GPU Marketplace

The GPU Marketplace provides access to dedicated baremetal GPU instances for maximum performance and resource isolation.

What is Baremetal?

Baremetal instances run directly on physical hardware without virtualization overhead. This provides:

• Maximum Performance - No hypervisor layer
• Dedicated Resources - Hardware exclusively for your use
• Consistent Performance - No noisy neighbor issues
• Full Hardware Access - Direct GPU memory and NVLINK

Available GPU Types

The marketplace offers various NVIDIA GPUs:

Availability varies based on inventory and demand. Use filters to find currently available GPU types.

Browsing the Marketplace

Viewing Inventory

1. Navigate to Compute > GPU Marketplace
2. Browse available instance configurations
3. Filter by GPU type
4. View pricing and specifications

Instance Details

Each listing shows:

• GPU Configuration - Type and count
• System Specs - CPU, memory, storage
• Pricing - Hourly rate (including GST)
• Availability - In stock or waitlist

Reservation Process

Step 1: Browse and Filter

1. Navigate to GPU Marketplace
2. Filter available inventory:
   • GPU Type: Select specific GPU models
   • GPU Count: Number of GPUs needed
   • Search: Find by name or configuration
3. Sort by price, GPU count, or availability

Step 2: Select an Instance

1. Review instance specifications:
   • GPU configuration
   • CPU cores and memory
   • Storage capacity
   • Network bandwidth
2. Check pricing (hourly rate including GST)
3. Click Reserve

Step 3: Configure Reservation

1. Select Project: Choose which project to bill
2. Operating System: Select your preferred OS:
   • Ubuntu 22.04 LTS (recommended)
   • Ubuntu 20.04 LTS
   • CentOS 8
   • Debian 11
   • Rocky Linux 8
3. SSH Key: Select key for access
4. Billing Period: Hourly or monthly commitment
5. Review total cost

Step 4: Complete Reservation

1. Confirm reservation details
2. Payment is processed (deducted from wallet or charged)
3. Status changes to Provisioning

Step 5: Provisioning

After payment:

1. Instance is prepared (may take 30 minutes to a few hours for baremetal)
2. OS is installed and configured
3. SSH credentials are provided when ready
4. Status changes to Active
5. Email notification sent when ready

Managing Reservations

View Reservations

Navigate to GPU Marketplace > My Reservations to see:

• All active and pending reservations
• Provisioning status
• Payment status
• Connection details

Reservation Status

Cancel a Reservation

Before provisioning completes:

1. Find the reservation
2. Click Cancel
3. Refund processed if applicable

Active Instances

Viewing Instances

Go to GPU Marketplace > My Instances for:

• All provisioned baremetal instances
• Connection information
• Instance specifications

Connecting

Baremetal instances provide SSH access:

ssh root@<instance-ip>

Connection details including IP and credentials are shown on the instance detail page.

Terminating

To end a baremetal instance:

1. Find the instance
2. Click Terminate
3. Confirm the action

Billing stops after termination.

Pricing

Baremetal pricing is shown per hour and includes:

• GPU costs
• System resources (CPU, memory, storage)
• Network bandwidth
• GST (for India)

Pricing is displayed transparently in the marketplace listing.

Waitlist

When desired GPU types are unavailable:

Joining the Waitlist

1. Select the unavailable GPU type
2. Click Join Waitlist
3. Enter your requirements
4. Submit

Notification

When inventory becomes available:

• You’ll receive an email notification
• Act quickly as inventory is first-come-first-served

Managing Waitlist

View and manage your waitlist entries:

• See your position
• Cancel waitlist entries
• Update requirements

Use Cases

Large Model Training

• Multi-GPU instances with NVLink
• H100/A100 for transformer models
• Long reservation periods

Batch Processing

• Reserve for specific time periods
• Process large datasets
• Release when complete

Production Inference

• Dedicated resources for SLA
• Consistent latency
• No resource contention

Best Practices

1. Plan Ahead - Popular GPUs sell out quickly
2. Right-Size - Choose appropriate GPU for your workload
3. Monitor Usage - Ensure GPUs are utilized efficiently
4. Use Waitlist - Get notified when preferred GPUs available

Next Steps

• Set up Storage for your training data
• Configure Billing for reservations