H100 SXM GPU
The H100 SXM is the flagship Hopper GPU with HBM3 memory and NVLink 4.0, designed for maximum AI training and HPC performance in data centers.
Cloud Pricing
Cheapest on Verda — 71% below avg
Prices updated daily. Last check: 4/21/2026
Performance
Strengths & Limitations
- 80GB HBM3 memory capacity supports large model training and inference
- 3,350 GB/s memory bandwidth enables efficient data movement for memory-intensive workloads
- 4th Generation Tensor Cores with FP8 precision support via Transformer Engine
- 990 TFLOPS FP16 performance for AI training and inference acceleration
- Multi-Instance GPU (MIG) technology allows workload partitioning and isolation
- NVIDIA Confidential Computing provides hardware-level security for sensitive workloads
- NVLink interconnect supports high-bandwidth multi-GPU configurations
- 700W TDP requires substantial power delivery and cooling infrastructure
- SXM form factor limits deployment to compatible server platforms
- Previous-generation architecture compared to current GB300 series
- High computational power may be excessive for smaller AI models or basic inference tasks
- Memory capacity constraints for the largest current language models exceeding 70B parameters
Key Features
About H100 SXM
Common Use Cases
The H100 SXM targets large-scale AI training workloads, particularly for language models up to 70 billion parameters where its 80GB memory capacity and high memory bandwidth prove essential. Its 990 TFLOPS FP16 performance and Transformer Engine make it well-suited for training and fine-tuning transformer-based models, while the substantial CUDA core count supports traditional HPC simulations and scientific computing. The MIG capability enables cloud providers to partition the GPU for multiple concurrent workloads, making it valuable for multi-tenant AI inference serving and development environments.
Full Specifications
Hardware
- Manufacturer
- NVIDIA
- Architecture
- Hopper
- CUDA Cores
- 14,592
- Tensor Cores
- 456
- TDP
- 700W
Memory & Performance
- VRAM
- 80GB
- Memory Bandwidth
- 3350 GB/s
- FP32
- 67 TFLOPS
- FP16
- 990 TFLOPS
- FP64
- 34 TFLOPS
- Release
- 2022
Frequently Asked Questions
How much does an H100 SXM cost per hour in the cloud?
H100 SXM pricing varies by provider, region, and commitment level. Check the pricing table above for current rates across all providers.
What is the H100 SXM best used for?
The H100 SXM excels at large language model training and inference, particularly for models up to 70 billion parameters. Its 80GB memory capacity and high bandwidth make it suitable for AI research, fine-tuning large models, and high-performance computing workloads requiring substantial memory and computational resources.
How does the H100 SXM compare to the current GB300 series?
The H100 SXM represents NVIDIA's previous-generation Hopper architecture, while GB300 series uses the newer Blackwell Ultra architecture. The H100 SXM offers 80GB HBM3 memory and 990 TFLOPS FP16 performance, whereas newer GB300 GPUs typically provide higher performance and updated architectural features, though specific comparisons depend on the exact GB300 model.