Home / Hardware / NVIDIA / Hopper / H100

NVIDIA Hopper H100

Arch: HopperProduction ReadyTDP 700W
Compute Performance (FP8 Tensor)▲ 3.3x vs A100
2.0 PetaFLOPS
FP8 Tensor performance with Transformer Engine
H100 Hopper2.0 PF
A100 Ampere0.6 PF
V100 Volta0.1 PF
Memory System
80GB HBM3
5 Hi-Stacks / 5120-bit interface
3.35 TB/s Bandwidth
Interconnect & I/O
900 GB/s NVLink 4
Bi-directional total bandwidth (18 links)
PCIe Gen 5.0 x16
Real-World Applications
Foundation Model Training

The industry workhorse for training large language models. H100 clusters power the majority of frontier AI labs, with the Transformer Engine automatically managing mixed-precision training for optimal throughput on models from 7B to 175B+ parameters.

Production AI Inference

Deploy production inference endpoints with native FP8 quantization. The H100's Transformer Engine delivers up to 30x inference speedup over A100 for large language models while maintaining model accuracy.

Drug Discovery & Genomics

Accelerate molecular dynamics simulations, protein structure prediction, and genomic analysis. The H100's 3.35 TB/s memory bandwidth handles the massive datasets required for computational biology research.

Computer Vision & Video AI

Train and deploy vision transformers, video understanding models, and autonomous driving perception stacks. The H100 is the standard platform for production computer vision workloads across cloud providers.

Full Technical Specifications
GPU ArchitectureNVIDIA Hopper
Transistor Count80 Billion (4N Process)
CUDA Cores16,896
Tensor Cores4th Gen (528 cores)
Memory Capacity80 GB HBM3
Memory Interface5120-bit
Memory Bandwidth3.35 TB/s
L2 Cache50 MB
NVLink Bandwidth900 GB/s
Form FactorSXM5 / PCIe
Thermal Design Power700W (SXM) / 350W (PCIe)