The reference board at 850 MHz?The GeForce GTX 460 has very high overclocking capacity, but does the reference board allow you to fully take advantage of it without damaging the board? This is something we’ve been wondering about since the problems we came across on the Point of View TGT The Beast and EVGA FTW cards.
They both use the reference PCB and cooler and are specially selected by the manufacturers who validate them at the very high clocks of 850 MHz and 855 MHz. In comparison to the stock clocks, this represents a factory overclocking of 26%! This is pretty unusual for cards based on the reference design…
We have no doubts about the GF104’s being able to stabilize at this clock, nor about the stock cooler’s capacity to handle the temperature levels. Our tests have however highlighted a problem around the power stage, which hasn’t been designed to stand up to such clocks. Compared to the stock card clocked at 675 MHz, energy consumption is similar at idle but hits great heights in load. For example, on the EVGA FTW model, from 145W in 3DMark06 (PS), we were up to 202W, and from 151W in Furmark, we’re at 217W.
Athough NVIDIA did provide for a significant overclocking margin, this does have its limits and here we’re talking energy consumption increases of 39% and 44% respectively. Obviously, the power stage’s three phases struggle at this sort of stress. To remedy things a bit, EVGA has put a small radiator under the sensitive components:
With the EVGA card, the 3 stock PCB phases for the GPU power supply have a small radiator. The 4th phase, top right, supplies the memory.
We tested the heat and noise levels on these cards in a closed casing, with a surrounding temperature of 26°C, using the 3DMark06 Pixel Shader for 45 minutes straight to give a continuously high graphics load. When we measured the temperature, the CPU was also in load, with Prime95, which has the effect of bringing performance down a bit (5%) in 3Dmark06 and therefore also reducing graphics load. We were, then, a long way off an extreme test in which the card would be saturated by Furmark or OCCT.
When run through this test, the Point of View TGT The Beast failed after 35 minutes. One of the power stage capacitors detached itself from the PCB… probably trying to run off. We thought this was probably a one-off and asked a second card of the same model. We’re still waiting for it. A little later, EVGA supplied us with a similar FTW model. It kept crashing after 5 minutes of the test. EVGA then supplied us with another FTW card, but the verdict was the same. EVGA did tell us that they had also observed this issue but said that it hadn’t come up in the games and benches recommended by NVIDIA, which is considered to be sufficient to validate the card at these clocks.
To confirm our suspicions about the power stage at these extreme clocks, we launched the same test with the casing open and used the infrared camera to follow temperature developments. The temperature at the back of the graphics card, where the power stage is, rapidly rose over 130 °C. After ten minutes, it was in excess of 150 °C! Things didn’t stop there however and just before the card failed we recorded a temperature of 160 °C. Remember that this is measured at the back of the PCB and that the components at the front of it could be even hotter. We managed to get a picture of the temperature at 158 °C:
We observed similar behaviour with a reference board that we overclocked ourselves at 850 MHz. Some models which display a lower GPU voltage and/or leakages, will have a slightly reduced energy consumption and may be able to survive this test, but most won’t.
Of course, during gaming, the power stage isn’t run through its paces so extremely. Energy consumption is lower in general and less constant, which reduces stress on the components. That said, some of the power stage components have been designed to function at maximum temperatures of 125 °C (150 °C for the MOSFETs), which can be exceeded after playing a demanding game for some time, which does put a question mark over how long these models will last, even if you’re not using the sort of heavy test tools we’ve used here.
We therefore strongly advise you against these cards. More especially, we advise you not to put your trust in models based on the stock PCB and clocked beyond 800 MHz, if they haven’t at least been provided with a dedicated radiator for the power stage. If this has been provided, going up to 825 MHz is probably alright and gives you a small safety margin.
To go beyond that, you’ll need a better adapted design. This is something that we neglected to comment on at first with respect to models such as the MSI GTX 460 Hawk, which, with six phases for the GPU, won’t have any problem at 850 MHz. We’re thinking about how we can improve our test protocol in the future, for example by putting all the cards through temperature tests when overclocked.