Skip to main content

Bare Metal Servers

Substrate AI provides high-performance bare metal servers dedicated entirely to your workloads, offering raw power without virtualization overhead.

Overview​

Bare metal instances are ideal for:

  • High-performance computing (HPC)
  • Large-scale AI model training
  • Latency-sensitive applications
  • Custom hypervisor deployments

Step 1: Provisioning​

  1. Navigate to Enterprise > Bare Metal
  2. Click Deploy New Server
  3. Select configuration:
FeatureOptions
GPUH100, A100, L40S configurations
CPUAMD EPYC / Intel Xeon Scalable
RAMUp to 2TB DDR5
Network2x 100GbE / InfiniBand
  1. Launch instance

Step 2: Configuration​

Network Setup​

  • Servers come with 2x public IPv4 addresses
  • Private networking available for cluster setups
  • Configure VLANs via Enterprise > Networking

Operating System & Cloud-Init​

Default images available:

  • Ubuntu 22.04 / 24.04 LTS
  • Rocky Linux 9
  • Custom ISO support (via support request)

User Data Example​

You can provide a cloud-init script during provisioning to pre-configure your instance:

#cloud-config
package_update: true
packages:
  - nvidia-driver-535
  - build-essential
runcmd:
  - [ sh, -c, "echo 'Performance mode enabled' > /etc/motd" ]

Step 3: Access & Management​

SSH Configuration​

Connect via SSH using your verified keys. For easier access, configure your local ~/.ssh/config:

Host baremetal-01
    HostName <server-ip>
    User root
    IdentityFile ~/.ssh/id_ed25519
    StrictHostKeyChecking no

Then simply run:

ssh baremetal-01

Verifying Connectivity​

Once connected, verify your environment:

  1. Check GPU Status:

    nvidia-smi

  2. Test Network Connectivity:

    # Test public internet access
    ping -c 4 8.8.8.8

    # Test private network (if configured)
    nc -zv <private-ip-neighbor> 22

IPMI Access​

For out-of-band management, access the IPMI console via the dashboard to troubleshoot boot issues or configure BIOS settings.

Performance Tuning​

Recommended settings for AI workloads:

  • Enable Performance Mode in BIOS
  • Configure HugePages for memory-intensive applications
  • properly configure NVIDIA Fabric Manager for multi-GPU interconnects
# Example: Check NVIDIA Fabric Manager status
systemctl status nvidia-fabricmanager

Pricing​

Bare metal servers are billed hourly or monthly with significant discounts for reserved instances (1-year / 3-year commitments).