Product Overview
The NVIDIA H100 is a flagship data center GPU based on the Hopper architecture GH100 chip, released in 2022. As of 2025, it remains the de facto standard for large language model training and inference. It introduces the Transformer Engine (dynamic FP8 precision acceleration) and the DPX instruction set (dynamic programming algorithm acceleration).
Core Specifications
| Parameter | Value |
|---|
| Architecture | Hopper GH100 |
| Process Node | TSMC 4N (custom 5nm) |
| Transistor Count | 80 billion |
| Memory | 80 GB HBM3 |
| Memory Bandwidth | 3.35 TB/s (3,352 GB/s) |
| CUDA Cores | 16,896 |
| Tensor Cores | 528 (4th Gen) |
| FP32 | 67 TFLOPS |
| FP64 | 34 TFLOPS (important for HPC) |
| TF32 Tensor Core | 989 TFLOPS (sparse) |
| FP16/BF16 Tensor Core | 1,979 TFLOPS (sparse) |
| FP8 Tensor Core | 3,958 TFLOPS (sparse) |
| INT8 Tensor Core | 3,958 TOPS (sparse) |
| TDP | 700 W (SXM5) |
| Interconnect | NVLink 4.0 (900 GB/s), PCIe 5.0 |
| MIG | Up to 7 instances |
| Form Factor | SXM5 / PCIe 5.0 |
Software & Drivers
Key Features
- Transformer Engine: Automatically switches between FP8 and FP16 to accelerate Transformer training
- 4th Gen Tensor Cores: Support FP8 (E4M3, E5M2)
- DPX Instructions: Hardware-accelerated dynamic programming algorithms
- MIG: Virtualizes a single GPU into up to 7 independent GPU instances
- NVLink 4.0 + NVSwitch 3.0: 900 GB/s intra-server interconnect
Use Cases
- LLM training and fine-tuning
- Recommendation systems and multimodal AI
- HPC scientific computing
- Large-scale distributed training clusters