Interesting bits: part 1

Before we move on to the actual overclocking results, there are a couple of things I need to share with you about this card. This will help you understand the end results and some other, very interesting results.

1. Voltage scaling

This card does not scale very well with voltage. At least, not beyond 1.5V; between 1.2V and 1.35V there's significant scaling, but also an increase in temperature. However, even on stock cooling there's a significant scaling as displayed in notes I made on the forums.

Without voltage modifications, the card is capable of running 675/1450/940 MHz (core/shader/memory frequency). If the memory is clocked lower, to 900MHz for instance, the shader clock can be pushed to 1500MHz.

The voltage modifications are confirmed and, as we speak, I am running 1.2V Vgpu and 1.8V Vmem. Even on stock cooler, this makes quite a difference as I'm now capable of running 750/1600/955 MHz (core/shader/memory frequency). Although the voltage has been increased, the card still running below 70°C load.

On extreme cooling (read: liquid nitrogen), we tested up to 1.75V, which quite a lot for 40nm, but simply no gain past 1.5V. To be safe, however, we ran 1.6V Vgpu. As for the memory voltage, it seems that a small increase to 1.8V to 1.85V is enough to max out the memory overclockability. Again, we tried higher voltages (up to 2.05V) but this didn't result in a higher clock frequency, only in a more unstable setup.

2. Temperature scaling

Next to the voltage, decreasing the temperature is one option to increase the maximum stable frequencies. After realizing that the stock Leadtek cooler wasn't enough, we hooked up a Zalman VF900 cooler and re-tried overclocking. How well it performed you can read in underneath quote:

The Zalman cooler does wonders for this card. At the same voltages as before (1.2V Vgpu and 1.8V Vmem), I increase the core frequency from 750MHz to 800MHz (shaders are linked), the memory frequency from 955MHz to 1GHz artifact-free and, in addition, the temperature under load decreased from 80°C to 45°C!

Under liquid nitrogen, the temperature scaling was visible as well, maybe even more visible than on air cooling as shown in the next table:

1000MHz at -40°C

1100MHz at -60°C

1200MHz at -80°C

1300MHz (almost) at -100°C

(stable frequencies at 1.6V Vgpu, running through 3DMark2001SE Nature subtest).

3. Temperature pain-in-the-ass

Up until this paragraph I'm sure you all believe this article is one fairytale story in which princess "performance" is saved by prince "voltage and temperature". Well ... back to reality, I'm afraid, as this card has a big flaw: it's a low-profile card. Being low-profile means that the PCB is smaller and, more importantly, that everything is much closer to each other; for instance: the memory chips to the core. This created a problem when using liquid nitrogen to cool down the core. Why?

As you know, cold spreads through the PCB. As the memory chips are soldered onto the PCB (duh!), the cold also spreads to the memory chips. The problem is, however, that some memory chips don't really like cold,which is the case with these particular chips. Looking at the datasheet of the memory, it becomes apparent that these chips are rated to a guaranteed operating temperature between 0°C and 85°C. Below 0°C, the memory is not guaranteed to work, which is exactly what we experienced.



To keep a (very long) story short, when decreasing the core temperature below -50°C, the performance drops from (what should be) 260FPS to a mere 40FPS. Dropping the temperature below -100°C makes the memory artifact in 2D desktop environment which, for me, is a clear indication that the memory is not capable of withstanding this kind of temperature. Later into the session, we removed the insulation from the memory chips and tried to keep the memory temperature above 0°C. This worked perfectly for the memory on the right side of the card, but not for the two memory chips right above the core. As the LN2 container was mounted so closely to the PCB, it was impossible to heat the memory up.



As you can see in above picture, the back of the card has turned into ice, which indicates that cold has been traveling across and through the PCB. The upper section of the ice is where the two “cold-bugging” memory chips are located.