NCA-AIIO - Essential AI Knowledge - Section 1.8

Compare and contrast GPU and CPU architectures.

Compare CPU architecture, which uses a small number of high-clock-speed cores optimised for sequential tasks, with GPU architecture, which uses thousands of smaller cores designed for parallel floating-point computation. Recognise why this parallelism makes GPUs vastly more efficient than CPUs for the matrix operations at the heart of deep learning.

GPUCPU

Practice question for this objective

Free sampleEssential AI Knowledgemedium

A data-centre engineer is evaluating whether to assign a large matrix-multiplication workload from a deep learning training job to a GPU or a CPU. Which architectural characteristic most directly explains why a GPU handles this workload more efficiently?

  • AGPUs contain a small number of high-frequency cores optimised for low-latency branch prediction and serial execution.
  • BGPUs use a DPU co-processor to offload matrix arithmetic from the main processing units, freeing those units for other tasks.
  • CGPUs expose thousands of simpler cores that execute the same instruction across many data elements simultaneously, matching the uniform parallel structure of matrix operations. Correct
  • DGPUs prioritise large per-core cache hierarchies that keep frequently accessed matrix rows resident, reducing instruction latency.
Explain why the GPU SIMT throughput model makes it better suited than a CPU for the data-parallel workloads found in deep learning training. A GPU contains thousands of relatively simple cores grouped into streaming multiprocessors. Using the Single Instruction, Multiple Thread (SIMT) execution model, the GPU issues one instruction that runs across hundreds of threads simultaneously. Matrix multiplication involves applying identical multiply-accumulate operations to large arrays, so every thread does the same work on a different data element at the same time. This maps directly onto SIMT throughput. A CPU, by contrast, has a small number of powerful cores each optimised for low-latency sequential execution and complex branch prediction, making it better suited to control-flow-heavy serial code than to massive parallel arithmetic.

Why A is wrong: This describes the CPU design philosophy, not the GPU. CPUs favour latency-optimised cores suited to sequential, branchy code rather than the throughput model used for matrix work.

Why B is wrong: A DPU (Data Processing Unit) offloads networking, storage, and security tasks from the host CPU; it is not a component inside a GPU and plays no role in matrix arithmetic acceleration.

Why C is correct: The GPU Single Instruction, Multiple Thread model runs thousands of threads in lock-step, directly exploiting the uniform, data-parallel nature of matrix multiplication where the same arithmetic is applied to large arrays of values.

Why D is wrong: CPUs invest heavily in large, multi-level caches to reduce latency for individual threads. GPUs instead hide latency by switching between many concurrent threads rather than caching for a few fast ones.

See more NCA-AIIO practice questions, answers explained.

More in this domain

Back to all Essential AI Knowledge objectives, or the NCA-AIIO cert hub.

Examworthy is not affiliated with or endorsed by NVIDIA. Original, blueprint-aligned practice material only.