Do data center CPUs have shorter lifespans?

Will data center GPUs have their lifespans reduced by constantly running at high utilization rates? That is an issue raised by an anonymous poster on X, but at this point it’s just a theory.

The posting copied a screenshot of an unsourced message from a “GenAI principal architect at Alphabet Inc.” who said assuming a utilization rate of 60% to 70% – which CoreWeave and Lambda Labs reportedly achieve – would lead to a lifespan of three years instead of the anticipated five years. With lower utilization, operators could get five years of productive life out of a GPU.

It’s not an unreasonable claim. GPUs run extremely hot. The Hopper generation draws over 700 W, while the upcoming Blackwell will draw up to 1000 W of power. They run so hot that they can’t be air cooled efficiently and require water cooling.

And anecdotally, it’s not unprecedented. I’ve heard stories of gamers who bought second-hand, high-end GPU cards that had been used for crypto mining. Those cards ran 24/7 for months if not years and later died on the gamers who bought them, not realizing they’d purchased heavily used cards.

Still, the process of powering on and powering off a PC is actually worse than leaving it running steadily, notes Jon Peddie, who is president of Jon Peddie Research, which specializes in all things graphics (full disclosure: I do some work for JPR), and has a degree in electrical engineering.

“The thing that will damage a chip is turning it on and off; that causes temp cycle which impacts the connection,” he told me. “The only reason I can think of as to why a [data center add-in board] should fail is the contributed heat from the adjoining AIBs. I heat my little lab with one RTX 4090 – not a joke.”

Google must be feeling some heat from this (no pun intended) because it issued quite a vociferous denial.

“Recent purported comments about Nvidia GPU hardware utilization and service life expressed by an ‘unnamed source’ were inaccurate, do not represent how we utilize Nvidia’s technology, and do not represent our experience. Nvidia GPUs are a critical part of our infrastructure, both for internal systems and our Cloud offerings, and our experience with Nvidia GPUs is consistent with industry standards,” a Google spokesperson said in an email.

And while there are stories from gamers who got burned off by an eBay purchase of a used mining card, claims of an early death of a data center GPU by actual customers have yet to manifest. So file this under speculation, not fact.

Source:: Network World