Skip to content

GPU Accelerator Configuration Guide

GPU accelerator configuration enables you to specify dedicated GPU resources for your SaaS Products. This feature allows your services to leverage hardware acceleration for AI/ML workloads, high-performance computing, graphics processing, and other GPU-intensive tasks.

The acceleratorConfiguration feature is part of the configurationOverrides section in your service compute configuration. It provides a declarative way to specify the type and count of GPU accelerators that should be attached to your compute instances.

Warning

GPU accelerator configuration is currently only supported on Google Cloud Platform (GCP). Shared CPU instances do not support GPU accelerators, and built-in GPU instances cannot have additional accelerator configuration.

Configuration Properties

Each accelerator configuration can be configured with the following properties:

  • Type: The GPU accelerator type to attach to the instance (required)
  • Count: Number of GPU accelerators to attach (required, minimum: 1)

Supported Accelerator Types

The following GPU accelerator types are currently supported:

  • nvidia-tesla-t4 - NVIDIA Tesla T4 (16GB GDDR6)
  • nvidia-tesla-v100 - NVIDIA Tesla V100 (16GB/32GB HBM2)
  • nvidia-tesla-p100 - NVIDIA Tesla P100 (16GB HBM2)
  • nvidia-tesla-p4 - NVIDIA Tesla P4 (8GB GDDR5)
  • nvidia-a100 - NVIDIA A100 (40GB/80GB HBM2e)
  • nvidia-a40 - NVIDIA A40 (48GB GDDR6)

Compose Format (Docker Compose Style)

x-omnistrate-compute:
  instanceTypes:
    - name: n1-standard-4
      cloudProvider: gcp
      configurationOverrides:
        acceleratorConfiguration:   
          type: "nvidia-tesla-v100"
          count: 1

Plan Format (Omnistrate Spec)

services:
  - name: gpu-service
    compute:
      instanceTypes:
        - name: n1-standard-4
          cloudProvider: gcp
          configurationOverrides:
            acceleratorConfiguration:   
              type: "nvidia-tesla-v100"
              count: 1

Cloud Provider Support

GPU accelerator configuration is currently supported only on Google Cloud Platform (GCP).

Google Cloud Platform (GCP)

Fully Supported - GCP provides comprehensive GPU accelerator support

  • General-purpose instances: Compatible with most instance families (n1, n2, e2, etc.) that can have GPU accelerators attached via acceleratorConfiguration
  • GPU-enabled instances: Instance types with built-in GPUs (a2-highgpu, a2-ultragpu, a3-highgpu series) already include GPUs and don't require acceleratorConfiguration
  • Automatic resource allocation: pods with accelerator configuration automatically receive nvidia.com/gpu resources
  • Multi-GPU support: respects the configured GPU count

Validation and Constraints

Instance Type Compatibility

  • GCP: Most general-purpose instance families support GPU accelerators
  • Shared CPU instances: Do not support GPU accelerators
  • Built-in GPU instances: Instances with built-in GPUs cannot have additional accelerator configuration

Examples of Built-in GPU Instance Types (GCP)

The following instance types already have built-in GPUs and cannot use acceleratorConfiguration:

  • a2-highgpu-1g - 1x NVIDIA A100 40GB (12 vCPUs, 85GB RAM)
  • a2-highgpu-2g - 2x NVIDIA A100 40GB (24 vCPUs, 170GB RAM)
  • a2-highgpu-4g - 4x NVIDIA A100 40GB (48 vCPUs, 340GB RAM)
  • a2-highgpu-8g - 8x NVIDIA A100 40GB (96 vCPUs, 680GB RAM)
  • a2-ultragpu-1g - 1x NVIDIA A100 80GB (12 vCPUs, 170GB RAM)
  • a2-ultragpu-2g - 2x NVIDIA A100 80GB (24 vCPUs, 340GB RAM)
  • a2-ultragpu-4g - 4x NVIDIA A100 80GB (48 vCPUs, 680GB RAM)
  • a2-ultragpu-8g - 8x NVIDIA A100 80GB (96 vCPUs, 1360GB RAM)
  • a3-highgpu-1g - 1x NVIDIA H100 80GB (12 vCPUs, 208GB RAM)
  • a3-highgpu-2g - 2x NVIDIA H100 80GB (24 vCPUs, 416GB RAM)
  • a3-highgpu-4g - 4x NVIDIA H100 80GB (48 vCPUs, 832GB RAM)
  • a3-highgpu-8g - 8x NVIDIA H100 80GB (96 vCPUs, 1664GB RAM)

Examples of Compatible Instance Types (GCP)

The following instance types can use acceleratorConfiguration to attach GPUs:

  • n1-standard-4 - 4 vCPUs, 15GB RAM (general purpose)
  • n2-standard-8 - 8 vCPUs, 32GB RAM (general purpose)
  • e2-standard-16 - 16 vCPUs, 64GB RAM (cost-optimized)
  • c2-standard-30 - 30 vCPUs, 120GB RAM (compute-optimized)
  • m1-ultramem-40 - 40 vCPUs, 961GB RAM (memory-optimized)

Count Validation

  • Minimum count: 1
  • Maximum count: Varies by instance type
  • Must be an integer value

Type Validation

  • Must be one of the supported accelerator types
  • Must be compatible with the specified instance type

Common Use Cases

AI/ML Workloads

# For TensorFlow/PyTorch training using accelerator configuration
services:
  - name: ml-training
    compute:
      instanceTypes:
        - name: n1-standard-4
          cloudProvider: gcp
          configurationOverrides:
            acceleratorConfiguration:
              type: "nvidia-tesla-v100"
              count: 1
# Alternative: Using built-in GPU instance (no accelerator configuration needed)
services:
  - name: ml-training-builtin
    compute:
      instanceTypes:
        - name: a2-highgpu-1g  # Already has 1x NVIDIA A100 40GB
          cloudProvider: gcp
          # No acceleratorConfiguration needed - GPU is built-in

High-Performance Computing

# For CUDA applications requiring multiple GPUs
services:
  - name: hpc-workload
    compute:
      instanceTypes:
        - name: n1-standard-8
          cloudProvider: gcp
          configurationOverrides:
            acceleratorConfiguration:
              type: "nvidia-a100"
              count: 2

Graphics Processing

# For rendering workloads
services:
  - name: graphics-rendering
    compute:
      instanceTypes:
        - name: n2-standard-4
          cloudProvider: gcp
          configurationOverrides:
            acceleratorConfiguration:
              type: "nvidia-tesla-t4"
              count: 1

Development and Testing

# Cost-effective GPU for development
services:
  - name: gpu-development
    compute:
      instanceTypes:
        - name: e2-standard-4
          cloudProvider: gcp
          configurationOverrides:
            acceleratorConfiguration:
              type: "nvidia-tesla-t4"
              count: 1