Efficiency and cost impacts of cluster “slicing” for training

Characterize the efficiency and cost impacts of training AI models using large numbers of less powerful chips (cluster slicing) versus fewer more powerful chips with the same theoretical throughput, including implications for decentralized or disaggregated training configurations.

Background

Governance measures targeting compute should understand how cluster composition affects training feasibility and cost. If many smaller chips can substitute for fewer powerful chips without major penalties, restrictions on specific chip classes may be less effective.

Quantifying slicing trade-offs informs more precise hardware policies and helps anticipate circumvention via cluster design.