Google submits first TPUv7 FP8 vLLM results to SemiAnalysis inference benchmark

Google has submitted its first TPUv7 FP8 vLLM results to SemiAnalysis's Agentic Inference Benchmark, according to a SemiAnalysis post on X. SemiAnalysis says the TPUv7 delivers 1.67x better performance per dollar than the Nvidia GB300 NVL72 running FP4.

The only source is SemiAnalysis's X thread, which is the first of four posts. The post's claim rests on the benchmark's own results, and no other report or independent verification is included in this story.