Denvr AI Services Docs
  • Welcome to Denvr AI Services Docs!
  • OVERVIEW
    • Getting started
      • Launch a virtual machine
      • Secure Shell (SSH): Best Practices
      • API Usage samples
      • Registration
    • Data centers
    • Shared responsibility model
    • Technical support
    • What's new!
  • PLATFORM
    • Dashboard
    • Applications
    • Virtual machines
    • Bare metal
    • Storage
    • Networking
    • User management
    • Billing
  • API Reference
    • Authentication
    • Clusters
    • Applications
    • Virtual machines
    • VPCs
    • Bare metal
  • Additional Information
    • FAQs
      • Desktop vs data center GPUs
      • Differences of bare metal and virtual machines
      • GPU monitoring
      • Using Github with SSH keys
      • Data persistence and recovery
      • Do you support Kubernetes?
      • Installing GPU drivers
      • What is the network bandwidth?
      • What ports are publicly accessible?
      • What is persistent local storage?
      • Adding DAS to /etc/fstab
      • Provisioning States
    • Policies
      • Terms of Service
      • Privacy Policy
      • Maintenance policy
Powered by GitBook
On this page
  • Bare metal
  • Virtual machines
  • Feature comparison
  • Performance
  • Security
  • Management
  1. Additional Information
  2. FAQs

Differences of bare metal and virtual machines

Bare metal

Bare metal clusters provide direct access to physical hardware for greater control and customization of the environment. This includes but is not limited to:

  • physical CPUs, not virtualized CPUs

  • no hypervisor overhead which requires host CPU and memory resources

  • ability to change symmetric multi threading (SMT) and NUMA BIOS settings

  • full access to local NMVe block devices to configure filesystems and RAID

  • enhanced security by removing the hypervisor layer

Denvr Cloud has limited observability into bare metal hosts which which requires additional responsibility of the tenant to monitor and report system issues for analysis.

Virtual machines

Virtual machines provide access to GPUs and compute resources with a minor overhead required for the hypervisor. The primary benefits of virtualization, even for full nodes, are:

  • choice of operating system images to use during provisioning

  • 1-5 minutes to boot the instance (depends primarily on number of vCPUs assigned)

  • ability to resize allocated resources (CPU, memory disks) as requirements change

  • less likelihood of over-provisioning resources versus large bare metal hosts

  • usage of snapshots for backups and starting new instances

  • self-service management via Denvr Cloud console and APIs

GPU processing and InfiniBand/RoCE fabrics are not impacted by virtualization as the devices are passed-through to the guest O/S. This provides the same level of performance, control, and isolation benefits as a bare metal host.

Feature comparison

Performance

Bare Metal
Virtual Machines

Overhead

Hypervisor requires CPU and memory resources

Consistency

Potential variability in performance

Security

Bare Metal
Virtual Machines

Tenant isolation

Multi-user access

May require additional software to manage user isolation and Quality of Service

Security risk

Higher risk due to shared infrastructure. Virtualization is a mature technology in use since the early 2000's.

Data privacy

Management

Bare Metal
Virtual Machines

Node types

Scalability

Full node only

Software and drivers

Denvr only installs the GPU and fabric drivers.

Time to launch

Slower

Backup and restore

Difficult to backup the operating system for recovery

PreviousDesktop vs data center GPUsNextGPU monitoring

Last updated 7 months ago

None

High consistency in performance

Single tenant only

Single or multi-tenant

Quality of service is enforced

Lowest risk of cross-tenant impact

Denvr has no mechanism to access the system, including for operational support

Denvr can monitor and maintain the system hardware instead of the user

Full node only

Allows customization of machine resources including single-GPU instances

Instance sizes can be changed dynamically without re-provisioning

Pre-configured images are available with required software dependencies.

Fastest

Simple to snapshot, clone, copy, and duplicate machine images.

✅
✅
✔️
✔️
✅
✅
✔️
✔️
✔️
✔️
✅
✅
✅
✅