H100 Specifications
Fourth-generation Tensor Cores speed up all precisions, including FP64, TF32, FP32, FP16, INT8, and FP8, to reduce memory usage and increase performance while still maintaining accuracy for LLMs.
80 GB
HBM3 Memory
Run 70B+ parameter models on a single GPU.
900 GB/s
NVLink Bandwidth
Fourth-gen NVLink for multi-GPU scaling.
1,979
TFLOPS FP16
3,958 TFLOPS at FP8 with Transformer Engine.
9X
Faster Pre-Training
vs A100 on large language models.
NVIDIA H100 on Denvr AI Cloud
Distributed Training
Scale across 8-GPU NVLink nodes and multi-node clusters with 3,200G InfiniBand. Bare metal access for maximum training throughput with no virtualization overhead.
LLM Inference
Serve 70B+ parameter models on a single GPU, or scale to 8 GPUs with 640 GB VRAM for 500B+ parameter models at FP8 precision.
Production AI
H100 delivers 9x faster training vs. A100 at $2.10/hr on-demand. For extended context or larger models, consider the H200 with 141 GB HBM3e.
Managed Storage
High-performance Weka filesystem and local NVMe available. No external storage to provision for datasets, checkpoints, or model artifacts.
Platform
GPUs
GPU VRAM
vCPUs
Memory
Local Storage
Interconnect
On-Demand
NVIDIA H100 SXM
8
80 GB
208
1024 GB
6x 3.8TB NVMe
IB 3200G
$2.10 / GPU
Configurations
Per-minute billing with on-demand and reserved options. All configurations available as bare metal, VM, or model endpoints.
Related GPUs
Compare Denvr GPU options by workload and performance requirements.
Optimized For
Distributed training, multi-node scaling
Large model training, high-throughput inference
Extended context, large batch inference
VRAM
80 GB
80 GB
141 GB
Memory Bandwidth
2,039 GB/s
3,350 GB/s
4,800 GB/s
FP64/FP32
19.5 TFLOPS
67 TFLOPS
67 TFLOPS
FP16
312 TFLOPS
1,979 TFLOPS
1,979 TFLOPS
FP8
-
3,958 TFLOPS
3,958 TFLOPS
NVLink
600 GB/s
900 GB/s
900 GB/s
On-Demand Pricing
$1.30 / GPU
$2.10 / GPU
Reserved only
Infrastructure you can trust at scale
As an NVIDIA Cloud Partner we build and operate AI clusters following NVIDIA Reference Architectures. Your models and data are supported via strict privacy safeguards and SOC 2 Type 2 security practices.
