SemiAnalysis_· @SemiAnalysis_ · X·· 9 小时前AI 评分
TPUv7 FP8 性价比对比 GB200 NVL72
另一项对比是，在中高交互性场景下，TPUv7 FP8 的每美元性能比 GB200 NVL72 FP4 最高好 3.16 倍，两者均启用 MTP，且 GB200 还在使用分离式部署，而 TPUv7 仅使用聚合式服务。话虽如此，在低交互性场景下，GB200 FP4 目前仍保持 1.2 倍的微弱优势。(2/3)
Another comparison is that at medium to high interactivity, TPUv7 FP8 is up to 3.16x better perf per dollar than even GB200 NVL72 FP4, both with MTP enabled, and GB200 is even using disagg, yet TPUv7 is just using aggregated serving. That being said, at low interactivity, GB200 FP4 still maintains a slight 1.2x advantage for now. (2/3)
来源：SemiAnalysis_ · x.com