Essential AI Knowledge

A data-centre team is sizing GPU nodes for a large language model training job. Which characteristic most directly drives the decision to prioritise aggregate GPU memory capacity over single-GPU latency?

• ALow-latency GPU memory reduces the time-to-first-token metric that training pipelines report to operators.
• BQuantisation to INT8 precision shrinks memory footprint enough that capacity planning becomes irrelevant for training runs.
• CModel weights and optimizer states must all reside in GPU memory simultaneously during the forward and backward passes.check_circle Correct
• DThroughput-per-watt efficiency determines whether the cluster can sustain the sustained compute throughput needed for each training step.

Why A is wrong: Time-to-first-token is an inference serving metric and is not a primary training objective; confusing the two is a common mistake when candidates carry inference intuition into training contexts.

Why B is wrong: INT8 quantisation is primarily applied at inference to trade a small accuracy cost for speed and efficiency; mixed-precision training typically uses FP16 or BF16 with FP32 master weights, and capacity pressure remains significant.

Why C is correct: During training, the full model parameters, gradients, and optimizer states must coexist in accelerator memory at once, making total capacity the binding constraint rather than per-request latency.

Why D is wrong: Throughput-per-watt is a key metric when selecting inference hardware to maximise served requests within a power budget; for training, raw compute throughput and memory capacity take priority over energy efficiency ratios.

A machine learning engineer has trained a computer-vision model and now needs to reduce its inference latency for production deployment on an NVIDIA GPU. Which layer of the NVIDIA software stack performs graph optimisation, layer fusion, and precision calibration to produce a deployment-ready inference engine?

• ATriton Inference Server, which applies graph rewrites and precision reduction to each loaded model at startup before accepting requests.
• BcuDNN, which selects the fastest convolution algorithm for each layer and caches the plan, effectively fusing the graph at the kernel level.
• CTensorRT, which analyses the trained model graph, fuses compatible layers, and calibrates numerical precision to minimise inference latency on the target GPU.check_circle Correct
• DCUDA, which provides JIT compilation of GPU kernels that the runtime optimises and fuses when the inference workload is first profiled.

Why A is wrong: Triton Inference Server manages model lifecycle and routes inference requests to multiple backends; it does not itself perform graph optimisation or precision calibration. Those tasks are handled by TensorRT before the optimised engine is loaded into Triton.

Why B is wrong: cuDNN selects optimised kernel implementations for individual deep-learning operations such as convolutions, but it operates layer by layer within the training or inference framework; it does not perform whole-graph fusion, quantisation calibration, or produce a standalone serialised inference engine.

Why C is correct: TensorRT is NVIDIA's inference optimisation SDK. It takes a trained model, applies layer and tensor fusion to reduce kernel launch overhead, selects the fastest kernel implementations for the target GPU architecture, and can calibrate INT8 or FP16 precision to further reduce latency and memory footprint without significant accuracy loss.

Why D is wrong: CUDA's JIT compilation resolves kernel code for the target GPU architecture but does not perform model-level graph fusion or precision calibration across neural-network layers; those are responsibilities of the TensorRT optimiser that sits above the CUDA runtime.

An engineer is explaining to a new team member why a modern data-centre GPU dedicates far more of its silicon area to arithmetic logic units than a server CPU of comparable transistor count does. Which design priority does this allocation reflect?

• AThe GPU maximises aggregate throughput by running thousands of lightweight threads in parallel, so it spends transistors on many simple cores rather than on complex per-thread acceleration.check_circle Correct
• BThe GPU minimises single-thread latency by devoting most transistors to deep out-of-order pipelines and large branch predictors that keep one instruction stream moving quickly.
• CThe GPU reduces memory access latency by allocating most of its die to a very large shared last-level cache that holds the entire working set on chip.
• DThe GPU prioritises low idle power by packing transistors into aggressive power-gating and clock-gating control logic rather than into compute units.

Why A is correct: Correct. A GPU is a throughput-oriented design that hides latency with massive thread-level parallelism, so most of its die area goes to many simple arithmetic units rather than to control or large caches.

Why B is wrong: Tempting because deep pipelines and branch predictors are real techniques for speed, but those are CPU latency-optimisation features; GPUs deliberately spend that area on more arithmetic units instead.

Why C is wrong: Tempting because caches do reduce latency, but GPUs actually devote relatively little area to cache and rely on high-bandwidth memory and thread parallelism rather than a large cache to hide latency.

Why D is wrong: Tempting because power management matters in data centres, but power-gating logic is not where GPU die area goes; the dominant allocation is to parallel arithmetic units for throughput.

An operations team must categorise three components of an AI platform by how broad each concept is: the overall field of building systems that perform tasks needing human-like intelligence, the subset of that field whose systems learn from data, and the further subset that uses many-layered neural networks. Which ordering correctly arranges these three from the broadest concept to the narrowest?

• ADeep learning is the broadest field, machine learning sits inside it, and artificial intelligence is the narrowest specialisation.
• BMachine learning is the broadest field, artificial intelligence is a subset of it, and deep learning is a subset of artificial intelligence.
• CArtificial intelligence is the broadest field, machine learning is a subset of it, and deep learning is a subset of machine learning.check_circle Correct
• DArtificial intelligence, machine learning, and deep learning are three separate fields with no containment relationship between them.

Why A is wrong: Tempting because deep learning is the most advanced of the three and advanced can feel like broadest, but the nesting is the reverse: deep learning is the narrowest subset, not the containing field.

Why B is wrong: Tempting because machine learning dominates modern AI discussion, but ML is contained within AI rather than the other way around, so placing AI inside ML inverts the true relationship.

Why C is correct: Correct: AI is the umbrella field, machine learning is the subset whose systems learn from data, and deep learning is the further subset that uses many-layered neural networks, giving the nesting AI then ML then DL.

Why D is wrong: Tempting because each term has its own distinct meaning, but they are nested rather than separate: every deep learning system is a machine learning system, and every machine learning system is a form of AI.

Three weeks after deployment, a fraud-detection model's precision has fallen steadily even though its code is unchanged. Investigation shows that the statistical distribution of incoming transactions now differs from the data the model was trained on. Which lifecycle software capability is most directly responsible for catching this condition in production?

• AContainer image scanning, which inspects the deployed image for known vulnerabilities and outdated packages before it is promoted.
• BModel monitoring with drift detection, which compares the live input and prediction distributions against the training baseline and alerts when they diverge.check_circle Correct
• CHyperparameter tuning, which searches training configurations to maximise the validation metric of a candidate model.
• DA model registry, which catalogues trained model versions with their metadata so a specific version can be promoted or rolled back.

Why A is wrong: Image scanning is a sound production practice and runs against the deployed artefact, which makes it sound relevant, but it checks software security rather than the statistical behaviour of live inputs, so it cannot detect a shift in the data distribution.

Why B is correct: Drift detection within model monitoring continuously compares production input and output distributions to the training baseline and raises an alert when they diverge, which is exactly the silent, data-driven degradation described in the scenario.

Why C is wrong: Hyperparameter tuning influences how well a model fits its training data, which can seem connected to accuracy, but it runs during training and cannot observe or react to changes in the distribution of production traffic after deployment.

Why D is wrong: A model registry is essential for managing deployments and can roll back to an earlier version, which makes it feel like part of the fix, but it stores and versions artefacts and does not observe live traffic, so it cannot detect that distributions have shifted.