Describe Network Architecture Used with AI Workloads (Ethernet vs Infiniband, RDMA, GPUDirect Storage) CERTIFIED-AI-EXPERT Practice Questions (Page 1)

Questions 1–5

1. 1application · medium

   A data scientist is running a training job that reads large datasets from a parallel file system. The job is bottlenecked by the time it takes to move data from storage to GPU memory. They are considering using GPUDirect Storage (GDS) to improve performance. Which network architecture is required to fully benefit from GDS?

   • AA network that supports RDMA, such as InfiniBand or RoCEv2, because GDS requires RDMA to bypass the CPU
   • BA standard Ethernet network with TCP/IP, because GDS works over any network
   • CA dedicated storage network separate from the compute network, because GDS requires a separate physical fabric
   • DAn InfiniBand network, because GDS is only supported on InfiniBand
   Reveal Solution Discussion Select an answer first
2. 2foundation · easy

   What is the main purpose of GPUDirect Storage (GDS) in AI workloads?

   • ATo encrypt data stored on disk
   • BTo enable direct data transfer between GPU memory and storage, bypassing the CPU
   • CTo replace the need for network storage
   • DTo increase the storage capacity of GPUs
   Reveal Solution Discussion Select an answer first
4. 3application · medium

   A data scientist is running a training job that reads a large dataset from a network-attached storage system. They notice that the GPU utilization is low, and the CPU is heavily utilized. They suspect the data loading path is the bottleneck. Which technology would most directly help them improve GPU utilization?

   • AMore system memory, because it can cache more of the dataset
   • BA faster network connection, because it can transfer the data faster
   • CA faster CPU, because it can process the data loading faster
   • DGPUDirect Storage (GDS), because it enables direct data transfer from storage to GPU memory, bypassing the CPU
   Reveal Solution Discussion Select an answer first
5. 4expert · hard

   An AI company is planning to deploy GPUDirect Storage (GDS) to speed up data loading. Their storage system is connected via a standard TCP/IP network. They are considering upgrading to either InfiniBand or RoCEv2 to enable GDS. The storage system is older and only supports RoCEv1. What is the best course of action?

   • AUpgrade the storage system's NICs to support RoCEv2, then deploy GDS over the RoCEv2 network
   • BReplace the entire storage system with a new one that supports RoCEv2
   • CDeploy GDS over the existing TCP/IP network, because GDS can work with any network protocol
   • DDeploy GDS over the RoCEv1 network, because it is compatible with the storage system
   Reveal Solution Discussion Select an answer first
6. 5application · medium

   An AI platform team is evaluating ways to reduce CPU overhead on their GPU servers. They notice that during distributed training, the CPU is heavily loaded with copying data between kernel buffers and application memory for network communication. Which technology is specifically designed to address this overhead?

   • AA higher-speed Ethernet link, because it reduces the time the CPU spends on network processing
   • BA larger CPU cache, because it speeds up the memory copies required for network communication
   • CTCP/IP offloading, because it moves all network processing to the NIC, eliminating CPU involvement entirely
   • DRDMA, because it allows the NIC to transfer data directly between the application's memory and the network, bypassing the CPU and kernel
   Reveal Solution Discussion Select an answer first