Running GPUs at scale?

We’re working with a handful of operators on clusters that tune themselves and waste less. Tell us what you run and where it hurts.