AMD Thuban – Temperature Scaling

The second graph is the one where we show you the scaling in clock frequency as a function of the temperature. As said before, when dealing with hardware and extreme overclocking, there are many variables that determine whether or not the CPU is considered ok or not ok. One of the biggest problems on an Intel-based platform is that a vast majority of the processors have the so-called coldbug and coldbootbug problems.

These two problems are very defining in the evaluation of a particular processor sample, especially the delta between both. A processor with a coldbug of -160°C and coldbootbug of -50°C can be considered worse than a processor with coldbug -140°C and coldbootbug of -135°C, unless the maximum stable frequency is a lot lower.

Fortunately, the AMD-based platforms offer us not only processors without any form of coldbug, but also without any form of coldbootbug! This allows us, overclockers, to decrease the temperature to -190°C and keep the system running without the fear of having to warm up to boot again. In fact, the biggest problem of the AMD overclocker is that you have to remember that you need to pour LN2. All jokes aside, the AMD Phenom II X6 is just a perfect solution for this kind of articles because of the non-existent cold problems.



Since there are more steps in the range of the applied temperatures, the graph does show more in detail what’s going on exactly. At 1.1V, we notice that the maximum stable frequency is almost reached when the system is 0°C. A delta of 180°C results in a mere 300MHz increase. The higher the voltage is set, the more the temperature decrease is making a difference. Based on the increase in frequency from 0°C to -180°C, we can see that 1.1V is worse than 1.3V, which is again worse than 1.5V:

# 1.1V scaling: 3634MHz – 3386MHz = 248MHz
# 1.3V scaling: 4766MHz – 4013MHz = 753MHz
# 1.5V scaling: 5393MHz – 4542MHz = 851MHz

An interesting curve is the one of the 1.5V as we can clearly see where the architecture stops scaling, namely 5.4GHz. We suspect that going below -180°C with liquid helium might allow a little bit more MHz at 1.5V (similar to the 1.1V curve) to 5.5GHz, but this of course untested.

Also, as mentioned before, we did not test 1.7V with temperatures lower than -40°C. Critics may have a point when stating this makes our data incomplete (it does), but looking at the frequency scaling at 1.7V, we can see that it’s very likely that the scaling of would be similar to the 1.5V curve. Maybe even worse, because in the compiled graphs, we do not come across the negative scaling phenomenon: a state where more voltage or lower temperature results in a less stable configuration. This phenomenon is quite common in the overclocking game, but it’s possible that we don’t see it here because we are working within the ’safety’ range.

The biggest problem of the Thuban core, as far as I can see, seems to be the lack of scaling beyond 1.5V. Yes, adding 200mV does yield you a couple of MHz, but it’s not at all as impressive as going from 1.3V to 1.5V. The difference between 1.7V and 1.8V is even less significant, indicating that the scaling sweetspot is definitely below 1.7V.

Scaling from -80°C to -180°C (and -140°C)

# 1.1V scaling: 3634MHz – 3386MHz = 248MHz (145MHz)
# 1.3V scaling: 4766MHz – 4289MHz = 477MHz (226MHz)
# 1.5V scaling: 5393MHz – 5017MHz = 376MHz (376MHz)
# 1.7V scaling: 5769MHz – 5142MHz = 627MHz (501MHz)
# 1.8V scaling: 5902MHz – 5266MHz = 636MHz (503MHz)

The good part of the temperature scaling story is that the temperature still allows quite a bit of scaling. In fact, the higher the voltage, the more the temperature is determining the overclockability. The problem is that this scaling is only visible within the voltage category.