{"categories":[],"author_url":"https://blog.hatena.ne.jp/Vengineer/","provider_url":"https://hatena.blog","blog_title":"Vengineer\u306e\u5984\u60f3","published":"2024-02-12 08:00:00","provider_name":"Hatena Blog","version":"1.0","title":"NVIDIA H100 \u3068 H200 \u306e Inference Performance\u306e\u9055\u3044\u306b\u3064\u3044\u3066","blog_url":"https://vengineer.hatenablog.com/","type":"rich","url":"https://vengineer.hatenablog.com/entry/2024/02/12/080000_1","author_name":"Vengineer","description":"\u306f\u3058\u3081\u306b NVIDIA H100\u306b\u5bfe\u3057\u3066\u3001H200\u306e LLM Inference \u6027\u80fd\u304c\u5411\u4e0a\u3059\u308b\u3068\u3001NVIDIA\u306f\u4e0b\u8a18\u306e\u3088\u3046\u306b\u8a00\u3063\u3066\u3044\u307e\u3059\u3002 www.nvidia.com LIama2 13B \u3067\u30011.4X GPT-3 175B \u3067\u30011.6X LIma2 70B \u3067\u30011.9X \u3067\u3059\u3002 H100 => H200 \u306b\u306a\u3063\u3066\u5909\u308f\u3063\u305f\u3082\u306e H100\u304b\u3089H200\u306b\u306a\u3063\u3066\u5909\u308f\u3063\u305f\u306e\u306f\u3001\u30e1\u30e2\u30ea\u3067\u3059\u3002 H100 \u3067\u306f\u3001HBM3 80GB, 3TB/s H200 \u3067\u306f\u3001HBM3e 141GB, 4.8TB/s \u30e1\u30e2\u30ea\u5e2f\u57df\u304c\u30011.6\u500d\u306b\u306a\u3063\u3066\u3044\u307e\u3059\u3002HBM\u3078\u306e\u30a2\u30af\u30bb\u30b9\u304c\u30dc\u30c8\u30eb\u30cd\u30c3\u30af\u3067\u3042\u308b\u306e\u306a\u3089\u30011.6\u500d\u7a0b\u5ea6\u307e\u3067\u306f\u2026","height":"190","html":"<iframe src=\"https://hatenablog-parts.com/embed?url=https%3A%2F%2Fvengineer.hatenablog.com%2Fentry%2F2024%2F02%2F12%2F080000_1\" title=\"NVIDIA H100 \u3068 H200 \u306e Inference Performance\u306e\u9055\u3044\u306b\u3064\u3044\u3066 - Vengineer\u306e\u5984\u60f3\" class=\"embed-card embed-blogcard\" scrolling=\"no\" frameborder=\"0\" style=\"display: block; width: 100%; height: 190px; max-width: 500px; margin: 10px 0px;\"></iframe>","image_url":"https://cdn-ak.f.st-hatena.com/images/fotolife/V/Vengineer/20240211/20240211152252.png","width":"100%"}