NCA NVIDIA Networking & Interconnects 2 — Questions and Answers
Question 1: What is the maximum bandwidth per port offered by NVIDIA's Quantum-2 InfiniBand switch?
- 200 Gb/s
- 400 Gb/s (Correct answer)
- 800 Gb/s
- 100 Gb/s
Correct answer: 400 Gb/s
NVIDIA's Quantum-2 InfiniBand switches deliver 400 Gb/s per port (NDR InfiniBand).
Question 2: Which NVIDIA networking technology enables GPUs to directly access remote GPU memory without involving the CPU?
- NVLink
- GPUDirect RDMA (Correct answer)
- ConnectX RDMA
- InfiniBand SHARP
Correct answer: GPUDirect RDMA
GPUDirect RDMA allows GPUs to directly read and write remote GPU or NIC memory, bypassing the CPU.
Question 3: What does SHARP stand for in NVIDIA's InfiniBand networking context?
- Scalable Hierarchical Aggregation Reduction Protocol (Correct answer)
- Scalable High-speed Aggregation and Routing Protocol
- Switch Hardware Accelerated Reduction Protocol
- Synchronous High-bandwidth Aggregation Reduction Pipeline
Correct answer: Scalable Hierarchical Aggregation Reduction Protocol
SHARP stands for Scalable Hierarchical Aggregation Reduction Protocol, enabling in-network computing for collective operations.
Question 4: In NVIDIA's networking stack, what is the primary role of the MLNX_OFED driver package?
- GPU memory management
- OpenFabrics Enterprise Distribution for RDMA-capable NIC support (Correct answer)
- Switch firmware management
- Network topology discovery
Correct answer: OpenFabrics Enterprise Distribution for RDMA-capable NIC support
MLNX_OFED is NVIDIA's OpenFabrics Enterprise Distribution providing drivers and libraries for RDMA-capable ConnectX NICs.
Question 5: Which protocol does NVIDIA's RoCE technology use as its transport layer?
- InfiniBand transport
- TCP/IP
- UDP/IP (Correct answer)
- Ethernet without IP
Correct answer: UDP/IP
RoCE v2 uses UDP/IP as its transport layer, enabling RDMA semantics over standard Ethernet infrastructure.
Question 6: What is the function of NVIDIA's Scalable Fabric Manager (UFM) in an InfiniBand cluster?
- GPU workload scheduling
- Subnet management, routing, and fabric monitoring (Correct answer)
- Storage volume management
- Container orchestration
Correct answer: Subnet management, routing, and fabric monitoring
NVIDIA UFM (Unified Fabric Manager) manages InfiniBand subnet routing, monitors fabric health, and provides telemetry.
Question 7: Which NVIDIA ConnectX feature accelerates distributed deep learning by offloading collective communication operations to the NIC?
- RDMA Send/Receive
- SHARP in-network computing (Correct answer)
- SR-IOV virtualization
- DCQCN congestion control
Correct answer: SHARP in-network computing
SHARP offloads collective operations like AllReduce to the network switches and NICs, reducing CPU and GPU overhead.
NCA NVIDIA Networking & Interconnects 2
What is the maximum bandwidth per port offered by NVIDIA's Quantum-2 InfiniBand switch?