NVIDIA-Certified Associate - Data Science — Questions and Answers

Question 1: How should NCA professionals handle confidential information related to data center gpu solutions?

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

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 2: What is backpropagation used for?

Model deployment.
Data visualization.
Weight updates using gradients (Correct answer)
Data preprocessing.

Correct answer: Weight updates using gradients

Backpropagation is a fundamental algorithm used in training artificial neural networks to efficiently calculate the gradients of the loss function with respect to the network's weights. These gradients indicate the direction and magnitude by which the weights should be adjusted to minimize the loss. By iteratively updating the weights based on these gradients, backpropagation enables the network to learn from its errors and improve its performance.

Question 3: What does kernel occupancy indicate?

Percentage of CPU used.
Amount of memory allocated.
Ratio of active warps to max warps (Correct answer)
Number of threads created.

Correct answer: Ratio of active warps to max warps

Kernel occupancy refers to the ratio of active warps on a multiprocessor to the maximum number of active warps that the multiprocessor can support. High occupancy generally indicates that the GPU is being effectively utilized, as it helps hide memory latency by allowing the scheduler to switch between active warps. It's a key metric for understanding how well a kernel is leveraging GPU resources.

Question 4: What is the significance of the 'ea' tag suffix on some NGC container images?

It signifies an Enterprise Authorized version
It denotes an End-of-life Archived build
It marks an Encrypted Archive for secure distribution
It indicates an Early Access release not yet fully production-ready (Correct answer)

Correct answer: It indicates an Early Access release not yet fully production-ready

The 'ea' tag suffix stands for Early Access, indicating the container is a preview release that may not be fully stable for production use.

Question 5: Which NVIDIA software platform provides a unified ecosystem of GPU-accelerated libraries, frameworks, and tools for data center AI workloads?

NVIDIA Omniverse
CUDA-X AI (Correct answer)
NVIDIA RAPIDS only
NVIDIA Fleet Command

Correct answer: CUDA-X AI

CUDA-X AI is NVIDIA's collection of GPU-accelerated libraries and SDKs (cuDNN, cuBLAS, TensorRT, etc.) that power AI and HPC workloads across data centers.

Question 6: A developer notices that cudaMemcpy from device to host takes longer than the same-size host-to-device transfer. What is the most probable reason?

PCIe bandwidth is asymmetric by design
The GPU memory is fragmented
ECC correction is adding overhead
The host memory is not page-locked (pinned) (Correct answer)

Correct answer: The host memory is not page-locked (pinned)

Non-pinned (pageable) host memory requires an extra staging buffer copy through the OS, adding overhead especially noticeable on device-to-host transfers.

Question 7: What is the difference between cudaMemcpy and cudaMemcpyAsync?

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

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 8: What does the CUDA programming model term 'grid' refer to?

The shared memory region of an SM
A single thread executing on the GPU
The collection of all thread blocks launched by a kernel (Correct answer)
A group of 32 threads executing in lockstep

Correct answer: The collection of all thread blocks launched by a kernel

A grid is the complete set of thread blocks that are launched together when a kernel is invoked.

Question 9: For image segmentation tasks, which NVIDIA pretrained model foundation available in NGC is optimized for rapid fine-tuning on medical and scientific imaging?

NeMo
RAPIDS cuDF
TAO Toolkit SegFormer
MONAI (Correct answer)

Correct answer: MONAI

MONAI (Medical Open Network for AI) provides domain-specific transforms, pretrained models, and workflows optimized for medical image segmentation.

Question 10: When a NCA professional encounters an unfamiliar challenge in computer vision & nlp workloads, what is the recommended first course of action?

Apply the solution used for the most recent similar problem without adaptation
Proceed based on personal intuition alone
Postpone addressing the issue indefinitely
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 11: Which file must be configured on the host system to allow Docker to authenticate with the NGC container registry before pulling images?

/etc/docker/daemon.json
~/.ngc/config
~/.docker/config.json (Correct answer)
/etc/nvidia/ngc.conf

Correct answer: ~/.docker/config.json

Docker stores registry credentials in ~/.docker/config.json after running `docker login nvcr.io`, which is used for subsequent pull authentication.

Question 12: Which flag in TensorRT's BuilderConfig enables layer-level profiling to identify per-layer latency bottlenecks?

MAX_WORKSPACE_SIZE flag
STRICT_TYPES flag
PROFILING_VERBOSITY set to LAYER_NAMES_ONLY or DETAILED (Correct answer)
INT8 calibration mode

Correct answer: PROFILING_VERBOSITY set to LAYER_NAMES_ONLY or DETAILED

Setting ProfilingVerbosity to LAYER_NAMES_ONLY or DETAILED causes TensorRT to record per-layer timing data accessible via the IExecutionContext profiler.

Question 13: 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?

M.2 Key M slot (PCIe Gen 4) (Correct answer)
USB 3.2 Gen 2
UART
MIPI CSI-2

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.

Question 14: What does the term 'epoch' mean in deep learning training?

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

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 15: Why does using a batch size of 32 instead of 33 often improve Tensor Core efficiency?

The OS scheduler prefers power-of-two process counts
32 is a multiple of 8, aligning the batch dimension with Tensor Core tile requirements (Correct answer)
Odd batch sizes cause CUDA kernel launch errors
cuDNN only supports even batch sizes

Correct answer: 32 is a multiple of 8, aligning the batch dimension with Tensor Core tile requirements

Tensor Core tiles require dimensions aligned to multiples of 8 (or 16); batch size 32 aligns perfectly while 33 requires padding that wastes cycles.

Question 16: What happens when a CUDA kernel accesses global memory at an unaligned address?

The data is silently corrupted
The access may require multiple memory transactions, reducing throughput (Correct answer)
A hardware exception is thrown
The warp is terminated

Correct answer: The access may require multiple memory transactions, reducing throughput

Unaligned global memory access can break coalescing, requiring additional transactions and lowering effective memory bandwidth.

Question 17: Which NVIDIA GPU architecture first introduced Tensor Cores for accelerating matrix multiply-accumulate operations?

Turing
Volta (Correct answer)
Pascal
Maxwell

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 18: Which memory type in CUDA is best suited for read-only data broadcast to all threads in a warp simultaneously?

Constant memory (Correct answer)
Local memory
Global memory
Texture memory with 2D filtering

Correct answer: Constant memory

Constant memory is cached and optimized for scenarios where all threads in a warp read the same address simultaneously, providing broadcast with very low latency.

Question 19: What is the purpose of calibration data during TensorRT INT8 engine building?

To verify that the model meets latency SLAs on the target hardware
To retrain the model with lower-precision gradients
To benchmark GPU throughput before deployment
To determine optimal quantization scale factors by running representative inputs through the FP32 model (Correct answer)

Correct answer: To determine optimal quantization scale factors by running representative inputs through the FP32 model

The calibrator uses a small representative dataset to compute per-tensor dynamic ranges, which TensorRT uses to set quantization scales that preserve accuracy.

Question 20: What does the InfiniBand IBTA specification define as the maximum MTU for HDR/NDR links?

8192 bytes
4096 bytes (Correct answer)
2048 bytes
16384 bytes

Correct answer: 4096 bytes

The InfiniBand specification defines a maximum MTU of 4096 bytes for packet payloads on HDR and NDR links.

Question 21: Which NVIDIA tool is used to profile GPU kernel performance and identify bottlenecks in deep learning workloads?

cuSPARSE
Nsight Systems / Nsight Compute (Correct answer)
CUDA GDB
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 22: How does NVIDIA's Spectrum-4 Ethernet switch accelerate AI workloads compared to a generic Ethernet switch?

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

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 23: How do you specify the number of thread blocks and threads per block when launching a CUDA kernel?

Via a configuration struct passed as the first argument
Inside the kernel using blockDim and gridDim
Using cudaLaunchKernel() API only
With the triple angle bracket <<<grid, block>>> syntax (Correct answer)

Correct answer: With the triple angle bracket <<<grid, block>>> syntax

CUDA kernels are launched with kernel<<<gridDim, blockDim>>>(...) where the triple angle brackets specify the execution configuration.

Question 24: What does the CUDA event timing API measure that CPU timers cannot accurately capture?

Host-to-device transfer bandwidth
Number of CUDA cores utilized
Precise GPU kernel execution time excluding CPU overhead (Correct answer)
GPU temperature during execution

Correct answer: Precise GPU kernel execution time excluding CPU overhead

CUDA events are recorded on the GPU timeline and measure actual kernel execution time without CPU-side scheduling jitter or synchronization overhead.

Question 25: Which technique most effectively hides global memory latency in CUDA kernels?

Using larger thread blocks
Increasing the number of active warps per SM (Correct answer)
Enabling ECC memory
Reducing kernel launch overhead

Correct answer: Increasing the number of active warps per SM

Higher occupancy (more active warps per SM) allows the GPU scheduler to switch to ready warps while others wait for memory, hiding latency effectively.

Question 26: When using NVIDIA Multi-Instance GPU (MIG) with containers, which environment variable specifies the MIG device instance to expose?

CUDA_VISIBLE_DEVICES
MIG_INSTANCE_ID
NVIDIA_MIG_DEVICE
NVIDIA_VISIBLE_DEVICES (Correct answer)

Correct answer: NVIDIA_VISIBLE_DEVICES

NVIDIA_VISIBLE_DEVICES is used to specify which GPU or MIG device instance (e.g., MIG-GPU-<uuid>/0/0) is exposed to the container.

Question 27: What is the significance of the 'NVIDIA Deep Learning Frameworks' containers on NGC compared to upstream framework containers?

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

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 28: What is the primary purpose of the cuDNN library in the NVIDIA software stack?

Providing optimized primitives for deep neural network operations (Correct answer)
Enabling multi-GPU communication via collectives
General-purpose dense matrix multiplication for all workloads
Managing GPU memory allocation and deallocation

Correct answer: Providing optimized primitives for deep neural network operations

cuDNN (CUDA Deep Neural Network library) provides highly tuned implementations of primitives such as convolution, pooling, normalization, and activation functions used in deep learning.

Question 29: In CUDA, what is 'occupancy'?

The number of thread blocks per grid
The percentage of global memory in use
The ratio of active warps to maximum warps on an SM (Correct answer)
The fraction of time the GPU is not idle

Correct answer: The ratio of active warps to maximum warps on an SM

Occupancy measures how many warps are active on an SM relative to the hardware maximum, indicating how well latency can be hidden.

Question 30: In a multi-GPU inference deployment, what technique reduces latency by splitting a single large request's tensor operations across multiple GPUs simultaneously?

Batched inference with data parallelism
Post-training quantization
Tensor parallelism for inference (Correct answer)
Pipeline parallelism with micro-batch scheduling

Correct answer: Tensor parallelism for inference

Tensor parallelism for inference partitions weight matrices across GPUs so a single request's matrix multiplications run in parallel, reducing per-request latency.

Question 31: When running an NGC container with GPU access using Docker, which runtime flag must be specified?

--gpu-enable
--nvidia-runtime
--cuda-device 0
--gpus all (Correct answer)

Correct answer: --gpus all

The `--gpus all` flag (or `--gpus device=<id>`) is required to expose GPU devices to a Docker container via the NVIDIA Container Toolkit.

Question 32: What is the primary role of NVIDIA DeepStream SDK in an edge AI pipeline?

Training deep learning models on edge devices
Compiling CUDA kernels for embedded GPUs
Managing OTA firmware updates for Jetson devices
Streaming optimized inference on multi-sensor video inputs (Correct answer)

Correct answer: Streaming optimized inference on multi-sensor video inputs

DeepStream is a streaming analytics toolkit that enables real-time, multi-stream video inference on NVIDIA edge hardware.

Question 33: How many GPUs can be directly interconnected in a single NVLink domain using NVIDIA's H100 SXM platform?

16 GPUs
32 GPUs
4 GPUs
8 GPUs (Correct answer)

Correct answer: 8 GPUs

H100 SXM servers connect 8 GPUs within a single NVLink domain using NVSwitch chips for all-to-all connectivity.

Question 34: Which NVIDIA technology enables a single GPU kernel to directly initiate RDMA communication without CPU involvement?

cuDNN distributed
NVLink P2P
CUDA Unified Memory
GPUDirect Async (Correct answer)

Correct answer: GPUDirect Async

GPUDirect Async allows CUDA kernels to trigger RDMA operations on ConnectX NICs directly, enabling communication-compute overlap without CPU wake-ups.

Question 35: What distinguishes NVIDIA Nsight Compute from NVIDIA Nsight Systems?

Nsight Compute provides deep per-kernel GPU metrics; Nsight Systems delivers system-wide CPU+GPU timeline analysis (Correct answer)
Nsight Compute is for Windows only; Nsight Systems is cross-platform
Both tools perform identical analysis but target different GPU generations
Nsight Compute profiles system-wide CPU+GPU timelines; Nsight Systems analyzes individual kernel metrics

Correct answer: Nsight Compute provides deep per-kernel GPU metrics; Nsight Systems delivers system-wide CPU+GPU timeline analysis

Nsight Compute is a GPU kernel profiler providing detailed hardware counter metrics per kernel, while Nsight Systems is a system-wide performance analyzer showing CPU, GPU, and memory activity over time.

Question 36: What is 'knowledge distillation' in the context of deep learning model compression?

Quantizing weight precision from FP32 to INT8
Pruning redundant neurons from a large network to reduce size
Training a smaller 'student' model to mimic the outputs of a larger 'teacher' model (Correct answer)
Applying L1/L2 regularization to reduce weight magnitudes

Correct answer: Training a smaller 'student' model to mimic the outputs of a larger 'teacher' model

Knowledge distillation transfers knowledge from a large teacher model to a compact student model by training the student on the teacher's soft output probabilities.

Question 37: What is the role of the Subnet Manager (SM) in an InfiniBand fabric?

Manages storage volume allocation
Monitors GPU utilization across nodes
Initializes the fabric, assigns LIDs, and computes routing tables (Correct answer)
Assigns IP addresses to compute nodes

Correct answer: Initializes the fabric, assigns LIDs, and computes routing tables

The Subnet Manager discovers InfiniBand topology, assigns Local Identifiers (LIDs), and programs routing tables in all switches.

Question 38: What is the maximum number of threads per block on modern NVIDIA GPUs?

1024 (Correct answer)
2048
512
256

Correct answer: 1024

Modern NVIDIA architectures support up to 1024 threads per block.

Question 39: What is the primary purpose of CUDA shared memory in performance optimization?

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
To increase the number of concurrent thread blocks

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 40: What role does continuous improvement play in multi-gpu computing & scaling for NCA certified professionals?

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

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

Continuous improvement is fundamental to professional practice in multi-gpu computing & scaling, involving regular evaluation, feedback integration, and process enhancement to maintain high standards.

Question 41: When using GradScaler with AMP in PyTorch, what problem does it solve?

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

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 42: In the context of TensorRT deployment, what does 'binding' refer to?

A TensorRT plugin that binds custom layers to built-in operations
A CPU-side handle that links model weights to inference context
A CUDA event used to synchronize between streams
A GPU memory buffer associated with a named network input or output tensor (Correct answer)

Correct answer: A GPU memory buffer associated with a named network input or output tensor

Bindings are GPU device pointers mapped to each named input/output tensor that must be provided when calling execute.

Question 43: In NVIDIA's BlueField DPU, what distinguishes it from a standard ConnectX SmartNIC?

Built-in GPU compute cores
Higher port speeds
Embedded Arm cores that run a full OS and offload host CPU networking tasks (Correct answer)
Support for InfiniBand only

Correct answer: Embedded Arm cores that run a full OS and offload host CPU networking tasks

BlueField DPUs include Arm processor cores capable of running a full Linux OS, enabling complete networking stack offload from the host CPU.

Question 44: In the context of NVIDIA's data center portfolio, what does TDP (Thermal Design Power) of 700W for the H100 SXM5 imply for data center planning?

Each GPU needs 700W of cooling capacity and power provisioning (Correct answer)
TDP only affects PCIe slot selection, not rack planning
700W is the minimum power needed to boot the GPU
The GPU can sustain 700W performance boosts indefinitely

Correct answer: Each GPU needs 700W of cooling capacity and power provisioning

TDP of 700W means the data center must provision at least 700W of power and equivalent cooling capacity per H100 SXM5 GPU in thermal and electrical planning.

Question 45: Which technique helps identify race conditions in GPU code?

Using synchronization and detection tools (Correct answer)
Ignoring synchronization.
Increasing thread count.
Disabling threads.

Correct answer: Using synchronization and detection tools

Race conditions occur in GPU code when multiple threads access and modify shared data concurrently without proper synchronization, leading to unpredictable results. Identifying and preventing these issues requires careful use of synchronization primitives like atomic operations or barriers, along with specialized detection tools. These tools help ensure data integrity and correct program execution in parallel environments.

Question 46: Which activation function outputs values between 0 and 1?

Tanh.
ReLU.
Sigmoid (Correct answer)
Linear.

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 47: In a multi-node multi-GPU setup, what is the function of a 'rendezvous' backend (e.g., etcd, Redis) in PyTorch distributed training?

It coordinates process discovery so all ranks can find each other at startup (Correct answer)
It manages GPU power states across the cluster
It stores model checkpoints across nodes
It compresses gradients before AllReduce

Correct answer: It coordinates process discovery so all ranks can find each other at startup

The rendezvous backend provides a shared key-value store that distributed processes use to discover peers and synchronize initialization.

Question 48: Which NVIDIA library provides GPU-accelerated primitives specifically optimized for deep neural network operations like convolutions and RNNs?

NCCL
cuBLAS
cuDNN (Correct answer)
Thrust

Correct answer: cuDNN

cuDNN (CUDA Deep Neural Network library) provides highly optimized primitives for deep learning operations including convolutions, pooling, and recurrent layers.

Question 49: What is overfitting in deep learning?

Model is undertrained.
Model generalizes well.
Model performs poorly on training data.
Model learns noise and poorly generalizes (Correct answer)

Correct answer: Model learns noise and poorly generalizes

Overfitting in deep learning occurs when a model learns the training data too well, including its noise and specific idiosyncrasies, rather than the underlying general patterns. Consequently, while the model performs exceptionally on the training set, its performance significantly degrades when presented with new, unseen data, indicating poor generalization ability. This often happens when a model is too complex for the amount of training data available.

Question 50: What is the significance of ECC (Error Correcting Code) memory support in NVIDIA data center GPUs?

It doubles the effective GPU memory capacity
It corrects single-bit memory errors to ensure computational accuracy for critical workloads (Correct answer)
It enables faster memory clock speeds by reducing voltage requirements
It allows GPU memory to be shared across multiple physical servers

Correct answer: It corrects single-bit memory errors to ensure computational accuracy for critical workloads

ECC memory detects and corrects single-bit errors and detects double-bit errors, which is critical for HPC and AI training where silent data corruption is unacceptable.

Question 51: What is the purpose of gradient compression techniques (e.g., PowerSGD) in large-scale multi-GPU training?

To increase the learning rate dynamically based on GPU count
To reduce the size of gradient tensors communicated during AllReduce and lower communication overhead (Correct answer)
To prune low-magnitude weights from the model during training
To convert gradients from FP32 to INT8 for inference

Correct answer: To reduce the size of gradient tensors communicated during AllReduce and lower communication overhead

Gradient compression approximates gradient tensors with lower-rank representations, reducing the communication bandwidth needed for AllReduce.

Question 52: Which memory space in CUDA is read-only, cached, and optimized for broadcast to all threads?

Local memory
Texture memory
Shared memory
Constant memory (Correct answer)

Correct answer: Constant memory

Constant memory is a read-only, cached region optimal when all threads read the same address simultaneously.

Question 53: Which parallelism strategy in distributed training splits individual layers across multiple GPUs, with each GPU holding a subset of each layer's parameters?

Tensor parallelism (Correct answer)
Data parallelism
Expert parallelism
Pipeline parallelism

Correct answer: Tensor parallelism

Tensor parallelism (intra-layer model parallelism) partitions weight matrices across GPUs, with each GPU computing a portion of each layer's operation.

Question 54: Which TensorRT feature allows different network segments to run at different precisions (e.g., some layers in INT8, others in FP16)?

Separate engine compilation per precision mode
Mixed-precision execution via per-layer precision constraints (Correct answer)
Precision cascading through nested engine calls
cuDNN mixed-mode fallback

Correct answer: Mixed-precision execution via per-layer precision constraints

TensorRT supports setting per-layer precision and data-type constraints, enabling a single engine to mix INT8, FP16, and FP32 layers.

Question 55: Which framework introduced the concept of 'eager execution' as its default mode, executing operations immediately without building a static graph?

Theano
TensorFlow 2.x (Correct answer)
MXNet
Caffe2

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 56: What is the primary role of the L1 cache in NVIDIA GPU Streaming Multiprocessors (SMs)?

Synchronize data between the CPU and GPU
Store long-term model weights persistently
Provide low-latency data access for threads within an SM (Correct answer)
Cache outputs from the display pipeline

Correct answer: Provide low-latency data access for threads within an SM

The L1 cache in each SM provides fast, low-latency access to frequently used data for threads executing within that SM.

Question 57: Which library is popular for deep learning with Python?

Pandas.
Scikit-learn.
PyTorch (Correct answer)
NumPy.

Correct answer: PyTorch

PyTorch is an extremely popular open-source machine learning library widely adopted for deep learning research and development with Python. It provides a flexible framework for building and training neural networks, known for its dynamic computational graph and strong GPU acceleration capabilities. While other libraries like Scikit-learn, NumPy, and Pandas are essential for data science, PyTorch is specifically designed for deep learning.

Question 58: What is the most effective way to measure success in tensor core & mixed precision within NCA professional practice?

Compare only with industry averages without considering context
Count only the number of activities completed
Rely solely on supervisor opinion
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 59: In a convolutional neural network, what does the 'receptive field' of a neuron refer to?

The total number of trainable parameters in the network
The number of filters in a convolutional layer
The spatial dimensions of feature maps after pooling
The region of the input image that influences a particular neuron's output (Correct answer)

Correct answer: The region of the input image that influences a particular neuron's output

The receptive field is the spatial extent of the input that a neuron 'sees,' which grows deeper in the network through stacked convolutions.

Question 60: In multi-GPU training, which NVIDIA library uses Tensor Cores for the all-reduce collective communication pattern?

NCCL (Correct answer)
MPI
cuBLAS
NVLINK

Correct answer: NCCL

NCCL (NVIDIA Collective Communications Library) implements optimized all-reduce operations that leverage NVLink bandwidth, though compute remains on Tensor Cores via training kernels.

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