inf-201
Intermediate
Available
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.
inf-210
Intermediate
Waitlist
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.
inf-220
Intermediate
Waitlist
How to make Kubernetes a good home for GPU work: the GPU and Network Operators, device plugins versus Dynamic Resource Allocation, GPU sharing, RDMA with Multus, batch queueing and gang scheduling with Kueue, JobSet and KAI Scheduler, distributed training and multi-node inference patterns, and the observability and cost controls that keep a shared GPU platform fair.
inf-230
Intermediate
Waitlist
Slurm from the daemons up, with an AI focus: slurm.conf and gres.conf for GPUs, cgroup v2 confinement, topology-aware placement including block topology for NVLink racks, fair share, QOS and preemption, containers with Pyxis and Enroot, multi-node launch with PMIx and NCCL, accounting and reporting, and safe upgrades.
inf-240
Intermediate
Waitlist
A structured comparison of where to rent GPUs as of October 2026: the instance families on AWS, Azure, Google Cloud and Oracle, the specialist GPU clouds, how their cluster networks and managed storage differ, how to actually obtain scarce capacity, and the managed cluster services that sit on top.
inf-260
Intermediate
Waitlist
A practical method for the build-versus-rent decision: forecasting GPU demand, modeling the full cost of owned capacity (hardware, facilities, power, network, storage, staff and refresh), modeling cloud cost honestly (reservations, egress, idle time), finding the utilization break-even, and turning the answer into a staged, hybrid capacity plan.
inf-270
Intermediate
Waitlist
How to measure, account for and reduce the energy, carbon and water footprint of AI: GPU and node power telemetry, PUE and WUE, location- and market-based carbon accounting, embodied emissions, the SCI and SCI for AI specifications, US disclosure rules such as California's SB 253 alongside the EU requirements that reach US operators with EU sites or customers, and the engineering levers (efficient models, precision, batching, power capping and carbon-aware scheduling) that actually reduce impact.
inf-280
Intermediate
Waitlist
FinOps applied to AI: the cost anatomy of GPU infrastructure and model APIs, normalizing bills with the FOCUS specification, tagging and allocating shared GPU clusters and token spend, unit economics such as cost per inference and per resolved task, commitment and capacity strategy for GPUs, and the engineering levers (caching, batching, routing, right-sizing) that reduce AI spend without losing quality.
inf-310
Expert
Available
How to design a GPU cluster for AI from requirements to acceptance: compute node selection, rail-optimized fat-tree fabrics with real port and switch counts, InfiniBand versus RoCE, parallel storage sized for checkpoints and data loading, provisioning and scheduling, burn-in, power and cooling. Ends with a full 128-node reference design and an architecture review.
inf-320
Expert
Waitlist
The protocol and operations layer of AI networking: how RDMA and collectives load the fabric, InfiniBand subnet management, routing engines, partitions and SHARP, lossless RoCEv2 with PFC, ECN and DCQCN, Spectrum-X and Ultra Ethernet transport, NCCL and RCCL tuning, multi-tenant isolation and a disciplined method for finding the cause of a slow or flapping fabric.
inf-330
Expert
Waitlist
An operator's deep dive into parallel storage for AI: the I/O patterns of data loading, checkpointing and KV-cache offload; IBM Storage Scale internals and tuning; Lustre with LNet, DNE, PFL and Data-on-MDT; WEKA NeuralMesh and VAST's disaggregated shared-everything design; GPUDirect Storage; benchmarking that cannot fool you; and day-2 operations such as tiering, quotas, snapshots and replication.
inf-340
Expert
Waitlist
Power and cooling for AI racks measured in tens to hundreds of kilowatts: the electrical chain from utility to GPU, rack power shelves and the move toward 800 VDC distribution, GPU power dynamics and capping, direct-to-chip liquid cooling with CDUs and ASHRAE W- and S-class temperatures, immersion and rear-door alternatives, retrofitting air-cooled halls, commissioning and operations.
inf-350
Expert
Waitlist
Operating GPU clusters at production standard: defining availability and goodput SLOs, building a telemetry pipeline across GPUs, fabric, storage, schedulers and facilities, automating node health checks and drains, triaging GPU, fabric and storage failures at fleet scale, running incidents and post-incident reviews, and managing node lifecycle, spares and RMA.