NVIDIA H100 Tensor Core GPU

Architecture Overview

GPU Architecture: Hopper (H100)

Process Node: TSMC 4N (Custom 4nm)

Transistor Count: 80 billion

SMs (Streaming Multiprocessors): Up to 144 (varies by configuration)

CUDA Cores: Not explicitly stated (Hopper focuses on Tensor Cores rather than CUDA cores)

Tensor Cores: 4th Generation (with new FP8, FP16, TF32, FP64 acceleration)

PCIe Version: PCIe 5.0 (x16)

NVLink: 4th Gen NVLink (900 GB/s bidirectional bandwidth)

Multi-Instance GPU (MIG): Up to 7 instances per GPU