NVLink and GPU Interconnects Quiz

Answers and explanations appear as you take the quiz.

1. Why are high-speed interconnects like NVLink needed for large AI models?
   • Because the largest models are split across many GPUs that must constantly exchange data
   • Because one GPU is always enough but it is lonely
   • Because GPUs need to connect to printers
   • Because models are stored on the interconnect itself
2. What advantage does NVLink offer over the standard PCIe bus for GPU-to-GPU communication?
   • It is slower but cheaper
   • Much higher bandwidth and lower latency
   • It only works between a GPU and a CPU
   • It removes the need for any memory
3. What is the role of an NVSwitch in a multi-GPU server?
   • It cools the GPUs with fans
   • It stores the operating system
   • It acts as a high-speed crossbar so every GPU can reach every other at full bandwidth
   • It converts power from the wall outlet
4. Which collective operation, common in distributed training, benefits strongly from fast interconnects?
   • Defragmenting a hard drive
   • Rendering a single pixel
   • Compiling source code
   • All-reduce, which sums gradients across GPUs
5. Beyond a single server, what technologies typically connect thousands of GPU nodes into one cluster?
   • InfiniBand and RDMA-over-Ethernet
   • USB cables and Wi-Fi
   • Bluetooth and HDMI
   • Dial-up modems
6. What does a rack-scale system like NVIDIA's NVL72 achieve?
   • It is a single GPU in a small box
   • It connects dozens of GPUs into one unified NVLink domain
   • It replaces all software with hardware
   • It is a type of cooling fluid
7. Why can a fast GPU fabric improve distributed data-parallel training?
   • It removes the need to compute gradients
   • It converts training into a database lookup
   • It can reduce the time spent exchanging gradients in collectives such as all-reduce
   • It prevents GPUs from synchronizing
8. How does an NVLink scale-up fabric differ from InfiniBand or Ethernet used across a cluster?
   • NVLink is only a file format
   • InfiniBand runs inside a GPU’s tensor core
   • Ethernet is used only for local memory access
   • NVLink connects GPUs within supported systems or rack domains; InfiniBand and Ethernet commonly connect cluster nodes