NVIDIA NCP-AIN study plan

Map the OSI and TCP/IP layers to concrete NVIDIA networking components, writing one sentence per layer explaining what runs there

Explain how RoCEv2 encapsulates InfiniBand transport over UDP/IP, including the roles of IB BTH, GRH, and the UDP destination port

Differentiate lossless Ethernet mechanisms: PFC priority flow control, ECN marking, and DCQCN congestion control, and state when each one engages

Trace a packet from a GPU host NIC through a leaf-spine fabric to a remote GPU, naming every hop and the decision made at each hop

Configure and verify a basic VLAN and MTU 9000 jumbo frame path on a lab switch or simulator

Describe InfiniBand fabric concepts: subnet manager, LID vs GID addressing, partitions, and adaptive routing

Compare InfiniBand and Ethernet for GPU-to-GPU traffic using latency, throughput, and congestion-management criteria

Identify which NVIDIA products serve which role: Spectrum Ethernet switches, Quantum InfiniBand switches, ConnectX adapters, and BlueField DPUs

Explain GPUDirect RDMA and GPUDirect Storage data paths and the prerequisites each one needs

Diagnose three given fabric symptoms (credit stalls, PFC pause storms, ECN threshold misconfiguration) and propose a fix for each

Compare air cooling and direct-to-chip liquid cooling for high-density GPU racks, citing practical rack power density ranges and the tradeoffs of each

Explain why rear-door heat exchangers and coolant distribution units exist and where each fits in a datacenter topology

Compute usable power capacity for a rack given PDU rating, redundancy mode, and power factor, showing the arithmetic

Describe redundant power topologies (N, N+1, 2N) and map each to an availability tier and failure response

Identify the environmental monitoring signals (inlet temperature, humidity, leak detection, differential pressure) that precede a GPU thermal throttle event

Explain how rail-optimized network topology maps to physical rack layout and why rail alignment reduces cross-switch hops

Describe the cabling and optics choices for short intra-rack, medium intra-row, and long inter-row links, naming transceiver classes and reach limits

Plan cable management and airflow containment so that hot-aisle recirculation does not degrade GPU inlet temperatures

Draft the physical port map for a leaf-spine fabric spanning two rows, assigning each GPU rail to a leaf switch

State how a single leaf switch failure propagates through the fabric and what redundancy absorbs it

Distinguish training from inference workloads by memory, compute, and interconnect demands, and match each to an appropriate GPU class

Explain mixed-precision training: FP32, TF32, FP16, BF16, and FP8, including where loss scaling is required and where it is not

Describe data parallelism, tensor parallelism, pipeline parallelism, and fully sharded data parallelism, and identify the communication pattern each one generates

Compute the memory footprint of a transformer training step given parameter count, optimizer state, gradients, and activations, showing the formula

Explain how KV cache size scales with batch size, sequence length, and model dimension, and why it constrains inference throughput

Explain NVIDIA AI Enterprise software stack components and which parts are licensed versus open source

Describe the purpose of NVIDIA NIM microservices and Triton Inference Server in deploying models, and how they differ

Outline how GPU scheduling works in Kubernetes via the device plugin and GPU operator, including time-slicing and MIG partitioning

Compare MIG partitioning with whole-GPU allocation for multi-tenant inference, citing isolation and utilization tradeoffs

Identify the metrics that reveal GPU underutilization and memory pressure during a serving workload

Compare self-managed, cloud-hosted, and managed Kubernetes GPU offerings on control, cost model, and operational burden

Describe how GPU instance types differ across major cloud providers in interconnect, memory, and availability constraints

Explain identity, network isolation, and data protection controls relevant to a GPU training cluster in a cloud tenancy

Estimate total cost of ownership for a small training cluster on cloud versus on-premises, including networking and storage

Describe how NVIDIA NGC catalog, containers, and model artifacts move between on-premises and cloud environments

Re-derive the core quantitative relationships from memory: training memory footprint, KV cache size, rack power budget, and fabric oversubscription ratio

Diagnose five integrated scenarios that combine networking, datacenter, AI, and cloud factors, and write the first diagnostic step for each

Close every gap flagged during earlier weeks by rewriting the weakest deliverable in each domain

Time yourself on a mixed-domain practice set and classify every miss as knowledge gap, misread, or reasoning error