A weather-balloon shed is told “the training job uses CUDA.” An intern asks whether they must write a custom GPU kernel. What should they understand?
Select an answer to reveal the explanation.
Short Explanation
The intern asks whether they must write a custom GPU kernel. CUDA is the NVIDIA platform that lets existing frameworks run on the GPU. Awareness is enough; they do not author device code.
Full Explanation
At associate depth, CUDA is the acceleration layer that existing deep-learning frameworks use to run on NVIDIA GPUs. It is not a requirement to write device kernels or NCCL code. It is not a clustering method and it does not replace the loss.