What is NVIDIA H100?
The NVIDIA H100 Tensor Core GPU is NVIDIA's flagship data center GPU, built on Hopper architecture. Launched in 2022, delivers unprecedented performance for AI, HPC, and LLMs.
80GB HBM3 memory with 3.35 TB/s bandwidth. Transformer Engine provides up to 6x faster LLM training vs A100.
Innovations: 4th Gen Tensor Cores, FP8 (1,979 TFLOPS), MIG (up to 7 instances), NVLink 900 GB/s.
Specifications
ArchitectureHopper
CUDA Cores16,896
Tensor Cores528 (4th Gen)
Memory80GB HBM3
Bandwidth3.35 TB/s
FP16989 TFLOPS
TDP700W
Best Use Cases NVIDIA H100
- ✓LLM Training - Models with hundreds of billions of parameters
- ✓LLM Inference - Production AI with ultra-low latency
- ✓Generative AI - GPT-4, Stable Diffusion, content creation
- ✓HPC - Scientific simulations, weather modeling
- ✓Recommendation Systems - Personalization at scale
- ✓NLP - Translation, chatbots, sentiment analysis
- ✓Computer Vision - Object detection, autonomous vehicles
- ✓Drug Discovery - Molecular modeling
NVIDIA H100 vs GPU
| Comparison | Performance | Preço | Ideal Para |
|---|
💡 Provider Tips
Lambda Labs, CoreWeave: $3.50/hr. RunPod: 31 regions. Vast.ai: spot pricing (watch for interruptions).
FAQs
What is H100 best for?
LLM training, generative AI, HPC. Transformer Engine offers 3-6x speedup vs A100.
How much does H100 cloud cost?
Starts at $3.50/hour on Lambda Labs and CoreWeave. Spot instances can be 30-50% cheaper.
H100 worth it over A100?
For large LLMs, yes (3-6x better). For smaller models, A100 has better price/performance.
How much VRAM?
80GB HBM3 with 3.35 TB/s - ideal for LLMs up to 175B parameters.
H100 for Stable Diffusion?
Works, but overkill. RTX 4090 ($0.35/hr) is better value.