I have read the scientific article and checked their methods and graphs and this is my observation:
1. They tested GPU, not CPU.
There are almost no OS freezes, because graphics driver is restarted and recovered in case of GPU computation errors caused by low voltage.
In case of CPU computation errors, you may notice immediatelly blue screen/system restart, because CPU is managing whole system.
2. They choice GTX 980 which has 1126MHz base and 1216MHz boost frequency. That is how are set from Nvidia. Each Nvidia's partner is setting the boost frequencies even higher. In the article they started to check errors from core frequency 1404MHz up to 1806 which is considered as pretty high overclocking on default cooling solution.
You can clearly see(graphs figure 3 and 4), that if you do ~7.44% reduction of voltage(nice decrease of power consumption) for max. boost frequency 1404MHz(set by partner) there are no errors on this scientific level of measurement. It is sad, they did not publish exact volatage values.
The higher you go with frequency the lower posibilities are to undervolt, because for these frequencies the voltage/frequency in GPU's bios is not set. That scientific undervolting test doest not have much sense for overclocked frequencies.
What I am advising to cyrilp is easy and safe way where this scientific testing methods are not needed and the OS freezing test plus small voltage reserve after finding correct voltage for CPU is 100% enough for stable system with correct calculations. Otherwise as I said, I would be noticing constantly errors and bad strange results on everything I am doing on my PC with undervolted CPUs and GPUs almost 15 years.
My advised CPU undervolting knowledge are not falsehood.
The RAM errors caused by electromagnetic radiation are common, when users put their notebooks to sleep state over night(RAM modules still under voltage) instead of power off state. After two or more days strange things happens.