Atlas

Surge

Accelerate your AI inference workloads with instant access to thousands of H100/H200 GPUs through our on-demand infrastructure. Surge provides enterprise-grade autoscaling for containerized AI models with on-demand billing and zero upfront commitments.

Why Choose Surge for AI Inference?

  • Instant GPU Scaling - Go from 10 to hundreds of H100/H200 GPUs in minutes
  • Cost-Efficient Inference - Pay only for active GPU time ($X/hr per H100) with automatic scale-to-zero
  • Seamless Integration - Works with existing ML pipelines through AWS SQS, RabbitMQ, and Redis triggers
  • Production-Ready - Built-in monitoring, logging, and automatic recovery for mission-critical workloads

Getting Started

Surge configuration is done through a YAML configuration file in a shared GitHub repository that has been set up for your organisation. The repository structure is

$.
>├── apps
>│   └── surge.yaml
>└── values.yaml

Configuration

The configuration file is available at path values.yaml. To update your container deployment, simply modify the image field in your values.yaml file with your pre-pushed image from our Fluidstack Harbor registry:

1image: image-name:v1.2.0

It is possible to specify resource specifications in your values.yaml file. This ensures optimal usage of cluster resources for your containers For example:

1resources:
2  limits:
3    nvidia.com/gpu: 1
4    memory: "16Gi"
5    cpu: "8"

This configuration requests 1 GPU, 16GiB of RAM and 8 vCPUs for each container.

Make sure to adjust these values (number of GPUs, memory, and CPU) according to your workload demands.

Deployment

  1. Push container image to the Atlas Container Registry.
  2. Update the image tag in your configuration.
  3. Commit changes and push to the main branch of the shared repository.
  4. Monitor deployment in ArgoCD.

Autoscaling

Your deployment automatically scales based on real-time demand using our pre-configured optimization rules. We monitor:

  • Queue depths in message brokers (Kafka, SQS, RabbitMQ, Redis etc)
  • API request rates
  • Resource utilization patterns

For special scaling requirements, our support team can discuss and quickly implement custom rules.

Interested in Surge, contact [email protected].