Fail-slow at scale: Evidence of performance faults in large systems (2018) (dl.acm.org)

As systems become more complex, so too do their failure modes. This paper documents over a half a dozen slow performance modes of SSDs with several different internal failure modes caused by the silicon, firmware, etc.

Contrast that with the modes they discovered for CPUs: no internal factors, only external factors such as power delivery or cooling capacity.

That's quite a testament to the construction of the most sophisticated manufactured goods on the planet.

I've seen this first hand, across datacenter providers. The most common one is some sort of power supply failure that causes the CPUs to under-volt and throttle; the machine won't go down entirely, the clock will just go through the floor. It's common enough we set up alerting to watch for it specifically.