AI Reports · Epoch AI · 2026-11-09 · Survey / index / data
Changing how we measure AI chip inference performance
Epoch AI proposes the Inference Performance Index (IPI), arguing that FLOP/s measures underestimate chips' inference performance because memory and inter-chip communication also matter. It estimates that Nvidia GB300 provides about 25 times the inference capacity of H100 per chip on the agentic workloads it models, despite only 2.5 times the 8-bit FLOP/s. IPI Version 1 focuses on agentic workloads, combines results across four representative models, and expresses scores in GB300 equivalents.