NCA Edge AI & Embedded Systems 5 — Questions and Answers
Question 1: A Jetson-based autonomous forklift must process lidar point clouds in real time. Which CUDA library is purpose-built for accelerating point cloud processing on NVIDIA GPUs?
- cuDNN
- cuSPARSE
- CUDA-PCL (CUDA-accelerated Point Cloud Library) (Correct answer)
- TensorRT
Correct answer: CUDA-PCL (CUDA-accelerated Point Cloud Library)
CUDA-PCL provides GPU-accelerated implementations of common point cloud algorithms (voxelization, filtering, clustering) optimized for NVIDIA hardware.
Question 2: What is 'model pruning' in the context of edge AI optimization, and what is its primary benefit?
- Removing entire dataset classes to simplify inference logic
- Eliminating near-zero weights or entire neurons from a trained network to reduce model size and inference compute (Correct answer)
- Converting model weights from FP32 to FP16 format
- Splitting a large model across multiple edge devices
Correct answer: Eliminating near-zero weights or entire neurons from a trained network to reduce model size and inference compute
Pruning removes redundant weights or neurons that contribute little to model outputs, shrinking the model's memory footprint and reducing multiply-accumulate operations.
Question 3: In NVIDIA Jetson's security architecture, what is the purpose of Secure Boot?
- Encrypting inference model weights stored on the device
- Verifying cryptographic signatures on bootloader and OS images to prevent unauthorized firmware from running (Correct answer)
- Isolating CUDA contexts between multiple concurrent AI applications
- Enforcing network firewall rules during device startup
Correct answer: Verifying cryptographic signatures on bootloader and OS images to prevent unauthorized firmware from running
Secure Boot uses a hardware root-of-trust and cryptographic signatures to ensure only authenticated firmware and OS images are executed at startup.
Question 4: Which technique allows a single Jetson device to run multiple independent AI models simultaneously with guaranteed latency isolation between them?
- CUDA Streams with priority scheduling
- Multi-Process Service (MPS)
- CUDA Multi-Instance GPU (MIG) — available on select Orin variants (Correct answer)
- Jetson Clocks frequency pinning
Correct answer: CUDA Multi-Instance GPU (MIG) — available on select Orin variants
MIG (on supported Orin variants) partitions the GPU into isolated instances with dedicated memory and compute, guaranteeing that one workload cannot starve another.
Question 5: A smart city edge node must aggregate inference results from 50 cameras and forward only event metadata to a cloud dashboard. Which edge computing pattern does this describe?
- Fog computing with a gateway aggregator node (Correct answer)
- Pure cloud offload
- Peer-to-peer mesh inference
- Batch cloud training with edge serving
Correct answer: Fog computing with a gateway aggregator node
Fog computing places an intermediate processing layer (gateway) between edge sensors and the cloud to aggregate, filter, and reduce data before upstream transmission.
Question 6: What is the effect of enabling 'workspace size' settings in TensorRT builder configuration on a Jetson device with limited DRAM?
- Larger workspace allows TensorRT to explore more kernel variants during optimization but consumes more DRAM at build time (Correct answer)
- Smaller workspace forces TensorRT to use FP16 precision for all layers
- Workspace size only affects CPU-side preprocessing and has no GPU memory impact
- Increasing workspace size reduces inference latency by pre-loading model weights into L2 cache
Correct answer: Larger workspace allows TensorRT to explore more kernel variants during optimization but consumes more DRAM at build time
A larger TensorRT workspace gives the optimizer more scratch memory to try additional kernel strategies, potentially improving performance, but at the cost of higher peak DRAM usage during engine building.
Question 7: Which Jetson AGX Orin interface is typically used to attach a high-bandwidth NVMe SSD for storing large AI model repositories and video datasets at the edge?
- USB 3.2 Gen 2
- M.2 Key M slot (PCIe Gen 4) (Correct answer)
- MIPI CSI-2
- UART
Correct answer: M.2 Key M slot (PCIe Gen 4)
The M.2 Key M slot on Jetson AGX Orin supports PCIe Gen 4 NVMe SSDs, delivering several GB/s of storage bandwidth suitable for large datasets and model files.
NCA Edge AI & Embedded Systems 5
A Jetson-based autonomous forklift must process lidar point clouds in real time.
Which CUDA library is purpose-built for accelerating point cloud processing on NVIDIA GPUs?