VOLTAGEGPU

NVIDIA H100 80GB

The world's most advanced GPU for AI and HPC. Hopper architecture with 4th-gen Tensor Cores.

...from / gpu / hour
...available now
<60sdeploy time
1x+multi-GPU

Technical Specifications

GPU Memory
80 GB
HBM3
Bandwidth
3,350 GB/s
Memory Bandwidth
CUDA Cores
16,896
Hopper
Tensor Cores
528
4th Gen
FP32
67 TFLOPS
Single Precision
Tensor Perf
3,958 TFLOPS (FP8)
Mixed Precision
NVLink
900 GB/s
GPU Interconnect
PCIe
PCIe 5.0
Host Interface
Transformer EngineFP8 PrecisionDPX InstructionsConfidential Computing

Ideal Use Cases

Large Language Models

  • GPT-4 class training
  • LLaMA 70B fine-tuning
  • Mixtral 8x7B inference

AI Inference at Scale

  • High-throughput serving
  • Real-time chatbots
  • Recommendation engines

Scientific Computing

  • Drug discovery
  • Climate modeling
  • Genomics analysis

Performance Comparison

This GPU
H100 80GB
VRAM80 GB HBM3
Bandwidth3,350 GB/s
FP3267 TFLOPS
Tensor3,958 TFLOPS
A100 80GB
VRAM80 GB HBM2e
Bandwidth1,555 GB/s
FP3219.5 TFLOPS
Tensor312 TFLOPS
V100 32GB
VRAM32 GB HBM2
Bandwidth900 GB/s
FP3215.7 TFLOPS
Tensor125 TFLOPS

Multi-GPU Configurations

2x
160 GB VRAM134 TFLOPS FP32
4x
320 GB VRAM268 TFLOPS FP32
8x
640 GB VRAM536 TFLOPS FP32

FAQ

H100 delivers up to 3x faster training and 6x faster inference compared to A100, thanks to 4th-gen Tensor Cores and the Transformer Engine.

Yes, H100 introduces native FP8 support via the Transformer Engine, enabling up to 2x throughput vs FP16 with minimal accuracy loss.

The Transformer Engine dynamically switches between FP8 and FP16 precision during training, optimizing both speed and accuracy for transformer-based models.

VoltageGPU bills per second with no minimum commitment. Run for 5 minutes or 5 months — pay only for what you use.

Other GPUs

Ready to Deploy H100 80GB?

$5 free credit. No credit card required. Deploy in under 60 seconds.

99.9% Uptime Per-second billing Global network