Describe Network Architecture Used with AI Workloads (Ethernet vs Infiniband, RDMA, GPUDirect Storage) CERTIFIED-AI-EXPERT Practice Questions (Page 2)

Questions 6–10

1. 6application · medium

   A company is planning to deploy a new AI cluster with GPUDirect Storage (GDS). They are evaluating network options and have narrowed it down to InfiniBand and RoCEv2. Their primary concern is the total cost of ownership (TCO), including hardware, management, and operational expertise. Which network choice is likely to have a lower TCO for their team?

   • AInfiniBand, because it requires less configuration than RoCEv2
   • BRoCEv2, because it leverages existing Ethernet skills and commodity hardware
   • CInfiniBand, because it is a mature technology with lower hardware costs
   • DRoCEv2, because it does not require any specialized configuration
   Reveal Solution Discussion Select an answer first
2. 7application · easy

   A data engineering team is loading a 10TB dataset from a storage array into GPU memory for a training job. The current process is slow because the data must be read from storage into system memory and then copied to the GPU. Which technology would allow them to bypass the system memory copy and improve performance?

   • ANVMe over Fabrics (NVMe-oF)
   • BA faster PCIe Gen5 interconnect for the storage array
   • CRDMA over InfiniBand
   • DGPUDirect Storage (GDS)
   Reveal Solution Discussion Select an answer first
4. 8foundation · easy

   Which network architecture is designed with a lossless, low-latency fabric that is often used in high-performance computing (HPC) clusters for AI training?

   • AWi-Fi
   • BInfiniBand
   • CEthernet
   • DToken Ring
   Reveal Solution Discussion Select an answer first
5. 9application · medium

   A company is designing a new AI cluster and wants to use RDMA to reduce CPU overhead during data transfers. Their existing data center networking team is highly proficient with Ethernet and has a strong monitoring stack for it. They also have a budget constraint that favors commodity hardware. Which network architecture best meets these constraints?

   • AInfiniBand, because it is the only network technology that supports RDMA
   • BEthernet with TCP/IP, because it provides the same low-latency benefits as RDMA without the need for specialized NICs
   • CInfiniBand, because it is cheaper than Ethernet for the same port count
   • DEthernet with RoCEv2, because it provides RDMA capabilities while leveraging existing operational expertise and commodity hardware
   Reveal Solution Discussion Select an answer first
6. 10application · medium

   An AI engineer is tuning a training pipeline. They notice that the GPU is often idle while waiting for data to be loaded from a remote storage system. They suspect the bottleneck is the CPU copying data from the network buffer to system memory and then to GPU memory. Which configuration would most directly address this issue?

   • AUse a faster network protocol like UDP instead of TCP
   • BIncrease the number of CPU cores on the GPU servers
   • CIncrease the GPU memory size
   • DEnable GPUDirect Storage (GDS) on the storage and GPU servers
   Reveal Solution Discussion Select an answer first