Describe Network Architecture Used with AI Workloads (Ethernet vs Infiniband, RDMA, GPUDirect Storage) CERTIFIED-AI-EXPERT Practice Questions (Page 4)

Questions 16–20

1. 16expert · hard

   A company is deploying a new AI cluster with RoCEv2. They are experiencing packet drops and performance degradation during heavy training traffic. The network team has verified that all switches and NICs support RoCEv2. What is the most likely root cause of the packet drops?

   • AThe network is using a fat-tree topology, which is not suitable for RoCEv2
   • BThe NICs are not using the correct RDMA drivers
   • CThe network is not configured with lossless QoS policies like Priority Flow Control (PFC)
   • DThe application is not using the RDMA API correctly
   Reveal Solution Discussion Select an answer first
2. 17application · medium

   A storage engineer is configuring a new AI storage system. They want to enable GPUDirect Storage (GDS) to improve data loading performance. The storage system is connected to the GPU servers via a RoCEv2 network. What is the primary requirement for the storage system's network interface to support GDS?

   • AIt must use a proprietary driver from the GPU vendor
   • BIt must be configured with a larger MTU than the compute network
   • CIt must be connected to the same switch as the GPU servers
   • DIt must support RDMA operations
   Reveal Solution Discussion Select an answer first
4. 18foundation · easy

   Which network architecture is typically associated with lower per-port cost and broader ecosystem compatibility, making it a common choice for Ethernet-based AI clusters?

   • AOmni-Path
   • BInfiniBand
   • CEthernet
   • DFibre Channel
   Reveal Solution Discussion Select an answer first
5. 19foundation · easy

   How does GPUDirect Storage impact the performance of AI training workloads?

   • AIt reduces the time needed to load training data into GPU memory.
   • BIt reduces the amount of GPU memory required for training.
   • CIt increases the GPU's compute clock speed.
   • DIt eliminates the need for data preprocessing.
   Reveal Solution Discussion Select an answer first
6. 20expert · hard

   A large enterprise is building a multi-tenant AI platform. The platform will host both tightly coupled distributed training jobs (requiring low latency) and loosely coupled inference workloads (requiring high bandwidth but tolerant of latency). The networking team is proficient in Ethernet, but the budget is not a constraint. They also plan to use GPUDirect Storage (GDS) for data loading. Which architecture best balances performance and operational complexity?

   • AA single InfiniBand fabric for all workloads, because it provides the best performance for both training and inference
   • BTwo separate fabrics: InfiniBand for training and Ethernet for inference, because it isolates performance-sensitive traffic
   • CTwo separate fabrics: Ethernet for training and InfiniBand for inference, because it aligns with the team's expertise
   • DA single Ethernet fabric with RoCEv2 for all workloads, because it provides RDMA benefits and simplifies management
   Reveal Solution Discussion Select an answer first