Question 1. A manager asks why the team wants GPUs rather than more CPU servers for deep learning training. Which statement matches how NVIDIA's CUDA documentation describes the two architectures?
- Option A: GPUs have much higher single-thread performance than CPUs, so each individual operation finishes sooner
- Option B: A GPU runs thousands of threads in parallel, trading single-thread speed for much greater total throughput
- Option C: CPUs cannot perform the floating-point matrix operations that neural networks need
- Option D: GPUs dedicate more transistors to data caching and flow control, which is why they hide memory latency so well
Show answer and explanation
Correct answer: B. A GPU runs thousands of threads in parallel, trading single-thread speed for much greater total throughput
CPUs are designed for low latency on a serial sequence of operations and run a few tens of threads in parallel. GPUs devote more transistors to data processing and run thousands of threads at once, which suits the highly parallel arithmetic of deep learning.
- A (incorrect)The opposite is true: a GPU accepts lower single-thread performance in exchange for higher total throughput.
- B (correct)Correct. The CUDA guide describes the CPU as optimised to run a serial thread as fast as possible and the GPU as optimised for thousands of parallel threads and much greater total throughput.
- C (incorrect)CPUs can perform these operations; they simply run far fewer of them in parallel.
- D (incorrect)Dedicating more transistors to caching and flow control is how the CUDA guide characterises CPUs, not GPUs.
Reference: NVIDIA CUDA Programming Guide: Introduction (opens in a new tab)

