Inicio › Chips › NVIDIA H100

NVIDIA H100

Chip GPU Envío

The GPU of the ChatGPT era — the unit of account for the 2023–25 AI build-out.

Estás viendo un registro de chip: qué es este procesador, sus especificaciones clave, quién diseña, fabrica y suministra sus componentes, y dónde se utiliza.

Hopper's H100 became the defining commodity of the AI boom: clusters were sized in H100s, startups raised funding denominated in them, and export policy was written around them. Its Transformer Engine pairs fourth-generation tensor cores with FP8 precision, roughly tripling A100 training throughput per chip.

xAI's Colossus assembled roughly 100,000 of them in 122 days; Meta, Microsoft, OpenAI, Google and every major cloud deployed them in the hundreds of thousands. Blackwell succeeded it in 2025, but H100/H200 fleets still carry a large share of world AI compute.

Especificaciones y datos clave

Cada cifra está etiquetada según su base: especificación del fabricante, declaración del fabricante, dato publicado, estimación u oficial. "No divulgado públicamente" significa exactamente eso — nunca inventamos valores.

Overview

Designer NVIDIA

Architecture Hopper (GH100)

Announced March 2022 (GTC); shipping volume from early 2023

Status Shipping — succeeded by Blackwell at the frontier

Manufacturing

Foundry / process TSMC 4N (custom 5 nm-class)Vendor spec

Transistors 80 billionVendor spec

Die size 814 mm²Vendor spec

Packaging CoWoS-S with 5 or 6 HBM stacksVendor spec

Memoria

Memory 80 GB HBM3 (SXM); 94 GB HBM3 (NVL); 80 GB HBM2e (PCIe)Vendor spec

Memory bandwidth 3.35 (SXM) / 3.9 (NVL) TB/sVendor spec

Cómputo

TF32 tensor (dense) 495 TFLOPSVendor spec NVIDIA datasheets often headline throughput with structured sparsity; figures here are dense unless noted.

BF16/FP16 tensor (dense) 990 TFLOPSVendor spec

FP8 tensor (dense) 1,979 TFLOPSVendor spec FP8 via the Transformer Engine was Hopper's signature feature

INT8 tensor (dense) 1,979 TOPSVendor spec

FP64 tensor 67 TFLOPSVendor spec

AI Performance

MLPerf Set the reference results for training and inference benchmarks across 2023–24 submissionsMLPerf

Energía

TDP up to 700 (SXM) / 350 (PCIe) WVendor spec

Physical

Form factors SXM5, PCIe, NVL (paired PCIe)

Interconnect NVLink 4 — 900 GB/s; PCIe 5.0Vendor spec

Precio

Launch price Not publicly disclosed

Market price ~$25,000–40,000 per GPU at the 2023–24 peakReported

Cloud rental Roughly $2–7 per GPU-hour across major and specialist clouds during 2025 — falling steadilyReported varies with commitment and provider

Cómo se conecta con el resto del stack

Los enlaces de esta ficha de un vistazo: haz clic en un grupo para ir directamente a él. +11 más grupos a continuación.

Diseñado por

NVIDIA

Fabricado por

TSMC

Nodo de proceso

TSMC N5 / 4N (5 nm class)

Empaquetado

CoWoS

Proveedor de empaquetado

TSMC

Tecnología de memoria

HBM3

Memoria suministrada por

SK hynix

Interconexión y redes

NVLink & NVSwitch InfiniBand

Presente en

Dell PowerEdge XE9680/XE9712 Supermicro NVL72 & 4U Liquid-Cooled Systems NVIDIA HGX H100/H200

Ejecuciones

Megatron-Core vLLM Triton (OpenAI) PyTorch

Respaldado por

CUDA

Usado para entrenar

Llama 3.1 405B Llama 4 Grok 3

Disponible desde

Amazon Web Services Microsoft Google (Alphabet) Oracle CoreWeave Lambda

Implementado en

xAI Colossus

Base para

NVIDIA H800

Sucedido por

NVIDIA H200

Sucesor de

NVIDIA A100

Compite con

AMD Instinct MI300X Google TPU v5p AWS Trainium2 Intel Gaudi 3 Huawei Ascend 910C Groq LPU

Relacionado

Export Controls & the Chip War Training & Pre-training MLPerf

Fuentes

NVIDIA H100 page ↗ · Hopper architecture whitepaper ↗

Registro revisado a fecha de 2026-01-15 — ver Metodología.