CUDA and GPU Programming Quiz

Answers and explanations appear as you take the quiz.

1. In CUDA terminology, what is a 'kernel'?
   • A function run in parallel by many GPU threads
   • The core of an operating system
   • A type of memory chip
   • The GPU's cooling system
2. What does the SIMT execution model mean?
   • Slow Input, Many Transfers
   • Single Instruction, Multiple Threads run the same instruction on different data
   • Single Image, Multiple Textures
   • Sequential Instructions, Manual Timing
3. Why do PyTorch and TensorFlow rarely require users to write raw CUDA?
   • GPUs cannot run user code
   • CUDA is illegal to use directly
   • They call optimized CUDA libraries like cuDNN and cuBLAS internally
   • They only run on CPUs
4. What is 'warp divergence' and why does it hurt performance?
   • A networking error that drops packets
   • The GPU physically bending under heat
   • A bug that deletes memory
   • Threads in a warp take different branches and must run serially
5. Why is 'coalesced' memory access important in CUDA?
   • Adjacent threads reading adjacent addresses is far more efficient
   • It encrypts data automatically
   • It frees up disk space
   • It disables the GPU fans
6. What is OpenAI Triton in relation to CUDA programming?
   • A CUDA driver that replaces NVIDIA GPUs
   • A Python-like language and compiler for writing efficient GPU kernels
   • A model-serving server that only loads TensorFlow checkpoints
   • A framework that converts GPUs into CPUs
7. What does a CUDA kernel launch configuration specify?
   • A model tokenizer vocabulary
   • The number of training labels
   • The grid and block dimensions used to organize thread execution
   • The host operating system
8. Why might a CUDA kernel use shared memory?
   • To make data available across every cluster without communication
   • To replace all registers
   • To store the model permanently on the GPU
   • To let threads in a block reuse data from a faster on-chip memory space