profit than in the first test mode. However, GPU overclocking is more profitable even here.
I made my own record on the Training level of Far Cry:
It’s the same in Far Cry as in Doom 3: GPU overclocking is more rewarding than memory overclocking in each test mode. This might be different if I had achieved the same frequency growth with the memory as with the GPU, but 10% was the maximum I could get from the memory chips on that graphics card.
Diagrams below represent performance of NVIDIA GeForce 6600 GT graphics card with overclocking of core, memory as well as core in addition to memory.
So, although the GeForce 6600 GT seems to be “misbalanced” in the characteristics (i.e. its performance is oftener limited by the memory bandwidth rather than by the GPU), I enjoyed more gains from GPU overclocking than from the memory speedup. The reason is simple: I managed to increase the GPU clock rate by 28%, while the memory clock rate grew by 10% only.
This section of my report is about the efficiency of the standard cooling system of the GeForce 6600 GT. I kept track of the GPU temperature with the help of RivaTuner and of the memory temperature with a Fluke 54-II thermometer.
I wrote the numbers down after half an hour of tests in two modes: Idle and Burn. There were no running applications in the Idle mode; the Windows Desktop was on the screen in 1280x1024@75Hz display mode. For the Burn mode I launched Doom 3 in 1600x1200 resolution and forced 4x full-screen antialiasing and 8x anisotropic filtering.
I post the results of the card at its regular frequencies and with the standard cooling system as well as during extreme overclocking.
The GPU and memory temperatures of the GeForce 6600 GT are perfectly normal in the Idle mode. They grew up in the Burn mode – up to 70-72°C. These are allowable, although very high values. Such temperatures shouldn’t bother you much, but you’d better choose a card with a good cooling system for both the GPU and memory if you’re into overclocking. There’s no doubt such cards will appear in the market – GeForce 6600 GT is a mass product after all!
The water cooling system works very well – the GPU temperature doesn’t exceed 45°C even after the volt modification. The temperature of the graphics memory grew up considerably, though. Its frequency is increased, while it is now left even without the scanty cooling the standard cooler provided.
The sector of mainstream graphics cards has accepted a worthy successor to the GeForce FX 5700/5700 Ultra series. As we already know, NVIDIA’s new GeForce 6600 GT GPU surpasses last-generation mainstream graphics cards, but stands on just a slightly higher level of power consumption. It takes almost all of its power from the 12v power rail, and consumes just slightly less power from this line than the NVIDIA GeForce 6800 Ultra does. That’s why my recommendations about the PSU for the GeForce 6800 Ultra (see my report on the power consumption of NVIDIA’s GPUs in contrast to ATI’s chips) are true for the GeForce 6600 Ultra – the power supply unit must have a good current reserve on the 12v line.
The overclockability of the new graphics processor, manufactured with the most advanced 0.11-micron tech process, does not impress me, but does not disappoint either. 23% frequency growth at ordinary overclocking and 28% frequency growth after volt-modding are nice results for a newly-baked reference card. I have reasons to think that overclockability of GPUs on off-the-shelf cards will be at least no worse – the manufacture will be constantly improved, and the cards themselves will be coming out with good cooling systems.
As for the memory, its low overclockability was no surprise for me – practice shows that GDDR3 chips from Samsung do not like to be clocked higher than specified.
The temperatures of the GPU and memory of the reference card were normal in the idle mode and somewhat disturbing in the burn mode. But you should make allowances for the fact that I tested a reference card from NVIDIA with a surprisingly feeble cooler. Off-the-shelf products will surely have more efficient cooling systems.
So, this is what I’m going to be busy with now – waiting for those off-the-shelf graphics cards!