Standalone

Individual GPU machines provide direct hardware access for training, inference, data processing, and research workloads. Standalone systems offer maximum flexibility and performance without orchestration overhead.

Access

  • Console, dashboard, CLI and/or API available to manage resources
  • SSH key upload and management
  • Root/sudo access available on the system
  • API automation support for provisioning and management
  • Snapshot and image curation capabilities

Configuration

  • NVIDIA Container Toolkit with Docker (NVIDIA systems)NVIDIA Container Toolkit
  • ROCm Container Toolkit with --gpus=all support (AMD systems)ROCm install guide
  • Container performance optimization (import, execution, pull speeds)
  • PyTorch and AI framework readiness validation
  • Native drivers and CUDA/ROCm toolkits installed
  • Custom driver installation supported
  • Essential packages available: python, git, curl, wget, apt, vim, nano

Hardware Support

  • NVIDIA: H100/H200/B200, A100, L40S, RTX 6000 Pro
  • AMD: MI300X, MI325X, MI355X (when available)
  • Flexible billing granularity (second, minute, hour, day)
  • External storage and network egress cost transparency
  • Spot instances available
  • Multi-node interconnects supported

Monitoring and Health Checks

  • nvidia-smi, DCGM, or AMD equivalents with full metrics accessDCGM docs
  • GPU clocking, power management, and thermal monitoring
  • ncu (Nsight Compute) usable with hardware counters accessible. Needs NVreg_RestrictProfilingToAdminUsers=0 on the host (containers also need --cap-add=SYS_ADMIN), otherwise ERR_NVGPUCTRPERM. A full GPU rental should include full profiling access, bare metal, VM, or container.Fixing ERR_NVGPUCTRPERM
  • perf top / perf stat usable for CPU-side profiling without root: perf_event_paranoid <= 1 and kptr_restrict = 0.perf security docs
  • Integration with health check systems
  • Integration with monitoring tools

Networking and Storage

  • InfiniBand/RDMA configuration and testingRDMA perftest
  • NVMe direct access with multiple filesystem support
  • Storage performance benchmarking
  • Bandwidth testing and NCCL validationnccl-tests
  • Interconnect optimization

Security

  • BMC/IPMI restriction and security validation
  • Firewall management
  • Audit logging
  • Container Toolkit vulnerability assessmentNVIDIA Container Toolkit CVE-2024-0132
  • CVE compliance verification
  • Isolation options: bare metal, VM-based, container-based, or hybrid

General Expectations

  • Console, dashboard, CLI and/or API available to manage resources
  • 24x7 support availability
  • Process for security fixes and upgrades exists, proactive notifications are clear
  • Integration with comprehensive monitoring and alerting systems
  • Integration with active and passive health check systems

All expectations