According to SemiAnalysis, Nvidia’s GB300 submission in FP4 format shows 1.64 times lower cost efficiency than Google’s TPUv7 on agentic inference benchmarks. This performance-per-dollar gap, ranging from 1.42 to 1.67 times in Google’s favor, persists despite Nvidia’s engineering efforts to improve its results. The finding is particularly striking given that Nvidia used the leaner FP4 format designed to maximize throughput, while Google submitted its TPUv7 in FP8. In terms of configurations, the TPUv7 required only 4 chips compared to 72 GPUs for Nvidia’s comparison. For cost-conscious enterprises, the TPUv7 presents a structural advantage on agentic inference workloads, where long-context and multi-turn interactions are predominant.
Source: Read the original article
