NVIDIA

NVIDIA H100 SXM

80GB HBM3 Hopper accelerator for large-scale AI training, inference and HPC workloads

Reference GPU-hour unit rate for full-server rental

$3.30/GPU-hour

Reference rate only. Actual configuration, billing method and total price are set in the customer contract.

NVIDIA H100 SXM combines the Hopper architecture, 80GB HBM3 memory and fourth-generation Tensor Cores for large-scale model training, high-throughput inference and mixed-precision scientific computing.

Illustrative whole-server configuration. Component combinations vary by GPU edition, server platform and available deployment capacity; manufacturer GPU figures are not a performance commitment for a TATC deployment.

Memory

80GB HBM3

3.35 TB/s bandwidth

Architecture

Hopper

Family · NVIDIA Hopper

Tensor compute

3,958 TFLOPS

FP8 with sparsity · FP16/BF16: 1,979 TFLOPS

Power

Up to 700W (configurable)

Manufacturer specification

Highlights

80GB HBM3 and 3.35 TB/s memory bandwidth per GPU
Fourth-generation Tensor Cores and Transformer Engine with FP8 precision
Second-generation MIG support for up to seven 10GB GPU instances

Software stack

CUDA 12.x

PyTorch 2.x

Triton

vLLM

TensorRT-LLM

NCCL

Best-fit scenarios

Large-scale LLM training and fine-tuning
High-throughput inference (TensorRT-LLM / vLLM)
Mixed-precision scientific computing

Hardware gallery

NVIDIA H100 SXM and reference performance visuals

Charts show manufacturer comparisons and projections. Actual results depend on workload, software and system configuration and are not a TATC performance guarantee.

Explore NVIDIA H100 SXM Bare Metal

Share your workload profile to compare GPU memory, server capacity, networking and software-stack options.

Compare GPU models

NVIDIA

NVIDIA H100 SXM

80GB HBM3 Hopper accelerator for large-scale AI training, inference and HPC workloads

Reference GPU-hour unit rate for full-server rental

$3.30/GPU-hour

Reference rate only. Actual configuration, billing method and total price are set in the customer contract.

Compare GPU models

Memory

80GB HBM3

3.35 TB/s bandwidth

Architecture

Hopper

Family · NVIDIA Hopper

Tensor compute

3,958 TFLOPS

FP8 with sparsity · FP16/BF16: 1,979 TFLOPS

Power

Up to 700W (configurable)

Manufacturer specification

Highlights

80GB HBM3 and 3.35 TB/s memory bandwidth per GPU
Fourth-generation Tensor Cores and Transformer Engine with FP8 precision
Second-generation MIG support for up to seven 10GB GPU instances

Software stack

CUDA 12.x

PyTorch 2.x

Triton

vLLM

TensorRT-LLM

NCCL

Best-fit scenarios

Large-scale LLM training and fine-tuning
High-throughput inference (TensorRT-LLM / vLLM)
Mixed-precision scientific computing

Hardware gallery

NVIDIA H100 SXM and reference performance visuals

Charts show manufacturer comparisons and projections. Actual results depend on workload, software and system configuration and are not a TATC performance guarantee.

Explore NVIDIA H100 SXM Bare Metal

Share your workload profile to compare GPU memory, server capacity, networking and software-stack options.

Compare GPU models