Skip to content

Multi-GPU Workload Registration

Connect multiple GPUs to a single workload to handle large-scale AI tasks.

Tip

Multi-GPU workloads are used for large model training, inference, and rendering tasks that are difficult to handle with a single GPU. Parallel processing performance scales with the number of GPUs.

Single GPU vs Multi-GPU Comparison

Item Single GPU Multi-GPU
Suitable for Small-scale inference, testing Large-scale training, large models
VRAM Based on 1 GPU Number of GPUs × individual VRAM
Cost Low Proportional to number of GPUs
Setup complexity Low Moderate

Before You Begin

  • Make sure you have enough points charged. (Number of GPUs × hourly rate)
  • Prepare a container image that supports multi-GPU.
  • Decide on the number and model of GPUs you need in advance.

How to Register

  1. Complete the container setup on the new workload registration screen. (The screen below is an example.)

    Screenshot: Container setup screen

  2. Select your desired GPU model in the GPU selection section.

    Screenshot: Replica count screen

  3. Choose how many GPUs to allocate for the selected model.

    Screenshot: GPU selection screen

    Warning

    Make sure the selected GPU model is available in the required quantity on the same node. If the quantity is insufficient, deployment may be delayed.

    Heterogeneous GPU Selection

    You can select multiple GPUs of the same model or mix different GPU models.

  4. Confirm the estimated total cost. Costs increase proportionally with the number of GPUs.

    Screenshot: Estimated total cost screen

  5. Select whether to deploy immediately and click the Register button.

Multi-GPU Usage Examples

Model Recommended GPU Config Use Case
LLaMA3 70B A100 × 2 (80GB VRAM) Large LLM inference
Stable Diffusion XL RTX 3090 × 2 High-resolution batch image generation
Training (Fine-tuning) RTX 4090 × 4 Model fine-tuning