Question 1

What AI models can run on GH200?

Accepted Answer

The GH200 (96 GB VRAM) can run models including Llama-2-7b-chat, Llama-3.2-1B-Instruct-QLORA_INT4_EO8, MediPhi-Instruct, Llama-3.2-3B-Instruct-SpinQuant_INT4_EO8, Llama-3.2-1B-Instruct-SpinQuant_INT4_EO8 and more — 10 compatible models indexed so far. Compatible models are determined by whether the model's memory footprint fits within the GPU's 96 GB VRAM, potentially using quantization (Q8, Q6, or Q4) to reduce memory usage.

Question 2

What is the cheapest cloud option for GH200?

Accepted Answer

The cheapest on-demand option for the GH200 is Lambda at $2.29/hr (gpu_1x_gh200). We track pricing across 17 providers to help you find the best deal.

Question 3

Is GH200 good for fine-tuning?

Accepted Answer

Yes, the GH200 is excellent for fine-tuning. With 96 GB of VRAM, it can handle full fine-tuning of models up to ~30B parameters and LoRA/QLoRA fine-tuning of much larger models. It's a top-tier choice for training workloads.

Question 4

How much VRAM does GH200 have?

Accepted Answer

The GH200 has 96 GB of VRAM (HBM3). It also offers 4000 GB/s memory bandwidth, 989.5 TFLOPS of FP16 compute, FP8 support for efficient inference.

Question 5

What quantization formats work on GH200?

Accepted Answer

The GH200 supports the following quantization formats: FP16 (full precision), FP8, Q8 (8-bit), Q6 (6-bit), Q4 (4-bit). FP16 works for models up to ~40B parameters. FP8 halves memory vs FP16 with minimal quality loss. Q4 quantization lets models up to ~326B parameters (by VRAM estimate) fit on the GH200.

GH200

Specifications

Cloud pricing

Compatible models

Compare this GPU

Frequently asked questions

VRAM	96 GB
Memory type	HBM3
Bandwidth	4000 GB/s
FP16 TFLOPS	989.5
FP8 support	Yes