NVIDIA-Certified Associate - Data Science — Questions and Answers

Question 1: What is the purpose of a neural network in deep learning?

To store data.
To manage hardware.
To model complex data relationships (Correct answer)
To encrypt data.

Correct answer: To model complex data relationships

The primary purpose of a neural network in deep learning is to model and learn complex patterns and relationships within data. Inspired by the human brain, these networks consist of interconnected layers of artificial neurons that process information, allowing them to perform tasks like classification, regression, and pattern recognition, often with high accuracy on large and intricate datasets.

Question 2: What does the term 'epoch' mean in deep learning training?

Validation phase.
Model initialization.
One full training data pass (Correct answer)
Single batch processed.

Correct answer: One full training data pass

In deep learning training, an 'epoch' signifies one complete pass through the entire training dataset. During an epoch, the model processes every single training example, updates its internal parameters (weights and biases), and learns from the data. Multiple epochs are typically required for a model to converge and achieve good performance.

Question 3: A retail edge AI system must detect shoplifting events locally and store video clips without sending PII to the cloud. Which architectural pattern best satisfies this requirement?

Batch upload of raw video to cloud for offline analysis
Federated learning with raw video shared across store nodes
On-device inference with local encrypted storage and metadata-only cloud sync (Correct answer)
Cloud-first inference with edge buffering

Correct answer: On-device inference with local encrypted storage and metadata-only cloud sync

Running inference locally and storing only encrypted clips with anonymized metadata keeps PII on-device, satisfying privacy regulations while still enabling analytics.

Question 4: What is the primary purpose of CUDA shared memory in performance optimization?

To increase the number of concurrent thread blocks
To serve as a software-managed cache reducing repeated global memory accesses (Correct answer)
To enable inter-GPU communication
To store kernel code for faster instruction fetch

Correct answer: To serve as a software-managed cache reducing repeated global memory accesses

Shared memory is on-chip SRAM that threads in a block can use to cache frequently accessed data, dramatically reducing costly global memory transactions.

Question 5: What role does continuous improvement play in edge ai & embedded systems for NCA certified professionals?

It is optional and only necessary during certification renewal
It focuses exclusively on cost reduction
It applies only to new professionals in their first year
It drives ongoing enhancement of practices, processes, and outcomes through systematic evaluation (Correct answer)

Correct answer: It drives ongoing enhancement of practices, processes, and outcomes through systematic evaluation

Continuous improvement is fundamental to professional practice in edge ai & embedded systems, involving regular evaluation, feedback integration, and process enhancement to maintain high standards.

Question 6: How does NVIDIA's Spectrum-4 Ethernet switch accelerate AI workloads compared to a generic Ethernet switch?

By bypassing the Ethernet standard entirely
By using InfiniBand protocols internally
By implementing RoCE-aware adaptive routing and per-packet telemetry for congestion visibility (Correct answer)
By running NCCL on the switch CPU

Correct answer: By implementing RoCE-aware adaptive routing and per-packet telemetry for congestion visibility

Spectrum-4 combines hardware-based adaptive routing with granular telemetry that gives operators visibility into per-flow congestion for AI traffic optimization.

Question 7: In the context of NVIDIA GPU execution, what is a 'CUDA stream' and how does it enable performance optimization?

A hardware queue inside each SM that prefetches the next kernel's thread blocks
A shared memory buffer that pipelines data between producer and consumer kernels
A dedicated NVLink channel between two GPUs for overlapping communication and computation
A sequence of CUDA operations that execute in order, allowing overlap of independent streams for concurrency (Correct answer)

Correct answer: A sequence of CUDA operations that execute in order, allowing overlap of independent streams for concurrency

A CUDA stream is an ordered queue of GPU operations; operations in different streams can overlap (kernel execution, memory copies), enabling better GPU utilization through concurrency.

Question 8: Which type of memory is fastest on a GPU?

Shared memory.
Texture memory.
Registers (Correct answer)
Global memory.

Correct answer: Registers

Registers are the fastest type of memory on a GPU, located directly within each CUDA core. They provide extremely low-latency access for individual threads to store and retrieve data during computation. While registers offer the highest speed, their capacity is very limited, making them suitable for frequently accessed variables within a thread's scope.

Question 9: Which NVIDIA interconnect technology enables high-bandwidth GPU-to-GPU communication within a single server node?

InfiniBand HDR
RoCE v2
NVLink (Correct answer)
PCIe Gen 4

Correct answer: NVLink

NVLink is NVIDIA's proprietary high-speed interconnect that allows GPUs within the same node to share memory and data at much higher bandwidth than PCIe.

Question 10: Which CUDA API function launches kernels?

cudaStart().
cudaLaunch().
kernel<<<grid, block>>> (Correct answer)
kernelLaunch() function.

Correct answer: kernel<<<grid, block>>>

In CUDA, kernels are launched using a special syntax called the "execution configuration" or "chevron syntax": `kernel_name<<<gridDim, blockDim>>>(arguments)`. `gridDim` specifies the dimensions of the grid of thread blocks, and `blockDim` specifies the dimensions of each thread block. This syntax is fundamental for defining how a kernel will be executed across the GPU's parallel architecture.

Question 11: In a multi-GPU NLP training setup, what does the 'tensor parallelism' strategy split across GPUs?

Gradient accumulation steps
Individual weight matrices within a single layer (Correct answer)
The tokenizer vocabulary
The dataset batches across GPUs

Correct answer: Individual weight matrices within a single layer

Tensor parallelism partitions individual weight matrices (e.g., attention projections) across multiple GPUs, enabling parallelism within a single transformer layer.

Question 12: When deploying a TensorRT engine on a Jetson device, why must the engine be rebuilt on the target device rather than transferred from a desktop GPU?

Engine files are architecture-specific and not portable across different GPU architectures (Correct answer)
TensorRT engines are encrypted per GPU serial number
TensorRT does not support cross-compilation for ARM targets
Jetson devices use a different version of CUDA than desktop GPUs

Correct answer: Engine files are architecture-specific and not portable across different GPU architectures

TensorRT engines are optimized for the specific GPU architecture and are not portable; an engine built on an Ampere desktop GPU will not run on a Jetson Volta or Ampere variant without rebuilding.

Question 13: What is 'memory coalescing' in the context of NVIDIA GPU global memory access?

Combining multiple L2 cache lines into a single HBM access
Merging consecutive thread memory accesses into a single wide memory transaction (Correct answer)
Prefetching data from system RAM into GPU global memory
Compressing sparse tensors before storing them to VRAM

Correct answer: Merging consecutive thread memory accesses into a single wide memory transaction

Memory coalescing merges individual thread memory requests into a single wide memory transaction when threads in a warp access consecutive addresses, maximizing memory bandwidth utilization.

Question 14: How should NCA professionals handle confidential information related to edge ai & embedded systems?

Delete all records after project completion
Store information without any security measures
Follow established protocols for data protection, access control, and disclosure in accordance with applicable regulations (Correct answer)
Share freely with all colleagues for transparency

Correct answer: Follow established protocols for data protection, access control, and disclosure in accordance with applicable regulations

Confidential information must be handled according to established protocols, regulatory requirements, and professional ethics standards, including proper access control and disclosure procedures.

Question 15: What is the primary role of NVIDIA NVLink in a multi-GPU system compared to PCIe?

NVLink reduces GPU memory capacity to improve speed
NVLink is used only for power delivery between GPUs
NVLink provides higher bandwidth GPU-to-GPU interconnect than PCIe (Correct answer)
NVLink replaces the CPU-to-GPU connection entirely

Correct answer: NVLink provides higher bandwidth GPU-to-GPU interconnect than PCIe

NVLink delivers significantly higher bidirectional bandwidth between GPUs than PCIe, enabling faster peer-to-peer data transfers.

Question 16: What is the primary function of NVIDIA TensorRT in a data center inference pipeline?

Optimizes and compiles trained neural network models for high-throughput, low-latency GPU inference (Correct answer)
Distributes inference requests across multiple GPU nodes using load balancing
Provides a REST API gateway for serving machine learning models
Manages GPU memory allocation across containerized inference services

Correct answer: Optimizes and compiles trained neural network models for high-throughput, low-latency GPU inference

TensorRT takes a trained model and applies optimizations like layer fusion, precision calibration (INT8/FP16), and kernel auto-tuning to maximize inference throughput on NVIDIA GPUs.

Question 17: What is Multi-Instance GPU (MIG) technology introduced with NVIDIA A100, and what problem does it solve?

It merges multiple physical GPUs into a single logical GPU for large model training
It partitions a single GPU into isolated instances with dedicated memory and compute, enabling safe multi-tenant workloads (Correct answer)
It enables dynamic switching between compute and graphics workloads on the same GPU
It allows a single GPU to drive multiple independent display outputs simultaneously

Correct answer: It partitions a single GPU into isolated instances with dedicated memory and compute, enabling safe multi-tenant workloads

MIG partitions a single A100/H100 GPU into up to 7 isolated GPU instances, each with dedicated SM slices and HBM partitions, providing performance isolation for multi-tenant cloud deployments.

Question 18: What is the significance of the 'NVIDIA Deep Learning Frameworks' containers on NGC compared to upstream framework containers?

They are optimized with NVIDIA-specific performance patches, cuDNN tuning, and mixed-precision support not present in upstream images (Correct answer)
They include a paid license key that must be activated before use
They run exclusively on NVIDIA DGX systems and cannot be used on other hardware
They replace standard Python with NVIDIA's proprietary interpreter

Correct answer: They are optimized with NVIDIA-specific performance patches, cuDNN tuning, and mixed-precision support not present in upstream images

NGC framework containers include NVIDIA-specific optimizations such as cuDNN integration, TensorRT support, AMP tuning, and performance patches that are not in community upstream images.

Question 19: When using NVIDIA's TAO Toolkit for transfer learning on a custom object detection dataset, which step directly precedes model export?

Dataset annotation with CVAT
Engine calibration
Vocabulary tokenization
Model pruning (Correct answer)

Correct answer: Model pruning

In the TAO workflow, model pruning removes redundant weights after fine-tuning before the model is exported and optimized with TensorRT.

Question 20: Which of the following best describes a key competency required for computer vision & nlp workloads in NCA practice?

Reliance on a single methodology for all situations
Strong analytical skills combined with effective communication and ethical judgment (Correct answer)
Memorization of all relevant regulations without understanding context
The ability to work independently without any oversight

Correct answer: Strong analytical skills combined with effective communication and ethical judgment

NCA professionals working in computer vision & nlp workloads need analytical skills to assess situations, communication skills to convey findings, and ethical judgment to make sound decisions.

Question 21: When a company needs to ensure that only their NGC API key is used to pull specific private containers, which NGC feature enforces this access control?

CUDA driver-level authentication tokens
NGC Team Spaces with role-based access control (Correct answer)
Container image layer encryption
Docker content trust with Notary signing

Correct answer: NGC Team Spaces with role-based access control

NGC Team Spaces provide organizational access management where admins can assign roles and restrict which users or API keys can access specific private registry content.

Question 22: Which NVIDIA profiling tool shows Tensor Core utilization as a specific metric during training?

Nsight Systems
Nsight Compute (Correct answer)
nvidia-smi
cuobjdump

Correct answer: Nsight Compute

Nsight Compute provides kernel-level metrics including sm__pipe_tensor_cycles_active which directly measures Tensor Core utilization.

Question 23: What is the maximum number of threads per block on modern NVIDIA GPUs?

2048
512
1024 (Correct answer)
256

Correct answer: 1024

Modern NVIDIA architectures support up to 1024 threads per block.

Question 24: In Nsight Systems, what does a gap between consecutive kernel launches in the GPU timeline most likely indicate?

CPU-side launch overhead or synchronization stall between kernels (Correct answer)
The GPU is running out of global memory
Thermal throttling is reducing GPU clock speed
Warp divergence is delaying kernel completion

Correct answer: CPU-side launch overhead or synchronization stall between kernels

Gaps in the GPU timeline between kernels reveal that the CPU is not keeping the GPU fed—often due to synchronization calls, data preparation, or excessive launch overhead.

Question 25: In ONNX (Open Neural Network Exchange), what is the primary purpose of the format?

Defining a standard training loop for neural networks
Providing a GPU-optimized runtime for model inference
Enabling model interoperability across different deep learning frameworks (Correct answer)
Specifying a compression standard for model weights

Correct answer: Enabling model interoperability across different deep learning frameworks

ONNX defines a common format for representing neural network models, allowing models trained in one framework to be deployed in another.

Question 26: Which NVIDIA tool is used to profile GPU kernel performance and identify bottlenecks in deep learning workloads?

cuSPARSE
CUDA GDB
Nsight Systems / Nsight Compute (Correct answer)
NVIDIA-SMI

Correct answer: Nsight Systems / Nsight Compute

Nsight Systems provides system-level profiling while Nsight Compute offers detailed GPU kernel analysis for optimizing deep learning workloads.

Question 27: What is the purpose of cudaMemcpy with the flag cudaMemcpyDeviceToHost?

Copies data within GPU shared memory
Copies data between two GPU devices
Copies data from GPU global memory to host RAM (Correct answer)
Copies data from host RAM to GPU global memory

Correct answer: Copies data from GPU global memory to host RAM

cudaMemcpyDeviceToHost transfers data from device (GPU) memory back to host (CPU) memory.

Question 28: In NLP transformer models, what technique allows processing longer sequences than the model was originally trained on by adjusting positional encodings?

Flash Attention v2
Knowledge distillation
Quantization-aware training
Rotary Position Embedding (RoPE) scaling (Correct answer)

Correct answer: Rotary Position Embedding (RoPE) scaling

RoPE scaling modifies the rotation frequencies of positional encodings to extend context length beyond the training window.

Question 29: Which data type does TensorFloat-32 (TF32) use, and on which NVIDIA GPU generation was it introduced?

10-bit mantissa with 5-bit exponent, introduced on Pascal GPUs
10-bit mantissa with 8-bit exponent, introduced on Ampere GPUs (Correct answer)
23-bit mantissa with 5-bit exponent, introduced on Volta GPUs
16-bit mantissa with 8-bit exponent, introduced on Turing GPUs

Correct answer: 10-bit mantissa with 8-bit exponent, introduced on Ampere GPUs

TF32 uses a 10-bit mantissa and 8-bit exponent (19 bits total) and was introduced on NVIDIA Ampere architecture GPUs to accelerate deep learning.

Question 30: In an NVIDIA DGX H100 system, how many H100 GPUs are interconnected via NVLink?

16
4
32
8 (Correct answer)

Correct answer: 8

The DGX H100 contains 8 H100 SXM5 GPUs fully interconnected via NVLink 4.0 and NVSwitch for a combined 640 GB of HBM3 memory.

Question 31: What problem does 'activation recomputation' (gradient checkpointing) solve in multi-GPU training?

It eliminates gradient divergence between GPU ranks
It speeds up AllReduce by caching intermediate gradients
It reduces GPU memory usage by recomputing activations during the backward pass instead of storing them (Correct answer)
It prevents NaN loss values during mixed-precision training

Correct answer: It reduces GPU memory usage by recomputing activations during the backward pass instead of storing them

Gradient checkpointing trades compute for memory by discarding intermediate activations and recomputing them during backpropagation.

Question 32: When performing over-the-air (OTA) updates on an embedded Jetson device, which practice is most critical for maintaining system reliability?

Applying updates directly to the running root filesystem without rebooting
Using an A/B partition scheme so the device can revert to the previous known-good image if an update fails (Correct answer)
Downloading updates through an unencrypted HTTP channel to reduce latency
Disabling update verification to speed up the OTA process

Correct answer: Using an A/B partition scheme so the device can revert to the previous known-good image if an update fails

An A/B (dual-partition) scheme ensures the device always has a fallback boot partition, preventing a failed update from bricking the device.

Question 33: Which NVIDIA GPU architecture first introduced Tensor Cores for accelerating matrix multiply-accumulate operations?

Volta (Correct answer)
Turing
Maxwell
Pascal

Correct answer: Volta

NVIDIA Volta (V100) introduced the first generation of Tensor Cores, providing hardware acceleration for mixed-precision matrix operations used in deep learning.

Question 34: Which TensorFlow API is recommended for building models using a layer-by-layer approach with maximum flexibility for custom training loops?

tf.keras.Model subclassing (Correct answer)
tf.keras.Sequential
tf.estimator.Estimator
tf.compat.v1.Session

Correct answer: tf.keras.Model subclassing

Subclassing `tf.keras.Model` gives maximum flexibility, allowing custom forward passes and fine-grained control over training logic.

Question 35: Which activation function outputs values between 0 and 1?

Linear.
Sigmoid (Correct answer)
Tanh.
ReLU.

Correct answer: Sigmoid

The Sigmoid activation function outputs values that are squashed into a range between 0 and 1. This characteristic makes it particularly useful in the output layer of binary classification problems, where the output can be interpreted as a probability. While once popular in hidden layers, its vanishing gradient problem has led to ReLU being preferred for those layers.

Question 36: Which NGC CLI command lists all available versions of a specific container image?

ngc registry image list <org>/<image> (Correct answer)
ngc registry image info <org>/<image>
ngc container list --image <org>/<image>
ngc image versions <org>/<image>

Correct answer: ngc registry image list <org>/<image>

The `ngc registry image list <org>/<image>` command displays all available tags and versions for a specified container image in the NGC registry.

Question 37: What is a CUDA kernel?

A GPU function executed in parallel (Correct answer)
A graphics shader.
A memory allocation routine.
A CPU function.

Correct answer: A GPU function executed in parallel

In CUDA, a kernel is a function written in CUDA C/C++ that is executed on the GPU device by many threads in parallel. When a kernel is launched from the CPU (host), it specifies how many threads and blocks should execute it. Each thread then executes the kernel function independently, often operating on a different piece of data.

Question 38: For fine-tuning a large language model with limited GPU memory, which parameter-efficient method trains only a small set of adapter weights inserted between transformer layers?

Quantization-aware training
LoRA (Low-Rank Adaptation) (Correct answer)
Full fine-tuning
Pipeline parallelism

Correct answer: LoRA (Low-Rank Adaptation)

LoRA freezes the original model weights and injects trainable low-rank decomposition matrices into attention layers, drastically reducing the number of trainable parameters.

Question 39: What is the purpose of Tensor Cores in NVIDIA GPUs like Volta, Turing, Ampere, and Hopper?

Execute SIMD integer operations for scientific computing
Accelerate ray-triangle intersection tests for real-time rendering
Perform high-throughput matrix multiply-accumulate (MMA) operations for AI workloads (Correct answer)
Decode compressed video streams in hardware

Correct answer: Perform high-throughput matrix multiply-accumulate (MMA) operations for AI workloads

Tensor Cores are specialized hardware units designed to execute mixed-precision matrix multiply-accumulate operations at very high throughput, accelerating deep learning training and inference.

Question 40: In NVIDIA's networking stack, what is the primary role of the MLNX_OFED driver package?

Network topology discovery
Switch firmware management
OpenFabrics Enterprise Distribution for RDMA-capable NIC support (Correct answer)
GPU memory management

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 41: When authenticating to the NGC container registry via Docker, which registry hostname is used?

nvcr.io (Correct answer)
docker.ngc.io
hub.ngc.nvidia.com
registry.nvidia.com

Correct answer: nvcr.io

NGC container images are hosted at nvcr.io, so Docker login and pull commands must reference this hostname.

Question 42: How should NCA professionals handle confidential information related to nvidia networking & interconnects?

Share freely with all colleagues for transparency
Delete all records after project completion
Store information without any security measures
Follow established protocols for data protection, access control, and disclosure in accordance with applicable regulations (Correct answer)

Correct answer: Follow established protocols for data protection, access control, and disclosure in accordance with applicable regulations

Confidential information must be handled according to established protocols, regulatory requirements, and professional ethics standards, including proper access control and disclosure procedures.

Question 43: What does NVIDIA TensorRT primarily optimize when deploying deep learning models?

Multi-GPU gradient synchronization
Training throughput by fusing optimizer steps
Data pipeline preprocessing speed
Inference latency and throughput through layer fusion, precision calibration, and kernel auto-tuning (Correct answer)

Correct answer: Inference latency and throughput through layer fusion, precision calibration, and kernel auto-tuning

TensorRT optimizes trained models for deployment by fusing layers, selecting optimal kernels, and applying INT8/FP16 quantization to maximize inference performance.

Question 44: Which framework introduced the concept of 'eager execution' as its default mode, executing operations immediately without building a static graph?

Caffe2
Theano
MXNet
TensorFlow 2.x (Correct answer)

Correct answer: TensorFlow 2.x

TensorFlow 2.x made eager execution the default, allowing operations to run immediately like NumPy, which greatly simplified debugging and experimentation.

Question 45: When using GradScaler with AMP in PyTorch, what problem does it solve?

Gradient underflow (vanishing gradients) in FP16 (Correct answer)
Gradient explosion in FP32
Incompatibility between cuBLAS and cuDNN
Slow GPU memory allocation

Correct answer: Gradient underflow (vanishing gradients) in FP16

GradScaler scales the loss before backward pass to prevent FP16 gradients from underflowing to zero during backpropagation.

Question 46: In data parallelism across multiple GPUs, which operation is performed after each training batch to synchronize model parameters?

Gradient averaging (AllReduce) (Correct answer)
Weight pruning
Tensor transposition
Activation checkpointing

Correct answer: Gradient averaging (AllReduce)

AllReduce aggregates gradients from all GPUs and distributes the averaged result so each GPU updates identically.

Question 47: What is the primary role of NVIDIA's NVSwitch in large-scale GPU deployments?

Offloads network packet processing from GPUs
Acts as a high-bandwidth switch enabling all-to-all NVLink connectivity among multiple GPUs (Correct answer)
Manages GPU firmware updates across the data center
Provides power regulation for GPU clusters

Correct answer: Acts as a high-bandwidth switch enabling all-to-all NVLink connectivity among multiple GPUs

NVSwitch is a high-speed switch chip that connects multiple GPUs via NVLink, enabling full all-to-all bandwidth across all GPUs in a node like the DGX H100.

Question 48: Which precision format introduced in Hopper GPUs is specifically optimized for transformer inference?

TF32
FP16
FP8 (Correct answer)
INT4

Correct answer: FP8

Hopper (H100) added FP8 precision (E4M3 and E5M2 variants) Tensor Core support, targeting high-throughput transformer inference.

Question 49: What is the difference between cudaMemcpy and cudaMemcpyAsync?

cudaMemcpy uses pinned memory; Async uses pageable memory
They are identical in behavior
cudaMemcpy works only for host-to-device; Async works for all directions
cudaMemcpyAsync is non-blocking and can overlap with kernel execution in a stream (Correct answer)

Correct answer: cudaMemcpyAsync is non-blocking and can overlap with kernel execution in a stream

cudaMemcpyAsync returns immediately and can be issued to a stream, enabling overlap with compute or other transfers.

Question 50: What does NCCL stand for in the NVIDIA ecosystem?

NVIDIA CUDA Communication Link
NVIDIA CUDA Compute Library
NVIDIA Collective Communications Library (Correct answer)
NVIDIA Core Clustering Layer

Correct answer: NVIDIA Collective Communications Library

NCCL (NVIDIA Collective Communications Library) implements multi-GPU and multi-node collective operations such as AllReduce, Broadcast, and AllGather, optimized for NVIDIA interconnects.

Question 51: Which scaling strategy is most appropriate when a single GPU cannot hold the entire model due to memory constraints?

Batch size reduction only
Data parallelism
Model parallelism (pipeline or tensor) (Correct answer)
Gradient checkpointing without any parallelism

Correct answer: Model parallelism (pipeline or tensor)

Model parallelism partitions model layers or tensors across multiple GPUs, allowing models larger than a single GPU's memory to be trained.

Question 52: What is the attention mechanism in transformer models computing with its Query, Key, and Value matrices?

Applying recurrent state transitions to sequential input tokens
Convolving input features across spatial dimensions using learned filters
Normalizing feature distributions across the batch dimension
Computing weighted sums of value vectors, where weights derive from query-key dot product similarities (Correct answer)

Correct answer: Computing weighted sums of value vectors, where weights derive from query-key dot product similarities

Self-attention computes dot products between queries and keys to produce attention weights, then uses those weights to aggregate value vectors into context-aware representations.

Question 53: What does occupancy mean in CUDA programming?

Ratio of active warps to max warps (Correct answer)
The percentage of GPU clock speed.
Number of threads per block.
The amount of memory used.

Correct answer: Ratio of active warps to max warps

Occupancy in CUDA programming refers to the ratio of the number of active warps on a Streaming Multiprocessor (SM) to the maximum number of warps that the SM can support. High occupancy indicates that the SM is effectively utilized, as it has many warps ready to execute, which helps hide memory latency and keep the execution units busy. Achieving optimal occupancy is crucial for maximizing GPU performance.

Question 54: Which layer type is commonly used in Convolutional Neural Networks (CNNs)?

Dropout layer.
Recurrent layer.
Convolutional layer (Correct answer)
Fully connected layer.

Correct answer: Convolutional layer

The convolutional layer is a core component commonly used in Convolutional Neural Networks (CNNs), especially for tasks involving image and video processing. This layer applies a set of learnable filters (kernels) to the input data, performing a convolution operation to extract features such as edges, textures, or patterns. These extracted features are then passed to subsequent layers for further processing and analysis.

Question 55: What is the most effective way to measure success in tensor core & mixed precision within NCA professional practice?

Rely solely on supervisor opinion
Count only the number of activities completed
Compare only with industry averages without considering context
Use a combination of quantitative metrics, qualitative assessments, and stakeholder feedback aligned with defined objectives (Correct answer)

Correct answer: Use a combination of quantitative metrics, qualitative assessments, and stakeholder feedback aligned with defined objectives

Effective measurement combines multiple data sources — quantitative metrics, qualitative assessments, and stakeholder feedback — all aligned with clearly defined objectives for a comprehensive evaluation.

Question 56: Which type of memory has the highest latency in CUDA?

Texture memory.
Shared memory.
Register memory.
Global memory (Correct answer)

Correct answer: Global memory

Global memory has the highest latency among the commonly used CUDA memory types. It resides off-chip in the device's DRAM and is accessible by all threads across all blocks, but accessing it is significantly slower than on-chip memories like registers or shared memory. Minimizing global memory accesses and maximizing data reuse in faster memories is a critical optimization strategy in CUDA programming.

Question 57: When a NCA professional encounters an unfamiliar challenge in data center gpu solutions, what is the recommended first course of action?

Proceed based on personal intuition alone
Postpone addressing the issue indefinitely
Apply the solution used for the most recent similar problem without adaptation
Research applicable standards, consult with subject matter experts, and document the approach (Correct answer)

Correct answer: Research applicable standards, consult with subject matter experts, and document the approach

Professional practice requires a methodical approach to unfamiliar challenges: research the applicable standards, consult experts when needed, and document the reasoning for the chosen approach.

Question 58: What is the purpose of the CUDA occupancy calculator (cudaOccupancyMaxActiveBlocksPerMultiprocessor)?

To compute the optimal grid size for a given problem
To measure live warp count during execution
To predict memory bandwidth utilization
To determine how many blocks can simultaneously run on an SM given resource usage (Correct answer)

Correct answer: To determine how many blocks can simultaneously run on an SM given resource usage

This API computes the maximum number of active blocks per SM based on register and shared memory usage of the kernel.

Question 59: What is the effect of increasing the number of GPUs in a data-parallel setup if the per-GPU batch size is kept constant?

Per-GPU memory usage decreases because gradients are shared
Training throughput stays flat because AllReduce becomes the bottleneck immediately
Model accuracy automatically improves proportionally
Effective global batch size increases, which may require learning rate scaling (Correct answer)

Correct answer: Effective global batch size increases, which may require learning rate scaling

Keeping per-GPU batch size fixed while adding GPUs scales the global batch size, which typically requires linear or square-root learning rate scaling to maintain convergence.

Question 60: How does CUDA handle synchronization between threads?

Using mutexes only.
Using __syncthreads() (Correct answer)
No synchronization is possible.
Threads run independently.

Correct answer: Using __syncthreads()

In CUDA, `__syncthreads()` is a barrier synchronization primitive used to ensure that all threads within the same thread block have completed their operations up to that point before any thread proceeds further. This is essential for correctness when threads within a block need to share data via shared memory or depend on each other's results. It does not synchronize threads across different blocks.

NVIDIA-Certified Associate - Data Science

The NVIDIA-Certified Associate - Data Science certification validates foundational knowledge in data science, including data manipulation, machine learning, and deep learning using NVIDIA tools and platforms.

Questions

90 min

Time Limit

70%

To Pass

Checking for saved progress...

Exam Rules

You can skip questions and return to them later
Flag questions for review before submitting
No feedback shown until you submit the entire exam
Unanswered questions count as wrong — answer everything
10 pretest questions are mixed in and don't affect your score
Timer auto-submits when time runs out
Your progress is auto-saved every 30 seconds