A trolley-photo archive notices that stacking layers of matrix mult…

Question

A trolley-photo archive notices that stacking layers of matrix multiplies crawls on a laptop CPU but finishes far sooner when the same work is sent to a GPU. Why does that hardware help deep training?

Accepted Answer

GPUs accelerate neural-net training because they run highly parallel linear algebra. — Stacking matrix multiplies crawls on a laptop CPU and finishes sooner on a GPU because GPUs run highly parallel linear algebra. They do not replace the loss, and this is not a CUDA-kernel writing item.

Answer

GPUs replace the loss function, so a CPU-trained network has no training signal.

Answer

GPUs are required only for datacenter power, cooling, and job scheduling, not for matrix math.

Answer

A GPU is slower at matrix multiplies, so the speedup must come from writing a custom CUDA kernel on the exam.

A trolley-photo archive notices that stacking layers of matrix multiplies crawls on a laptop CPU but finishes far sooner when the same work is sent to a GPU. Why does that hardware help deep training?

Short Explanation

Full Explanation