NCA-AIIO - Essential AI Knowledge - Section 1.2

Compare and contrast training and inference architecture requirements and considerations.

Compare the compute, memory bandwidth, and storage demands of training against inference, noting that training is batch-oriented and throughput-bound while inference prioritises low latency and high concurrency. Distinguish which GPU configurations and precision modes suit each workload.

TrainingInference

Practice question for this objective

Free sampleEssential AI Knowledgemedium

A team is comparing the memory subsystems of a data-centre GPU and a server CPU to understand why each suits different work. The GPU pairs its cores with high-bandwidth memory, while the CPU pairs its cores with large-capacity DDR DIMMs. Which trade-off does this difference most directly express?

  • AThe GPU favours very high memory bandwidth to feed thousands of parallel cores, while the CPU favours large capacity and lower-latency access for general-purpose working sets. Correct
  • BThe GPU favours large memory capacity per device, while the CPU favours higher bandwidth to keep its few cores saturated during serial work.
  • CThe GPU favours non-volatile memory to retain model weights across power cycles, while the CPU favours volatile memory for transient computation.
  • DThe GPU favours error-corrected memory exclusively, while the CPU favours uncorrected memory to reduce access latency on each request.
GPU memory is engineered for high bandwidth to feed parallel cores, while CPU memory is engineered for capacity and low-latency general-purpose access. Thousands of GPU cores need data delivered at very high rates, so GPUs use wide high-bandwidth memory. CPUs serve diverse workloads where capacity and per-request latency matter more, so they pair with large DDR DIMMs, mirroring the throughput-versus-latency design split.

Why A is correct: Correct. HBM gives the GPU the wide memory bandwidth its many cores demand, whereas CPU DDR memory prioritises capacity and latency for diverse general-purpose workloads, reflecting the throughput-versus-latency split.

Why B is wrong: Tempting because it sounds like a clean inversion, but it reverses reality: GPUs lead on bandwidth and CPUs lead on capacity, so the roles described are swapped.

Why C is wrong: Tempting because persistence matters for model storage, but both HBM and DDR are volatile; persistence is not the distinction between GPU and CPU main memory.

Why D is wrong: Tempting because ECC is associated with GPUs in some contexts, but server CPUs also widely use ECC DDR, so error correction is not the defining trade-off here.

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.