GPU Cloud - InstantGPU Documentation

GPU Cloud

URL: /docs/gpu-cloud/

Deploy dedicated GPU infrastructure with the flexibility and control of a cloud environment.

GPU Cloud is designed for developers and teams that need direct access to GPU resources for development, inference, training, and other compute-intensive workloads. You get full virtual machine control—SSH in, install anything, attach persistent storage, and run it just like your own hardware.

Features #

InstantGPU GPU Cloud combines bare-metal speed with cloud flexibility:

Dedicated GPU Compute

100% dedicated hardware. No shared cores or CPU throttling during heavy jobs.

Full Environment Control

Root sudo privileges. Install any CUDA version, Linux packages, or drivers.

SSH Access

Direct terminal access via SSH keys or interactive web terminal in your browser.

Persistent Storage

Keep your datasets, models, checkpoints, and code intact between pod restarts.

Docker & Templates

Start from pre-built templates (PyTorch, vLLM, ComfyUI) or custom Docker images.

Pause & Scheduling

Pause instances when not in use to save money, with automated failover migration.

Create a GPU Instance #

Launch and configure a dedicated GPU pod in 5 simple steps:

Step 1: Choose a GPU

Select the GPU that matches your workload requirements:

NVIDIA H100 / H200 SXM

80GB–141GB VRAM. Frontier LLM training, massive batch inference, and FP8 acceleration.

NVIDIA RTX 4090 / 5090

24GB–32GB VRAM. Fast prototyping, image generation, ComfyUI, and fine-tuning.

NVIDIA L40S / A100

48GB–80GB VRAM. Cost-effective mid-sized LLMs, computer vision, and production servers.

Step 2: Choose an Environment

Start with a pre-built template (PyTorch 2.4, CUDA 12.4, TensorFlow, vLLM, ComfyUI) or specify your own custom container image from Docker Hub or GitHub Packages.

Step 3: Configure Storage

Add the storage volume required for:

Models: Safetensors, Hugging Face caches, GGUF/AWQ quantized weights.
Datasets: Training corpora, image datasets, tokenized parquet files.
Applications: Python virtual environments, Git repositories, custom scripts.
Generated files: Output weights, checkpoints, logs, and generated images/audio.

Step 4: Launch

Review your configuration, choose between Secure Cloud or Community Cloud, and launch the instance. Your GPU will boot in 15–30 seconds.

Step 5: Connect

Access your instance using SSH, web terminal, JupyterLab, or VS Code Remote SSH using your generated connection string.

When to Use GPU Cloud #

GPU Cloud is ideal for:

AI Development & Research

Interactive experimentation in Jupyter notebooks, interactive debugging, and quick iterative testing.

Model Training & Fine-Tuning

Long-running LoRA, QLoRA, full parameter fine-tuning, and multi-day distributed training workloads.

Dedicated High-Volume Inference

24/7 sustained production workloads where predictable flat-rate hourly billing is cheaper than per-second calls.

Custom Environments & Proprietary Code

Non-standard CUDA libraries, legacy dependencies, custom drivers, and private networked servers.

Explore Live GPU Cloud Inventory

Browse real-time pricing and stock status for H100, RTX 4090, A100, and more.

View Live GPU Pricing Table →