NCA NVIDIA Networking & Interconnects 3 — Questions and Answers
Question 1: What is the key advantage of NVLink Switch System (NVL72) compared to PCIe-based GPU interconnects?
- Lower cost per port
- Higher bandwidth and lower latency for all-to-all GPU communication (Correct answer)
- Longer cable reach
- Support for CPU-to-GPU connections only
Correct answer: Higher bandwidth and lower latency for all-to-all GPU communication
NVLink Switch System provides terabits-per-second all-to-all bandwidth between GPUs with microsecond latency, far exceeding PCIe.
Question 2: In InfiniBand networking, what is a 'fat-tree' topology primarily designed to achieve?
- Minimize cable length
- Provide non-blocking full bisection bandwidth (Correct answer)
- Reduce switch port count
- Simplify routing table management
Correct answer: Provide non-blocking full bisection bandwidth
Fat-tree topologies are designed to provide full bisection bandwidth, ensuring no bottleneck regardless of traffic pattern.
Question 3: Which NVIDIA technology allows network adapters to process Ethernet frames and perform RDMA operations simultaneously on the same port?
- Dual-port bonding
- RoCE (RDMA over Converged Ethernet) (Correct answer)
- VXLAN offload
- DPDK acceleration
Correct answer: RoCE (RDMA over Converged Ethernet)
RoCE enables RDMA semantics over standard Ethernet, allowing the same physical port to handle both standard Ethernet and RDMA traffic.
Question 4: What is the purpose of Priority Flow Control (PFC) in a RoCE deployment?
- Prioritize GPU traffic over CPU traffic
- Prevent packet loss on lossless queues required for RDMA (Correct answer)
- Load balance traffic across multiple paths
- Encrypt sensitive RDMA data
Correct answer: Prevent packet loss on lossless queues required for RDMA
PFC creates lossless Ethernet queues by sending pause frames when buffers fill, which is required for reliable RDMA operation.
Question 5: How many GPUs can be directly interconnected in a single NVLink domain using NVIDIA's H100 SXM platform?
- 4 GPUs
- 8 GPUs (Correct answer)
- 16 GPUs
- 32 GPUs
Correct answer: 8 GPUs
H100 SXM servers connect 8 GPUs within a single NVLink domain using NVSwitch chips for all-to-all connectivity.
Question 6: What distinguishes NVIDIA's Spectrum-X Ethernet platform from standard data center Ethernet switches?
- It uses fiber-only connections
- It includes AI-optimized congestion control and adaptive routing for GPU workloads (Correct answer)
- It operates at layer 2 only
- It requires proprietary transceivers
Correct answer: It includes AI-optimized congestion control and adaptive routing for GPU workloads
Spectrum-X combines RoCE-optimized adaptive routing and congestion control algorithms specifically tuned for AI/ML workload traffic patterns.
Question 7: In NVIDIA GPUDirect Storage, what path does data take when reading from NVMe SSDs to GPU memory?
- NVMe → CPU RAM → GPU memory
- NVMe → GPU memory directly via PCIe peer-to-peer (Correct answer)
- NVMe → NIC RDMA buffer → GPU memory
- NVMe → CPU cache → PCIe → GPU memory
Correct answer: NVMe → GPU memory directly via PCIe peer-to-peer
GPUDirect Storage enables direct DMA transfers from NVMe SSDs to GPU memory over PCIe, bypassing CPU RAM entirely.
NCA NVIDIA Networking & Interconnects 3
What is the key advantage of NVLink Switch System (NVL72) compared to PCIe-based GPU interconnects?