Normally when we buy a processor we do so with a view to lasting X time. In some cases this time is quite long, but when we consider that we must change, we may sell this CPU and it ends up in the hands of a chain of users who will continue to ask for the same performance. This could be over, or at least cut back with the processors to come, why will there be faster CPU aging ?
We are not going to discover anything if we say that the new lithographic processes are getting more and more complicated, that the entire industry is spending huge amounts of money and that the investment will be greater because the challenges to be faced so require it. But, although many of the problems are solved, there is one in particular that does not allow many engineers to sleep: the premature aging of the chips from certain nanometers.
CPU aging, why will it occur and to what extent?
There are several factors that will determine the premature aging of new CPUs and possibly new GPUs. The first is very simple: the designs of the chips are being brought very close to the physical limit of frequencies, cores and in general, of excessive designs that seek to gain performance thereby differentiating themselves from the competition.
The best current example we have in the i9-10900K, a CPU that barely has an overclock margin, where the tension wall for common temperatures comes very fast and where the design of its 14 nm + has reached the physical limit by Intel.
This supposes a second factor that must be taken into account: the so-called electrical stress. Raising the frequencies implies that the voltage has to do the same and with it, the chip suffers a greater electrical stress that forces Intel in this case to leave a minimum margin of safety with said processor. This generates two more adjacent effects, not only to this specific CPU, but to any other that is in the same situation and is known as self-heating and field strength.
The high temperature achieves a very curious effect on the electrons, since it constantly migrates them from the hardest metal to the softest (electromigration) and this exponentially changes the life of the chip, especially at certain temperature delta.
Negative bias instability
There is another curious fact aside from everything said and that is going to get even bigger as we get closer to the nanometer itself and it is the negative bias instability factor. This factor determines what voltage is applied to each type of node according to values such as density, frequency or field strength.
This means that, even if nanometers are reduced, the minimum applicable voltage is not reduced to the same extent. For example, at 28 nm the minimum voltage fluctuated on average by 1.2 volts or 1.3 volts, but for 3 nm it is estimated that we will be at 0.6 or 0.7 volts in idle. The density in this example has been multiplied by almost 8, but the voltage has hardly been reduced by 50%, which causes overheating problems between the interconnections of the devices.
Due to all this, the manufacturers of the wafers have had to create a very curious system of monitoring and testing of the chips, since each user gives a different use to each CPU. From the one that does not overclock and has a good cooling system, to the one that takes it to the limit with a chiller and manual voltage for 24/7. The system has been called burn-in testing and that name is quite correct.
Burn-in testing
The objective of this test in wafers is very simple, since what is sought is to age a series of wafers previously selected and registered, so that through this artificial process their functionality and reliability can be verified. In addition to this, manufacturers can add control methods using certain sensors, which have predefined times when the electrons must go around a ship. If the electron takes longer or does not pass, the sensor detects this problem and reports the chip degradation.
The problem logically is that these sensors are extremely expensive and are only used in foundries under Burn-in testing, but it would be interesting to have something like that on all processors.
Ultimately, both Samsung, Intel and TSMC have found that their devices, if subjected to higher electric fields, degrade much faster than before as transistors shrink. This is going to cause that the costs of checking the wafers are higher because they have to guarantee a minimum of useful life and will surely end up affecting the frequencies and voltages, so the architectures will have a lot to say in this regard.