#performance

Question 1: Single choice

mediumAI-checkedPyTorch

You time a function that multiplies small matrices many times, once on the CPU and once on a GPU. As you shrink the matrices and increase the number of iterations, the CPU version eventually becomes faster. What is the main reason?

Each GPU operation has a fixed overhead (kernel launch, synchronization, data transfer) that dominates tiny workloads1
GPUs switch to lower precision for small matrices, which forces extra correction steps2
Small matrices do not fit into GPU memory, so they are swapped out to the CPU3
The CPU caches the result of the first iteration and skips the remaining ones4

Timed test