SemiAnalysis@SemiAnalysis_ · XAI 摘要
在 GTC 2024 上，Jensen 表示 GB200 NVL72 比 Hopper 快 35 倍。没人相信，认为这是经典的 fake Jensen Math。当我们测试其性能时，它不仅快了 35 倍，即使面对采用了 MTP、Disagg prefill、wideEP 等所有推理优化组合的强大 Hopper 基线，也快了 50 倍以上。在 InferenceX.com 查看详细结果。
At GTC 2024， Jensen said that GB200 NVL72 was 35x faster than Hopper. Nobody believed it and thought it was classic fake Jensen Math. When we tested the performance of it， it wasn't just 35x faster， it was over 50x times faster even against an strong Hopper baseline with all of the inference optimization composed together like MTP， Disagg prefill， wideEP， etc. View the nuanced results at InferenceX dot com.