For two years the argument about data center GPUs was whether they wear out or go obsolete too quickly. The evidence now points the other way.
On October 1, NVIDIA published its case that AI hardware stays productive far longer than its depreciation schedule assumes. The A100 shipped in 2020 and is still in commercial service six years later. NVIDIA also notes that Microsoft’s V100 fleet ran 8.4 years against a six-year book life.
So whether these GPUs can keep working is no longer in question. What is still open is narrower and more practical.
Fleet numbers describe fleets
An 8.4-year service life is what happened across a fleet that stayed in one operator’s hands, in conditions that operator controlled. It does not describe any single card.
Nobody buys or redeploys a fleet average. An operator moving older accelerators to inference, or a buyer taking them on, is dealing with specific units. Each has its own history: the workloads it ran, the thermal conditions it ran in, how it was cooled, and how it was handled when it came out of the rack.
Two units of the same model, from the same data center, can come out in different condition. Model and age alone will not tell you which is which.
What it takes to know
The condition of a retired accelerator is something you measure. In practice that means three things.
Testing under load. A unit that powers on has passed the easiest test there is. Burn-in under sustained workload is what shows whether it still performs to specification, and whether it keeps doing so.
Grading against defined criteria. A grade is only useful if it means the same thing every time. The tests behind each grade should be written down and repeatable.
A record that travels with the unit. The next owner should be able to see what was tested, when, and what the result was.
Liquid-cooled systems raise the bar again. They have to be tested as they were built to run, which requires liquid-cooling test infrastructure that air-cooled benches do not provide.
Four questions to ask before you buy or redeploy
- Was each unit tested under load, or only powered on?
- What does each grade mean, and which tests sit behind it?
- Does the test record come with the unit?
- Can liquid-cooled hardware be tested as it was designed to run?
Where illumynt stands
illumynt’s Talorem® platform was built for this work: GPU burn-in and grading, and test capability for liquid-cooled hardware, applied to AI infrastructure coming out of hyperscale and data center environments.
The industry has shown that GPUs can last. Whether a particular one will is an engineering question, and it has an answer.