The Generative AI Revolution - Direct and Indirect Effects on HPC for Weather and Climate Prediction

In the past year, the computing landscape has change substantially as a result of the the generative AI boom. Suddenly we are in a new world, with high prices and long lead times for GPUs, because generative AI is compute-bound and access to more compute capability is a significant factor in the success of AI companies. With well-funded startups sometimes spending upwards of 80% of their capital on compute and hyperscalers buying as many GPUs as they can, prices have increased and HPC centers planning procurements must recalibrate.

What workloads should run on these potentially expensive accelerators? How much AI will you run, and what type of AI is it? Will GPUs be the price/performance leader for your accelerated NWP and Earth System models? Benchmarks must enable organizations and vendors to assess price/performance differences between CPUs, GPUs and the various hybrid chips with different architectures and memory capacities. Because of the wide diversity of accelerators and accelerator node architectures, benchmarks that can provide meaningful information using a small number of nodes will be especially valuable. In a time of rapid change, having the right information will enable organizations to choose architectures that provide the best value for their expected workloads.