NVIDIA HGX H100
Based on the NVIDIA Hopper™ architecture, the NVIDIA HGX H100 features fourth-generation Tensor Cores and a Transformer Engine with FP8 precision that provides up to 4x faster training over the prior generation for GPT-3 (175B) models.
Deploy nowMini Cluster 64 H100 GPUs
Base Cluster 24 H100 GPUs
Starting at $2.300 / hour
Powered by the breakthrough NVIDIA Hopper™ architecture and Tensor Core technology for accelerated AI model training
Connected by NVIDIA Quantum-2 3200Gb/s InfiniBand Networking with Non-Blocking InfiniBand Network Design
NVIDIA H100 SXM with FP8 Support available
Enterprise-ready at any scale and any location
Clusters at any size
Vultr's enterprise-ready infrastructure seamlessly supports any cluster size of NVIDIA HGX H100 GPUs. Whether you require a small cluster or a massive deployment, Vultr ensures reliable, high-performance computing to meet your specific needs.
Globally available, locally accessible
Large clusters of NVIDIA HGX H100 GPUs are available where you need them, thanks to Vultr’s extensive infrastructure. With 32 global cloud data center locations across six continents, we guarantee low latency and high availability, enabling your enterprise to achieve optimal performance worldwide.
Enterprise-grade compliance and security
Vultr ensures our platform, products, and services meet diverse global compliance, privacy, and security needs, covering areas such as server availability, data protection, and privacy. Our commitment to industry-wide privacy and security frameworks, including ISO and SOC 2+ standards, demonstrates our dedication to protecting our customers' data.
| Specifications | NVIDIA H100 SXM | |
|---|---|---|
| FP64 | 34 TFLOPS | |
| FP64 Tensor Core | 67 TFLOPS | |
| FP32 | 67 TFLOPS | |
| TF32 Tensor Core | 989 TFLOPS2 | |
| BFLOAT16 Tensor Core | 1,979 TFLOPS2 | |
| FP16 Tensor Core | 1,979 TFLOPS2 | |
| FP8 Tensor Core | 3,958 TFLOPS2 | |
| INT8 Core | 3,958 TFLOPS2 | |
| GPU Memory | 80 GB | |
| GPU Memory Bandwidth | 3.35TB/s | |
| Decoders | 7 NVDEC | 7 JPEG | |
| Interconnect | NVIDIA NVLink®: 900GB/s PCIe Gen5: 128GB/s |
|
| 1Preliminary specifications. May be subject to change. 2With sparsity. |
||
Additional resources
FAQ
These GPUs excel in:
- AI model training & inference
- Natural language processing (NLP)
- Large-scale data analytics
- Scientific computing
- High-performance computing
- High-frequency trading (HFT)
Yes, Vultr offers on-demand NVIDIA H100 GPUs, allowing businesses and developers to scale AI workloads without investing in expensive hardware.
You can deploy NVIDIA H100 GPUs in just a few clicks via Vultr’s cloud platform. Simply choose the GPU instance that fits your needs and start running your AI or HPC workloads immediately.
Up to 80GB HBM3 memory
These GPUs offer low-latency inference performance with FP8 Tensor Cores, making them ideal for real-time AI applications like chatbots, voice assistants, and recommendation engines.
Yes, Vultr supports multi-GPU clusters using NVIDIA H100 GPUs, enabling faster AI model training and inference across multiple nodes.
NVIDIA H100 GPUs are high-performance GPUs designed for AI, deep learning, and high-performance computing (HPC). They feature Tensor Cores optimized for machine learning workloads, offering superior speed and efficiency compared to previous generations.
Tensor Cores are specialized AI cores in NVIDIA GPUs that accelerate matrix multiplications, a key operation in deep learning. This technology enables faster AI model training and inference, significantly boosting performance for machine learning applications.
The NVIDIA H100 introduces FP8 precision and the Transformer Engine, which deliver up to 4x faster training and inference for transformer-based models like GPT and BERT. FP8 enables smaller data formats with minimal accuracy loss, while the Transformer Engine dynamically chooses the optimal precision for each layer – ideal for large-scale deep learning on Vultr’s cloud infrastructure.
H100 support Multi-Instance GPU (MIG), allowing up to seven isolated GPU instances per physical GPU. This enables better utilization, isolation, and security across AI, ML, and HPC workloads. On Vultr’s cloud GPU hosting platform, MIG makes it easy to allocate right-sized GPU resources without overprovisioning.
Yes – paired with Vultr’s globally distributed infrastructure, H100 GPUs deliver low-latency AI inference capabilities close to end-users. These GPUs are well-suited for edge inference workloads such as video analytics, recommendation engines, and speech recognition, where real-time performance and availability across global regions are critical.
What workloads are best suited for NVIDIA H100 GPUs?
These GPUs excel in:
- AI model training & inference
- Natural language processing (NLP)
- Large-scale data analytics
- Scientific computing
- High-performance computing
- High-frequency trading (HFT)
Reserve the NVIDIA HGX H100 GPU now