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.