Skip to content

NCP-AII AI Infrastructure Professional Study Strategy

Study Approach

Phase 1: Foundation (1-2 weeks)

  1. GPU Hardware
  2. Memorize H100 and A100 specifications
  3. Understand NVLink generations and bandwidth
  4. Learn NVSwitch topology within DGX nodes
  5. Study MIG profiles and use cases

  6. 📖 DGX Documentation

  7. Networking

  8. InfiniBand generations (HDR, NDR, XDR) and speeds
  9. Fat-tree and rail-optimized topologies
  10. GPUDirect RDMA, Storage, and Peer-to-Peer

  11. 📖 NVIDIA Networking

  12. Storage

  13. Parallel file systems and their trade-offs
  14. Storage tiering strategies
  15. Data loading optimization patterns

Phase 2: Software Stack (2-3 weeks)

  1. Kubernetes GPU Management
  2. GPU Operator components and installation
  3. GPU scheduling, time-slicing, and MIG in Kubernetes
  4. DCGM Exporter and monitoring

  5. 📖 GPU Operator Docs

  6. Job Scheduling

  7. Slurm GRES configuration and job submission
  8. Multi-tenant policies: partitions, fair-share, QOS
  9. Base Command Manager capabilities

  10. 📖 Base Command Docs

  11. Hands-On Practice

  12. Deploy GPU Operator on a Kubernetes cluster
  13. Submit Slurm jobs with GPU requests
  14. Monitor GPUs with DCGM and nvidia-smi

Phase 3: Performance and Exam Prep (1-2 weeks)

  1. Performance Tuning
  2. DCGM metrics and profiling with Nsight Systems
  3. Multi-GPU training strategies (data, tensor, pipeline parallel)
  4. NCCL tuning and network optimization
  5. Memory optimization (mixed precision, ZeRO, checkpointing)

  6. 📖 NCCL User Guide

  7. Review and Practice

  8. Work through scenario-based questions
  9. Focus on bottleneck identification and resolution
  10. Create flashcards for key specifications

Official NVIDIA Resources

Supplementary Resources

Exam Tactics

Question Strategy

  1. Read the full question - identify the infrastructure component being tested
  2. Look for NVIDIA-specific solutions - prefer NVIDIA tools and platforms
  3. Keywords to watch for:
  4. "Bandwidth" or "throughput" - think NVLink, InfiniBand, GPUDirect
  5. "Isolation" - think MIG
  6. "Multi-tenant" - think Slurm partitions, QOS, fair-share
  7. "Low GPU utilization" - identify the bottleneck (network, storage, CPU)
  8. "Scale" or "multi-node" - think InfiniBand, NCCL, parallelism strategy
  9. "Monitor" or "health" - think DCGM
  10. Eliminate wrong answers - verify numbers match known specifications
  11. When in doubt - choose the solution that uses NVIDIA-specific technology

Time Management

  • 120 minutes for 60-70 questions
  • Approximately 1.7-2 minutes per question
  • Flag specification-heavy questions for review
  • Reserve 15 minutes for reviewing flagged questions

Common Pitfalls

Hardware Confusion: - NVLink is GPU-to-GPU, InfiniBand is node-to-node - MIG requires A100 or newer - not supported on older GPUs - H100 has 4th-gen NVLink (900 GB/s), A100 has 3rd-gen (600 GB/s) - DGX H100 has 8 GPUs with NVSwitch, not PCIe connected

Software Confusion: - GPU Operator manages the Kubernetes GPU stack, not training frameworks - DCGM is for monitoring, Nsight is for profiling - Slurm GRES is for GPUs, partitions are for node groups - Base Command Manager is NVIDIA's enterprise tool, not open-source Slurm

Performance Mistakes: - Data parallelism is preferred when model fits on one GPU - Tensor parallelism needs NVLink bandwidth - do not use across nodes - Pipeline parallelism has bubble overhead - not always faster - NCCL tree algorithm is better for large messages, ring for small

Progress Tracking

Self-Assessment Questions

  • Can I list DGX H100 specifications including NVLink bandwidth?
  • Do I know InfiniBand generation speeds and ConnectX adapter capabilities?
  • Can I configure GPU Operator and schedule GPU workloads in Kubernetes?
  • Do I understand Slurm GRES configuration and multi-tenant policies?
  • Can I diagnose GPU performance bottlenecks from DCGM metrics?
  • Do I know when to use data vs tensor vs pipeline parallelism?