NVIDIA H200
芯片
GPU
发货中
Hopper with 141 GB of HBM3E — the memory upgrade inference was waiting for.
您正在查看一条芯片记录：该处理器是什么、其关键规格、谁设计它、谁制造它、谁供应其零部件，以及它被用在哪里。
The H200 pairs the H100 die with 141 GB of HBM3E at 4.8 TB/s — 76% more capacity and 43% more bandwidth. Because LLM inference is usually memory-bound, that alone delivered up to ~2x serving throughput on large models without any new compute. It was the first product to ship HBM3E and bridged the gap to Blackwell.
规格与关键参数
每个数据均标注来源依据：厂商规格、厂商声明、公开报道、估算或官方数据。"未公开披露"即字面含义——我们从不捏造数值。
Overview
Designer
NVIDIA
Architecture
Hopper
Announced
November 2023; shipping from mid-2024
Status
Shipping
Manufacturing
Foundry / process
TSMC 4NVendor spec
内存
Memory
141 GB HBM3EVendor spec
Memory bandwidth
4.8 TB/sVendor spec
算力
Compute
Same published tensor throughput as H100 (FP8 dense 1,979 TFLOPS)Vendor spec
电力
TDP
up to 700 WVendor spec
Physical
Form factors
SXM5, NVL
Interconnect
NVLink 4 — 900 GB/sVendor spec
它如何与整个技术栈相连
本记录的关联链接一览——点击分组可直接跳转。 +4 更多分组见下方。
设计方
由……制造
封装
内存技术
内存供应商
内部包含
部署于
后继产品
前代产品
来源
记录审阅截至 2026-01-15 — 查看 方法论.