Updated minutes ago
nvidia

GH200

nvidia · hopper · 96 GB HBM3 · 900W TDP

VRAM

96 GB

BF16 TFLOPS

990

Bandwidth

4000 GB/s

From

$2.99/hr

Calculate ROI with this GPU →

Spec Sheet

VRAM96 GB HBM3
Memory Bandwidth4000 GB/s
BF16 TFLOPS990
FP16 TFLOPS990
FP8 TFLOPS1980
INT8 TOPS1980
TDP900W
InterconnectNVLINK
NVLink Bandwidth900 GB/s
Max per Node8
PCIe Gen5
CUDA Compute Capability9
Tensor CoresYes

Pricing by Provider

ProviderOn-DemandReservedSpotBadge
coreweave$3.99/hr$2.99/hr-Cheapest
lambda$3.49/hr--

Compatible Models (278)

Single GPU (228 models)

Qwen 2.5 72B72.7B FP8Qwen 2.5 Math 72B72.7B FP8Qwen 2.5 VL 72B72.7B FP8Dolphin 2.9 72B72B FP8Molmo 72B72B FP8NVLM-D 72B72B FP8DeepSeek R1 Distill 70B70.6B FP8Llama 3 70B 1M Context70.6B FP8Llama 3 70B70.6B FP8Llama 3.1 70B70.6B FP8Llama 3.3 70B70.6B FP8Hermes 3 70B70.6B FP8HelpSteer2 Llama 3.1 70B70.6B FP8Llama 3.1 Nemotron 70B Instruct70.6B FP8Llama 3.1 Nemotron 70B Reward70.6B FP8Nemotron 70B70.6B FP8Llama 3.1 70B Turbo70.6B FP8Code Llama 70B70B FP8Llama 2 70B70B FP8WizardMath 70B70B FP8+208 more

Training Capabilities

Estimated GPU count for full fine-tuning (AdamW, BF16) and QLoRA

Model SizeFull Fine-TuneQLoRA
7B model2 GPUs1 GPU
13B model3 GPUs1 GPU
70B model14 GPUs1 GPU
Train on this GPU →

Energy Efficiency

Estimated tokens/second per Watt for popular models

Mistral 7B
0.61 t/s/WFP8
Qwen 2.5 7B
0.58 t/s/WFP8
Llama 3.1 8B
0.55 t/s/WFP8
Llama 3.1 70B
0.06 t/s/WFP8
Qwen 2.5 72B
0.06 t/s/WFP8

Similar GPUs

GPUVRAMBF16 TFLOPSBW (GB/s)From
H2096 GB1484000$0.99/hr
H100 NVL94 GB8353938$3.09/hr
H100 SXM80 GB9903350$1.89/hr
H100 PCIe80 GB7562000$1.89/hr
H200 SXM141 GB9904800$2.69/hr

Methodology Note

Performance estimates for the GH200are based on InferenceBench's roofline performance model with CUDA kernel-level optimization including FlashAttention v2 and PagedAttention. Memory calculations account for model weights (96 GB HBM3 available), KV-cache allocation, and activation memory. Throughput predictions use the GH200's rated 4000 GB/s memory bandwidth and 990 BF16 TFLOPS compute capacity as roofline ceilings, with empirical correction factors per GPU architecture (hopper). See our full methodology.

Frequently Asked Questions

How many AI models can run on GH200?

The GH200 can run 278 AI models from our database within a single node. Compatible models range across various parameter sizes depending on the quantization precision (BF16, FP8, INT4). Smaller models fit on a single GPU while larger models may require multi-GPU setups up to 8x GH200.

What is the GH200 inference throughput?

The GH200 delivers 990 BF16 TFLOPS and 1980 FP8 TFLOPS with 4000 GB/s memory bandwidth. Actual inference throughput (tokens/sec) depends on the model size, precision, and batch size. Use our calculator for model-specific throughput estimates.

How much does GH200 cost per hour?

The GH200 is available starting from $2.99/hour via coreweave. Prices vary by provider and pricing tier (on-demand, reserved, spot). Compare pricing across all providers in the table above.