Learning paths

The five learning paths

Five self-study routes through what a neocloud operator needs to know — fabric, operations, racks, facility power and cooling. Every step is sourced and marked core, optional, later or skip.

Cooling

liquid to the chip, heat to the air Above roughly 20 to 30 kW per rack, air stops being practical. NVL72-class racks are cooled by liquid at the chip, with a small air remainder. NVIDIA specifies 45 deg C inlet water for GB200, GB300 and Vera Rubin. Thirteen topics are described further down. 10 h · free

Fabric

InfiniBand and Spectrum-X Ethernet How GPUs talk to each other across racks. Two fabrics compete: InfiniBand, and NVIDIA Spectrum-X, which brings lossless Ethernet (RoCE), adaptive routing and telemetry to standard Ethernet. The exam weighs them equally at 30 % each; troubleshooting tools are another 20 %. about 24 h · USD 630 to 660 with the exam

Facility

medium voltage to the rack Grid intake, transformers, switchgear, UPS and batteries, redundancy (N+1 or 2N), PUE and commissioning. The aim here is to brief and challenge design engineers, not to replace them. Paid design credentials teach little beyond the free material at that level. 25 to 35 h · free

Operations

Slurm, Kubernetes and fleet management Running the fleet: provisioning nodes, scheduling jobs, watching GPU health, upgrading and repairing. The NCP-AIO exam is the only one graded partly on a live system, and its heaviest domain (31 %) is installation and deployment with Base Command Manager. 40 to 48 h · USD 550 with the exam

Racks

what an NVL72-class rack needs The numbers a building has to meet: power per rack, coolant flow, inlet temperature, weight, and how far cables can run. All of it is in free NVIDIA and OCP documents. NVIDIA had not published a data-centre design guide for Vera Rubin by September 2026. 15 to 20 h · free