Wirewalk Instituteof AI
inf-310 Expert AvailableFlagship

HPC Cluster Design for AI: Compute, Fabric and Storage

For infrastructure architects and senior HPC engineers who must turn an AI workload brief into a buildable, testable cluster design and defend it to a review board.

Level
Expert
Study time
20 hours
Lessons
18 in 6 modules
Exam
25 questions

About this course

Most AI cluster problems are designed in long before the first job runs: a fabric that collapses onto a few inter-switch links, a storage system that benchmarked well because the test fit in cache, a rack row that cannot shed the heat of the nodes it was bought for. This course teaches the design discipline that prevents them. You start from the workload — training, inference or mixed research HPC — and derive concrete targets for GPU count, memory, checkpoint bandwidth, data-loading bandwidth and failure tolerance.

From there you design each layer with numbers: compute nodes (HGX-class 8-GPU servers, rack-scale NVL72 systems and PCIe GPU servers), the three networks every cluster needs, rail-optimized fat-trees with leaf, spine, cable and transceiver counts, InfiniBand versus RoCE and Spectrum-X, and tiered storage across node-local NVMe, parallel file systems (IBM Storage Scale, Lustre, WEKA, VAST) and object. You learn to run benchmarks that cannot lie to you, provision with NVIDIA Base Command Manager, schedule with topology-aware Slurm alongside Kubernetes, accept the system with HPL, NCCL tests, IOR and mdtest, and plan power, cooling and day-2 operations. The final module walks through a complete hypothetical 128-node design, its variations, and how to defend it. Product facts are stated as of October 2026.

What you will be able to do

  • Translate an AI workload brief into quantitative design targets for GPU count, memory, checkpoint bandwidth, data-loading bandwidth and availability
  • Select a compute node platform and justify its GPU, CPU, memory, NIC and PCIe topology choices for a stated workload
  • Design a rail-optimized fat-tree fabric and calculate leaf, spine, cable and transceiver counts and oversubscription for a given node count
  • Choose between InfiniBand and RoCE/Spectrum-X for a cluster and explain the role of adaptive routing, SHARP and the subnet manager
  • Size a tiered storage design for checkpoint and data-loading throughput and run storage benchmarks that defeat cache effects and under-driven concurrency
  • Configure provisioning and topology-aware scheduling with Base Command Manager, Slurm and Kubernetes GPU operators
  • Write an acceptance plan using HPL, NCCL tests, IOR and mdtest, and size power, cooling and rack layout for the design
  • Produce and defend a complete reference design with a bill of materials, risk register and acceptance criteria

Syllabus

6 modules · 18 lessons · about 20 hours

Module 1: From Requirements to Design

Classify the workload, convert the brief into numbers, and choose the compute node that serves them.
3 lessons
  1. Training, Inference and Research HPC: Three Different Clusters
  2. Turning a Workload Brief into Design Targets
  3. Compute Node Selection: GPU Platform, CPU, Memory, NICs and PCIe

Module 2: Fabric Design

The three networks every cluster needs, fat-tree arithmetic with real switch counts, and the InfiniBand versus Ethernet decision.
3 lessons
  1. Three Networks: Compute Fabric, Storage Fabric, In-Band and Out-of-Band Management
  2. Rail-Optimized Fat-Trees and Oversubscription, with Port and Switch Counts
  3. InfiniBand vs RoCE and Spectrum-X: Adaptive Routing, SHARP and UFM

Module 3: Storage for AI Clusters

Tiering, sizing for checkpoints and data loading, and benchmarking storage without fooling yourself.
3 lessons
  1. Storage Tiers: NVMe Scratch, Parallel File Systems and Object
  2. Sizing Throughput for Checkpoints and Data Loading
  3. Storage Benchmarks That Lie: Cache Effects and Under-Driven Concurrency

Module 4: Software Stack and Provisioning

Image-based provisioning, topology-aware Slurm and running Kubernetes alongside it.
3 lessons
  1. Provisioning with Base Command Manager: Images, Node Categories and Rebuilds
  2. Slurm for AI Clusters: Topology-Aware Scheduling, GRES and Health Checks
  3. Kubernetes and Slurm Side by Side: GPU Operator, Network Operator and Hybrid Patterns

Module 5: Acceptance and Facilities

Proving the system works, fitting it into power and cooling, and making it operable on day two.
3 lessons
  1. Burn-In and Acceptance: HPL, NCCL Tests, IOR and mdtest
  2. Power, Cooling and Rack Density
  3. Day-2 Readiness: Monitoring, Spares and Failure Domains

Module 6: Reference Design and Review

A complete worked design, how it changes for other workloads, and how to defend it.
3 lessons
  1. Reference Design Walkthrough: A 128-Node HGX Training Cluster
  2. Variations: Inference Farm and Mixed Research HPC on the Same Budget
  3. Defending the Design to an Architecture Review Board

Exam and certificate

The exam unlocks once you have completed every lesson. It draws 25 questions from a larger bank, spread across the modules, with the answer options shuffled for each attempt. You have 60 minutes and need 80% to pass. After a failed attempt you can try again after 24 hours, up to 3 attempts in total. Feedback is given per module, never as the answer key.

Passing earns a certificate of completion and assessment from the Wirewalk Institute of AI with a unique ID that anyone can check at institute.wirewalk.com/verify. It is not an accredited qualification or a vendor certification. How certification works.

The Wirewalk Institute of AI is independent. This course teaches publicly documented products and is not affiliated with or endorsed by NVIDIA, IBM, DDN, WEKA, VAST Data, SchedMD, Kubernetes, Dell, HPE, or Supermicro. Product and company names are trademarks of their respective owners.

More in AI Infrastructure and HPC

All 14 courses in this track
inf-101 Basic Waitlist

AI Infrastructure Explained

A plain-language tour of the stack that runs AI: what a GPU server is, why clusters need special networks and storage, how jobs are scheduled, what an AI data center needs in power and cooling, and how cloud and on-premises options differ. No prior infrastructure background required.

6 h · 15 lessons Join the waitlist
inf-201 Intermediate Available

GPU Infrastructure for AI

The operator's view of data-center GPUs: what the current NVIDIA and AMD systems are, how they are packaged and connected, how the driver, CUDA/ROCm and container stack fit together, how to monitor and triage failing GPUs, how to share and size GPU memory, and what power, cooling and sourcing decisions follow.

12 h · 16 lessons See price at checkout
inf-210 Intermediate Waitlist

Linux Administration for AI Clusters

The Linux skills that differ on an AI cluster: choosing and imaging the OS, packaging GPU and RDMA drivers that survive kernel updates, NUMA and PCIe topology tuning, cgroup v2 isolation, shared identity and file systems, scientific software environments with Lmod, Spack and containers, and patching a fleet without breaking jobs.

14 h · 19 lessons Join the waitlist